{ "$schema": "./opportunity-map.schema.json", "schemaVersion": "1.0.0", "metadata": { "id": "diffusion-models-opportunity-map", "title": "Diffusion Models Opportunity Map", "anchorAtlasNodeId": "diffusion", "asOf": "2026-08-19", "status": "alpha", "summary": "Source-driven alpha mapping 60 developments, capabilities, complements, applications, constraints, outcomes, competitors, and bounded hypotheses across diffusion models. Historical influence is separated from later mathematical equivalence; branch maturity is scoped; opportunity cards are falsifiable hypotheses rather than novelty or investment claims.", "timeDomain": { "startYear": 1908, "endYear": 2026, "focusStartYear": 2015 }, "visualBands": [ { "id": "band-precursors", "label": "Precursors", "description": "Documented or contextual antecedents; chronology does not by itself imply causality.", "order": 0, "nodeTypes": [ "precursor" ] }, { "id": "band-core", "label": "Core development", "description": "The 2015 diffusion, 2019 score, 2020 DDPM, and 2021 score-SDE core.", "order": 1, "nodeTypes": [ "core_development" ] }, { "id": "band-capabilities", "label": "Capabilities, refinements and complements", "description": "What the core enables, how it was improved, and the independently useful complements that unlocked applications.", "order": 2, "nodeTypes": [ "capability", "refinement", "complement" ] }, { "id": "band-applications", "label": "Applications", "description": "Eight bounded application branches with scope-specific maturity assessments.", "order": 3, "nodeTypes": [ "application" ] }, { "id": "band-outcomes", "label": "Outcomes and alternatives", "description": "Locally stalled attempts and competing or substitute paradigms; no global dead-end claim is implied.", "order": 4, "nodeTypes": [ "demonstrated_outcome", "failed_or_stalled_attempt", "competing_or_substitute_approach" ] }, { "id": "band-constraints", "label": "Constraints", "description": "Active bottlenecks, partially mitigated limits, and evaluation gaps.", "order": 5, "nodeTypes": [ "constraint" ] }, { "id": "band-frontier", "label": "Open frontier", "description": "Falsifiable, evidence-linked research hypotheses. Inclusion does not establish novelty, patentability, feasibility, or expected value.", "order": 6, "nodeTypes": [ "open_opportunity" ] } ], "importStatus": { "state": "imported_unreviewed", "notes": "Mapped from the supplied source-driven report and passed through structural validation. The report itself requires manual bibliography, primary-lineage, patent, clinical/science, licensing, and matched-replication review before publication-level promotion.", "reportId": "deep-research-report(1).md", "reportAsOf": "2026-08-19" }, "pathWidthMode": "fixed" }, "nodes": [ { "id": "p01", "type": "precursor", "title": "Nonequilibrium transport and annealed importance sampling", "summary": "Slow Markov transport between distributions supplied the physics/SMC analogy explicitly used in the 2015 paper.", "year": 1997, "yearEnd": 2001, "yearPrecision": "range", "domain": "core", "tags": [ "precursor", "core" ], "bandId": "band-precursors", "status": { "state": "mature_but_useful", "scope": "Nonequilibrium transport and annealed importance sampling is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s001", "s002", "s007" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Slow Markov transport between distributions supplied the physics/SMC analogy explicitly used in the 2015 paper.", "sourceIds": [ "s001", "s002", "s007" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "mcmc", "relation": "related_to", "note": "Annealed importance sampling belongs to the atlas probabilistic-inference lineage." } ] }, { "id": "p02", "type": "precursor", "title": "Diffusion, Langevin, Fokker-Planck and time reversal", "summary": "Stochastic diffusion and reverse-process mathematics supplied the vocabulary for forward noising and reverse dynamics.", "year": 1908, "yearEnd": 1949, "yearPrecision": "range", "domain": "core", "tags": [ "precursor", "core" ], "bandId": "band-precursors", "status": { "state": "mature_but_useful", "scope": "Diffusion, Langevin, Fokker-Planck and time reversal is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s007" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Stochastic diffusion and reverse-process mathematics supplied the vocabulary for forward noising and reverse dynamics.", "sourceIds": [ "s007" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "precursor_of", "note": "Stochastic diffusion and time-reversal mathematics precede the atlas diffusion anchor." } ] }, { "id": "p03", "type": "precursor", "title": "Score matching", "summary": "Learns scores of unnormalized densities without evaluating a partition function.", "year": 2005, "yearPrecision": "year", "domain": "core", "tags": [ "precursor", "core" ], "bandId": "band-precursors", "status": { "state": "mature_but_useful", "scope": "Score matching is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s003" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Learns scores of unnormalized densities without evaluating a partition function.", "sourceIds": [ "s003" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "precursor_of", "note": "Score matching is a formal precursor to the score strand represented inside the diffusion anchor." } ] }, { "id": "p04", "type": "precursor", "title": "Denoising-score / denoising-autoencoder equivalence", "summary": "Shows that denoising corrupted samples can recover a score; later directly invoked in the DDPM interpretation.", "year": 2011, "yearPrecision": "year", "domain": "core", "tags": [ "precursor", "core" ], "bandId": "band-precursors", "status": { "state": "mature_but_useful", "scope": "Denoising-score / denoising-autoencoder equivalence is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s004", "s009" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Shows that denoising corrupted samples can recover a score; later directly invoked in the DDPM interpretation.", "sourceIds": [ "s004", "s009" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "related_to", "note": "The report supports a later denoising-score equivalence, not a direct 2015 causal-origin claim." } ] }, { "id": "p05", "type": "precursor", "title": "Variational latent-variable learning", "summary": "Supplies scalable variational bounds/reparameterized latent-variable learning; C01 says its bound is similar but its motivation/model differ.", "year": 2013, "yearEnd": 2014, "yearPrecision": "range", "domain": "core", "tags": [ "precursor", "core" ], "bandId": "band-precursors", "status": { "state": "mature_but_useful", "scope": "Variational latent-variable learning is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "partial", "sourceIds": [ "s005", "s007" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "partial", "claim": "Supplies scalable variational bounds/reparameterized latent-variable learning; C01 says its bound is similar but its motivation/model differ.", "sourceIds": [ "s005", "s007" ], "note": "Evidence confidence assigned by the supplied report: partial." } ], "atlasLinks": [ { "atlasNodeId": "vae", "relation": "same_as", "note": "Bounded record for variational latent-variable learning." } ] }, { "id": "p06", "type": "precursor", "title": "Energy-based and neural generative modeling context", "summary": "EBMs, neural generators, VAEs, flows, GANs and autoregressive models formed the competing generative landscape. The 2014–2019 visual placement operationalizes the report’s “pre-2020” label from its cited source window; it is not an origin date for the entire energy-based-model tradition.", "year": 2014, "yearEnd": 2019, "yearPrecision": "range", "domain": "core", "tags": [ "precursor", "core" ], "bandId": "band-precursors", "status": { "state": "mature_but_useful", "scope": "Energy-based and neural generative modeling context is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "contextual", "sourceIds": [ "s006", "s007", "s009" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "editorial_analysis", "grade": "contextual", "claim": "EBMs, neural generators, VAEs, flows, GANs and autoregressive models formed the competing generative landscape.", "sourceIds": [ "s006", "s007", "s009" ], "note": "Evidence confidence assigned by the supplied report: contextual." } ], "atlasLinks": [ { "atlasNodeId": "gan", "relation": "related_to", "note": "GANs are part of the competing pre-DDPM generative context." }, { "atlasNodeId": "flows", "relation": "related_to", "note": "Normalizing flows are part of the adjacent generative context." } ] }, { "id": "c01", "type": "core_development", "title": "Diffusion probabilistic models", "summary": "Defines a learned finite-time reverse Markov diffusion from simple noise to data.", "year": 2015, "yearPrecision": "year", "domain": "core", "tags": [ "core-development", "core" ], "bandId": "band-core", "status": { "state": "mature_but_useful", "scope": "Diffusion probabilistic models is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s007" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Defines a learned finite-time reverse Markov diffusion from simple noise to data.", "sourceIds": [ "s007" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "precursor_of", "note": "The 2015 diffusion probabilistic model predates the atlas anchor convention centered on DDPM-era diffusion." } ] }, { "id": "c02", "type": "core_development", "title": "Noise-conditioned score networks", "summary": "Learns scores over noise levels and samples with annealed Langevin dynamics.", "year": 2019, "yearPrecision": "year", "domain": "core", "tags": [ "core-development", "core" ], "bandId": "band-core", "status": { "state": "mature_but_useful", "scope": "Noise-conditioned score networks is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s008" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Learns scores over noise levels and samples with annealed Langevin dynamics.", "sourceIds": [ "s008" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "precursor_of", "note": "Noise-conditioned score networks form a direct precursor strand." } ] }, { "id": "c03", "type": "core_development", "title": "Denoising diffusion probabilistic models", "summary": "Establishes practical high-quality image diffusion and the noise-prediction/denoising-score connection.", "year": 2020, "yearPrecision": "year", "domain": "core", "tags": [ "core-development", "core" ], "bandId": "band-core", "status": { "state": "mature_but_useful", "scope": "Denoising diffusion probabilistic models is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s009" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Establishes practical high-quality image diffusion and the noise-prediction/denoising-score connection.", "sourceIds": [ "s009" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "same_as", "note": "Primary cross-view anchor for diffusion models." } ] }, { "id": "c04", "type": "core_development", "title": "Score-SDE and probability-flow formulation", "summary": "Unifies score and diffusion models in continuous time; adds reverse SDE, probability-flow ODE and inverse-problem tools.", "year": 2021, "yearPrecision": "year", "domain": "core", "tags": [ "core-development", "core" ], "bandId": "band-core", "status": { "state": "mature_but_useful", "scope": "Score-SDE and probability-flow formulation is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s010" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Unifies score and diffusion models in continuous time; adds reverse SDE, probability-flow ODE and inverse-problem tools.", "sourceIds": [ "s010" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "descendant_of", "note": "The score-SDE formulation is a later unification of diffusion and score models." } ] }, { "id": "cap01", "type": "capability", "title": "Local denoising as scalable generative learning", "summary": "Replaces one difficult global density model with many local corruption-reversal prediction tasks.", "year": 2015, "yearEnd": 2020, "yearPrecision": "range", "domain": "core", "tags": [ "capability", "core" ], "bandId": "band-capabilities", "status": { "state": "mature_but_useful", "scope": "Local denoising as scalable generative learning is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s007", "s009" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Replaces one difficult global density model with many local corruption-reversal prediction tasks.", "sourceIds": [ "s007", "s009" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "capability_of", "note": "Local denoising is a core capability of the atlas diffusion development." } ] }, { "id": "cap02", "type": "capability", "title": "Multiscale score field and stochastic refinement", "summary": "Provides coarse-to-fine stochastic generation and score estimates across noise scales.", "year": 2019, "yearEnd": 2021, "yearPrecision": "range", "domain": "core", "tags": [ "capability", "core" ], "bandId": "band-capabilities", "status": { "state": "mature_but_useful", "scope": "Multiscale score field and stochastic refinement is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s008", "s010" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Provides coarse-to-fine stochastic generation and score estimates across noise scales.", "sourceIds": [ "s008", "s010" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "capability_of", "note": "Multiscale score fields are a core capability in the diffusion/score family." } ] }, { "id": "cap03", "type": "capability", "title": "Strong conditional guidance", "summary": "Modifies conditional/unconditional score estimates to trade diversity for adherence.", "year": 2021, "yearEnd": 2022, "yearPrecision": "range", "domain": "multimodal", "tags": [ "capability", "multimodal" ], "bandId": "band-capabilities", "status": { "state": "mature_but_useful", "scope": "Strong conditional guidance is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s013", "s014" ], "note": "Report branch classification: multimodal." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Modifies conditional/unconditional score estimates to trade diversity for adherence.", "sourceIds": [ "s013", "s014" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "txt2img", "relation": "capability_of", "note": "Conditional guidance is central to text-to-image diffusion." } ] }, { "id": "cap04", "type": "capability", "title": "Reusable generative prior for posterior inference", "summary": "Combines a pretrained generative prior with measurements or likelihoods.", "year": 2015, "yearEnd": 2022, "yearPrecision": "range", "domain": "inverse/science", "tags": [ "capability", "inverse/science" ], "bandId": "band-capabilities", "status": { "state": "actively_improving", "scope": "Reusable generative prior for posterior inference is assessed as actively improving only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s007", "s010", "s031", "s032" ], "note": "Report branch classification: inverse/science." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Combines a pretrained generative prior with measurements or likelihoods.", "sourceIds": [ "s007", "s010", "s031", "s032" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "capability_of", "note": "Generative-prior reuse is a capability of diffusion and score models." } ] }, { "id": "cap05", "type": "capability", "title": "Scalable high-resolution multimodal conditional synthesis", "summary": "Latents, text encoders, cascades and large backbones make high-res semantic synthesis practical.", "year": 2021, "yearEnd": 2024, "yearPrecision": "range", "domain": "media", "tags": [ "capability", "media" ], "bandId": "band-capabilities", "status": { "state": "actively_improving", "scope": "Scalable high-resolution multimodal conditional synthesis is assessed as actively improving only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s015", "s016", "s020", "s027" ], "note": "Report branch classification: media." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Latents, text encoders, cascades and large backbones make high-res semantic synthesis practical.", "sourceIds": [ "s015", "s016", "s020", "s027" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "txt2img", "relation": "capability_of", "note": "Efficient high-resolution semantic synthesis underlies the text-to-image branch." } ] }, { "id": "cap06", "type": "capability", "title": "Structured sequence, trajectory and geometry generation", "summary": "Extends denoising to trajectories, point clouds, molecules, protein frames and action chunks.", "year": 2021, "yearEnd": 2024, "yearPrecision": "range", "domain": "3D/science/robotics", "tags": [ "capability", "3d/science/robotics" ], "bandId": "band-capabilities", "status": { "state": "rapidly_expanding", "scope": "Structured sequence, trajectory and geometry generation is assessed as rapidly expanding only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s046", "s052", "s058", "s060" ], "note": "Report branch classification: 3D/science/robotics." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Extends denoising to trajectories, point clouds, molecules, protein frames and action chunks.", "sourceIds": [ "s046", "s052", "s058", "s060" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "capability_of", "note": "Structured-object denoising extends the core diffusion capability." } ] }, { "id": "r01", "type": "refinement", "title": "Learned reverse variances and improved objectives", "summary": "Improves likelihood and permits substantial sampling-step reduction.", "year": 2021, "yearPrecision": "year", "domain": "core", "tags": [ "refinement", "core" ], "bandId": "band-capabilities", "status": { "state": "mature_but_useful", "scope": "Learned reverse variances and improved objectives is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s011" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Improves likelihood and permits substantial sampling-step reduction.", "sourceIds": [ "s011" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "refinement_of", "note": "Improved objectives and learned variances refine DDPM-style diffusion." } ] }, { "id": "r02", "type": "refinement", "title": "DDIM and solver-based acceleration", "summary": "Deterministic/non-Markovian paths and specialized solvers lower the required network evaluations.", "year": 2020, "yearEnd": 2022, "yearPrecision": "range", "domain": "core", "tags": [ "refinement", "core" ], "bandId": "band-capabilities", "status": { "state": "mature_but_useful", "scope": "DDIM and solver-based acceleration is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s012", "s022" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Deterministic/non-Markovian paths and specialized solvers lower the required network evaluations.", "sourceIds": [ "s012", "s022" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "refinement_of", "note": "DDIM and solvers accelerate diffusion sampling." } ] }, { "id": "r03", "type": "refinement", "title": "Classifier and classifier-free guidance", "summary": "Makes conditional fidelity tunable; CFG removes the auxiliary classifier.", "year": 2021, "yearEnd": 2022, "yearPrecision": "range", "domain": "multimodal", "tags": [ "refinement", "multimodal" ], "bandId": "band-capabilities", "status": { "state": "mature_but_useful", "scope": "Classifier and classifier-free guidance is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s013", "s014" ], "note": "Report branch classification: multimodal." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Makes conditional fidelity tunable; CFG removes the auxiliary classifier.", "sourceIds": [ "s013", "s014" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "txt2img", "relation": "refinement_of", "note": "Guidance refines conditional image generation." } ] }, { "id": "r04", "type": "refinement", "title": "Cascaded diffusion", "summary": "Stacks base generation with spatial/temporal super-resolution.", "year": 2021, "yearEnd": 2022, "yearPrecision": "range", "domain": "image/video", "tags": [ "refinement", "image/video" ], "bandId": "band-capabilities", "status": { "state": "mature_but_useful", "scope": "Cascaded diffusion is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s016", "s036" ], "note": "Report branch classification: image/video." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Stacks base generation with spatial/temporal super-resolution.", "sourceIds": [ "s016", "s036" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "video", "relation": "refinement_of", "note": "Cascades refine spatial and temporal generation." } ] }, { "id": "r05", "type": "refinement", "title": "Latent diffusion", "summary": "Runs denoising in a compressed perceptual latent instead of pixels.", "year": 2021, "yearEnd": 2022, "yearPrecision": "range", "domain": "media", "tags": [ "refinement", "media" ], "bandId": "band-capabilities", "status": { "state": "mature_but_useful", "scope": "Latent diffusion is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s015" ], "note": "Report branch classification: media." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Runs denoising in a compressed perceptual latent instead of pixels.", "sourceIds": [ "s015" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "txt2img", "relation": "refinement_of", "note": "Latent diffusion is a defining refinement of the atlas text-to-image branch." } ] }, { "id": "r06", "type": "refinement", "title": "Diffusion Transformers", "summary": "Replaces convolutional U-Nets with patch/token Transformers and creates a clearer scaling path.", "year": 2022, "yearEnd": 2024, "yearPrecision": "range", "domain": "media", "tags": [ "refinement", "media" ], "bandId": "band-capabilities", "status": { "state": "actively_improving", "scope": "Diffusion Transformers is assessed as actively improving only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s020", "s027" ], "note": "Report branch classification: media." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Replaces convolutional U-Nets with patch/token Transformers and creates a clearer scaling path.", "sourceIds": [ "s020", "s027" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "dit", "relation": "same_as", "note": "Direct mapping to the Diffusion Transformers atlas node." } ] }, { "id": "r07", "type": "refinement", "title": "Distillation and consistency generation", "summary": "Compresses many denoising evaluations into few or one-step generators.", "year": 2022, "yearEnd": 2024, "yearPrecision": "range", "domain": "core", "tags": [ "refinement", "core" ], "bandId": "band-capabilities", "status": { "state": "actively_improving", "scope": "Distillation and consistency generation is assessed as actively improving only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s023", "s024" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Compresses many denoising evaluations into few or one-step generators.", "sourceIds": [ "s023", "s024" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "refinement_of", "note": "Distillation and consistency methods reduce diffusion evaluations." } ] }, { "id": "r08", "type": "refinement", "title": "Discrete, categorical and masked diffusion", "summary": "Generalizes corruption to discrete transitions; large masked diffusion LMs show modern scaling.", "year": 2021, "yearEnd": 2026, "yearPrecision": "range", "domain": "discrete/language side-lineage", "tags": [ "refinement", "discrete/language side-lineage" ], "bandId": "band-capabilities", "status": { "state": "rapidly_expanding", "scope": "Discrete, categorical and masked diffusion is assessed as rapidly expanding only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s028", "s029" ], "note": "Report branch classification: discrete/language side-lineage." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Generalizes corruption to discrete transitions; large masked diffusion LMs show modern scaling.", "sourceIds": [ "s028", "s029" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusionllm", "relation": "same_as", "note": "Direct mapping to the diffusion-language-model side lineage." } ] }, { "id": "m01", "type": "complement", "title": "U-Net and residual convolutional denoisers", "summary": "Supplies multiscale image inductive bias to DDPM-era systems.", "year": 2015, "yearEnd": 2021, "yearPrecision": "range", "domain": "image/inverse", "tags": [ "complement", "image/inverse" ], "bandId": "band-capabilities", "status": { "state": "mature_but_useful", "scope": "U-Net and residual convolutional denoisers is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s009" ], "note": "Report branch classification: image/inverse." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Supplies multiscale image inductive bias to DDPM-era systems.", "sourceIds": [ "s009" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "complement_to", "note": "U-Net-like denoisers are a complement rather than an origin node." } ] }, { "id": "m02", "type": "complement", "title": "Large text encoders and paired multimodal data", "summary": "Makes open-vocabulary semantic conditioning strong.", "year": 2021, "yearEnd": 2023, "yearPrecision": "range", "domain": "image/video/audio", "tags": [ "complement", "image/video/audio" ], "bandId": "band-capabilities", "status": { "state": "mature_but_useful", "scope": "Large text encoders and paired multimodal data is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s016", "s017" ], "note": "Report branch classification: image/video/audio." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Makes open-vocabulary semantic conditioning strong.", "sourceIds": [ "s016", "s017" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "txt2img", "relation": "complement_to", "note": "Language encoders and paired data complement text-to-image diffusion." } ] }, { "id": "m03", "type": "complement", "title": "Perceptual autoencoder latents", "summary": "Compresses media into a cheaper space for denoising.", "year": 2021, "yearEnd": 2022, "yearPrecision": "range", "domain": "media", "tags": [ "complement", "media" ], "bandId": "band-capabilities", "status": { "state": "mature_but_useful", "scope": "Perceptual autoencoder latents is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s015", "s043" ], "note": "Report branch classification: media." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Compresses media into a cheaper space for denoising.", "sourceIds": [ "s015", "s043" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "txt2img", "relation": "complement_to", "note": "Perceptual autoencoders enable latent diffusion." } ] }, { "id": "m04", "type": "complement", "title": "Transformers and accelerator-scale compute", "summary": "Enables large token/patch denoisers, flow models and VLA action heads.", "year": 2017, "yearEnd": 2026, "yearPrecision": "range", "domain": "media/robotics", "tags": [ "complement", "media/robotics" ], "bandId": "band-capabilities", "status": { "state": "actively_improving", "scope": "Transformers and accelerator-scale compute is assessed as actively improving only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s020", "s027", "s062" ], "note": "Report branch classification: media/robotics." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Enables large token/patch denoisers, flow models and VLA action heads.", "sourceIds": [ "s020", "s027", "s062" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "transformer", "relation": "related_to", "note": "Transformer backbones are one component of this broader compute complement." }, { "atlasNodeId": "diffusion", "relation": "complement_to", "note": "Accelerator-scale Transformer compute complements diffusion objectives." } ] }, { "id": "m05", "type": "complement", "title": "Geometry, equivariance and differentiable rendering", "summary": "Adapts generation to 3D, molecules, proteins and structured scenes.", "year": 2022, "yearEnd": 2026, "yearPrecision": "range", "domain": "3D/science", "tags": [ "complement", "3d/science" ], "bandId": "band-capabilities", "status": { "state": "actively_improving", "scope": "Geometry, equivariance and differentiable rendering is assessed as actively improving only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s047", "s051", "s052", "s055" ], "note": "Report branch classification: 3D/science." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Adapts generation to 3D, molecules, proteins and structured scenes.", "sourceIds": [ "s047", "s051", "s052", "s055" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "neural3d", "relation": "related_to", "note": "Geometry and differentiable rendering connect the opportunity branch to neural 3D." } ] }, { "id": "m06", "type": "complement", "title": "Simulators, demonstrations and domain validators", "summary": "Provides supervision and external validity signals for science, robotics and world models.", "year": 2022, "yearEnd": 2026, "yearPrecision": "range", "domain": "science/robotics/world", "tags": [ "complement", "science/robotics/world" ], "bandId": "band-capabilities", "status": { "state": "rapidly_expanding", "scope": "Simulators, demonstrations and domain validators is assessed as rapidly expanding only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s060", "s061", "s064", "s065" ], "note": "Report branch classification: science/robotics/world." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Provides supervision and external validity signals for science, robotics and world models.", "sourceIds": [ "s060", "s061", "s064", "s065" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "simenv", "relation": "related_to", "note": "Simulators are one of the domain-validation complements in this record." } ] }, { "id": "a01", "type": "application", "title": "Image synthesis and editing", "summary": "Unconditional, text-conditioned, inpainting, structural control and image editing.", "year": 2020, "yearEnd": 2026, "yearPrecision": "range", "domain": "image", "tags": [ "application", "image" ], "bandId": "band-applications", "status": { "state": "mature_but_useful", "scope": "Image synthesis and editing is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s009", "s015", "s016", "s017", "s018", "s019", "s027" ], "note": "Report branch classification: image." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Unconditional, text-conditioned, inpainting, structural control and image editing.", "sourceIds": [ "s009", "s015", "s016", "s017", "s018", "s019", "s027" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "txt2img", "relation": "same_as", "note": "Direct mapping to text-to-image and editing." } ] }, { "id": "a02", "type": "application", "title": "Image restoration and inverse problems", "summary": "Inpainting, deblurring, super-resolution, MRI/CT and general inverse reconstruction.", "year": 2015, "yearEnd": 2026, "yearPrecision": "range", "domain": "inverse", "tags": [ "application", "inverse" ], "bandId": "band-applications", "status": { "state": "constraint_bound", "scope": "Image restoration and inverse problems is assessed as constraint-bound only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s007", "s030", "s031", "s032", "s033", "s069" ], "note": "Report branch classification: inverse." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Inpainting, deblurring, super-resolution, MRI/CT and general inverse reconstruction.", "sourceIds": [ "s007", "s030", "s031", "s032", "s033", "s069" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "application_of", "note": "Inverse reconstruction reuses diffusion and score models as priors." } ] }, { "id": "a03", "type": "application", "title": "Video generation", "summary": "Spatiotemporal denoising/transport for conditional video and visual prediction.", "year": 2022, "yearEnd": 2026, "yearPrecision": "range", "domain": "video", "tags": [ "application", "video" ], "bandId": "band-applications", "status": { "state": "rapidly_expanding", "scope": "Video generation is assessed as rapidly expanding only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s035", "s036", "s037", "s038", "s039" ], "note": "Report branch classification: video." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Spatiotemporal denoising/transport for conditional video and visual prediction.", "sourceIds": [ "s035", "s036", "s037", "s038", "s039" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "video", "relation": "same_as", "note": "Direct mapping to the video-generation atlas node." } ] }, { "id": "a04", "type": "application", "title": "Audio, speech and music", "summary": "Waveform diffusion, TTS, text-to-audio and music generation.", "year": 2020, "yearEnd": 2026, "yearPrecision": "range", "domain": "audio", "tags": [ "application", "audio" ], "bandId": "band-applications", "status": { "state": "mature_but_useful", "scope": "Audio, speech and music is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s040", "s041", "s042", "s043", "s044", "s045" ], "note": "Report branch classification: audio." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Waveform diffusion, TTS, text-to-audio and music generation.", "sourceIds": [ "s040", "s041", "s042", "s043", "s044", "s045" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "wavenet", "relation": "related_to", "note": "The atlas neural-audio node provides adjacent audio-generation context." } ] }, { "id": "a05", "type": "application", "title": "3D, geometry and scene generation", "summary": "Point clouds, implicit objects, score-distilled text-to-3D and multiview generation.", "year": 2021, "yearEnd": 2026, "yearPrecision": "range", "domain": "3D", "tags": [ "application", "3d" ], "bandId": "band-applications", "status": { "state": "actively_improving", "scope": "3D, geometry and scene generation is assessed as actively improving only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s046", "s047", "s048", "s049", "s050", "s051" ], "note": "Report branch classification: 3D." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Point clouds, implicit objects, score-distilled text-to-3D and multiview generation.", "sourceIds": [ "s046", "s047", "s048", "s049", "s050", "s051" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "neural3d", "relation": "related_to", "note": "The 3D diffusion branch overlaps the atlas neural-3D lineage." } ] }, { "id": "a06", "type": "application", "title": "Molecules, proteins, materials and scientific design", "summary": "Molecular poses, protein backbones and crystalline/material structures.", "year": 2022, "yearEnd": 2026, "yearPrecision": "range", "domain": "science", "tags": [ "application", "science" ], "bandId": "band-applications", "status": { "state": "rapidly_expanding", "scope": "Molecules, proteins, materials and scientific design is assessed as rapidly expanding only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s052", "s053", "s054", "s055", "s056" ], "note": "Report branch classification: science." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Molecular poses, protein backbones and crystalline/material structures.", "sourceIds": [ "s052", "s053", "s054", "s055", "s056" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "proteinlm", "relation": "related_to", "note": "RFdiffusion and protein design overlap the protein generative-design node." }, { "atlasNodeId": "materials", "relation": "related_to", "note": "Crystal and material diffusion overlap the materials-discovery node." } ] }, { "id": "a07", "type": "application", "title": "Robotics, policies, trajectories and planning", "summary": "Trajectory planning, offline RL and multimodal visuomotor action generation.", "year": 2022, "yearEnd": 2026, "yearPrecision": "range", "domain": "robotics", "tags": [ "application", "robotics" ], "bandId": "band-applications", "status": { "state": "reopened_by_a_new_complement", "scope": "Robotics, policies, trajectories and planning is assessed as reopened by a new complement only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s058", "s059", "s060", "s061", "s062", "s063" ], "note": "Report branch classification: robotics." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Trajectory planning, offline RL and multimodal visuomotor action generation.", "sourceIds": [ "s058", "s059", "s060", "s061", "s062", "s063" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "rtfm", "relation": "related_to", "note": "Diffusion and flow policies overlap the robot-foundation-model branch." }, { "atlasNodeId": "offlinerl", "relation": "related_to", "note": "Trajectory diffusion also overlaps offline RL and sequence policies." } ] }, { "id": "a08", "type": "application", "title": "Simulation, synthetic data and world modeling", "summary": "Synthetic training data and interactive learned visual simulators.", "year": 2023, "yearEnd": 2026, "yearPrecision": "range", "domain": "world/synthetic", "tags": [ "application", "world/synthetic" ], "bandId": "band-applications", "status": { "state": "rapidly_expanding", "scope": "Simulation, synthetic data and world modeling is assessed as rapidly expanding only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s064", "s065", "s066", "s067" ], "note": "Report branch classification: world/synthetic." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Synthetic training data and interactive learned visual simulators.", "sourceIds": [ "s064", "s065", "s066", "s067" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "worldmodels18", "relation": "related_to", "note": "Generative world models continue the learned-world lineage." }, { "atlasNodeId": "genie", "relation": "related_to", "note": "Interactive learned simulators overlap the atlas interactive-world-model node." }, { "atlasNodeId": "datacentric", "relation": "related_to", "note": "Synthetic-data use overlaps data-centric AI." } ] }, { "id": "k01", "type": "constraint", "title": "Iterative sampling latency", "summary": "Repeated network evaluations impede real-time, video, edge and control use.", "year": 2020, "yearEnd": 2026, "yearPrecision": "range", "domain": "cross-cutting", "tags": [ "constraint", "cross-cutting" ], "bandId": "band-constraints", "status": { "state": "actively_improving", "scope": "Iterative sampling latency is assessed as actively improving only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s009", "s012", "s022", "s023", "s024" ], "note": "Report branch classification: cross-cutting." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Repeated network evaluations impede real-time, video, edge and control use.", "sourceIds": [ "s009", "s012", "s022", "s023", "s024" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "constraint_on", "note": "Iterative sampling latency constrains diffusion deployment." } ] }, { "id": "k02", "type": "constraint", "title": "Training/inference compute, memory and energy", "summary": "High resolution and temporal context make training/inference expensive.", "year": 2021, "yearEnd": 2026, "yearPrecision": "range", "domain": "media/world", "tags": [ "constraint", "media/world" ], "bandId": "band-constraints", "status": { "state": "constraint_bound", "scope": "Training/inference compute, memory and energy is assessed as constraint-bound only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s015", "s020", "s027", "s038" ], "note": "Report branch classification: media/world." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "High resolution and temporal context make training/inference expensive.", "sourceIds": [ "s015", "s020", "s027", "s038" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "video", "relation": "constraint_on", "note": "Compute and memory are especially binding for video/world scale." } ] }, { "id": "k03", "type": "constraint", "title": "Data provenance, memorization and privacy", "summary": "Large diffusion models can reproduce examples from training sets.", "year": 2022, "yearEnd": 2026, "yearPrecision": "range", "domain": "media/synthetic", "tags": [ "constraint", "media/synthetic" ], "bandId": "band-constraints", "status": { "state": "constraint_bound", "scope": "Data provenance, memorization and privacy is assessed as constraint-bound only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s068" ], "note": "Report branch classification: media/synthetic." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Large diffusion models can reproduce examples from training sets.", "sourceIds": [ "s068" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "constraint_on", "note": "Memorization and provenance constrain diffusion use." } ] }, { "id": "k04", "type": "constraint", "title": "Controllability and identity/spatiotemporal consistency", "summary": "Prompt adherence improved faster than exact spatial relations and long-term identity persistence.", "year": 2022, "yearEnd": 2026, "yearPrecision": "range", "domain": "image/video/3D", "tags": [ "constraint", "image/video/3d" ], "bandId": "band-constraints", "status": { "state": "actively_improving", "scope": "Controllability and identity/spatiotemporal consistency is assessed as actively improving only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s018", "s037", "s051" ], "note": "Report branch classification: image/video/3D." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Prompt adherence improved faster than exact spatial relations and long-term identity persistence.", "sourceIds": [ "s018", "s037", "s051" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "video", "relation": "constraint_on", "note": "Identity and spatiotemporal consistency constrain video generation." } ] }, { "id": "k05", "type": "constraint", "title": "Physical, causal and hard-constraint validity", "summary": "Plausible generations can violate measurements, geometry, chemistry, kinematics or physics.", "year": 2022, "yearEnd": 2026, "yearPrecision": "range", "domain": "inverse/video/3D/science/robotics", "tags": [ "constraint", "inverse/video/3d/science/robotics" ], "bandId": "band-constraints", "status": { "state": "constraint_bound", "scope": "Physical, causal and hard-constraint validity is assessed as constraint-bound only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s032", "s052", "s053", "s079" ], "note": "Report branch classification: inverse/video/3D/science/robotics." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Plausible generations can violate measurements, geometry, chemistry, kinematics or physics.", "sourceIds": [ "s032", "s052", "s053", "s079" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "materials", "relation": "constraint_on", "note": "Physical and hard constraints limit scientific-design validity." } ] }, { "id": "k06", "type": "constraint", "title": "Distribution shift, hallucination and calibration", "summary": "Strong priors can invent plausible detail when observational evidence is insufficient.", "year": 2021, "yearEnd": 2026, "yearPrecision": "range", "domain": "inverse/robotics/world", "tags": [ "constraint", "inverse/robotics/world" ], "bandId": "band-constraints", "status": { "state": "constraint_bound", "scope": "Distribution shift, hallucination and calibration is assessed as constraint-bound only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s033", "s069", "s075", "s076" ], "note": "Report branch classification: inverse/robotics/world." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Strong priors can invent plausible detail when observational evidence is insufficient.", "sourceIds": [ "s033", "s069", "s075", "s076" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "medicine", "relation": "constraint_on", "note": "OOD hallucination and calibration constrain high-stakes medical use." } ] }, { "id": "k07", "type": "constraint", "title": "Competition from GAN, autoregressive and flow paradigms", "summary": "Diffusion is one generative formulation among increasingly strong substitutes.", "year": 2020, "yearEnd": 2026, "yearPrecision": "range", "domain": "cross-cutting", "tags": [ "constraint", "cross-cutting" ], "bandId": "band-constraints", "status": { "state": "evidence_mixed", "scope": "Competition from GAN, autoregressive and flow paradigms is assessed as evidence mixed only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s006", "s025", "s026", "s027", "s062", "s063" ], "note": "Report branch classification: cross-cutting." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Diffusion is one generative formulation among increasingly strong substitutes.", "sourceIds": [ "s006", "s025", "s026", "s027", "s062", "s063" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "flowmatching", "relation": "competitor_to", "note": "Flow matching is an active partial substitute in several contexts." } ] }, { "id": "k08", "type": "constraint", "title": "Evaluation and likelihood-perception mismatch", "summary": "Likelihood, FID, preference and downstream utility measure different things.", "year": 2020, "yearEnd": 2026, "yearPrecision": "range", "domain": "cross-cutting", "tags": [ "constraint", "cross-cutting" ], "bandId": "band-constraints", "status": { "state": "constraint_bound", "scope": "Evaluation and likelihood-perception mismatch is assessed as constraint-bound only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "partial", "sourceIds": [ "s009", "s010", "s020", "s069", "s079" ], "note": "Report branch classification: cross-cutting." }, "evidence": [ { "type": "primary_source", "grade": "partial", "claim": "Likelihood, FID, preference and downstream utility measure different things.", "sourceIds": [ "s009", "s010", "s020", "s069", "s079" ], "note": "Evidence confidence assigned by the supplied report: partial." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "constraint_on", "note": "Metric mismatch constrains defensible diffusion comparisons." } ] }, { "id": "f01", "type": "failed_or_stalled_attempt", "title": "Progressive diffusion compression", "summary": "DDPM's conceptual compression interpretation depended on impractical high-dimensional coding machinery.", "year": 2020, "yearPrecision": "year", "domain": "core", "tags": [ "failed-or-stalled-attempt", "core" ], "bandId": "band-outcomes", "status": { "state": "locally_saturated", "scope": "Progressive diffusion compression is assessed as locally saturated only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s009" ], "note": "Report branch classification: core." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "DDPM's conceptual compression interpretation depended on impractical high-dimensional coding machinery.", "sourceIds": [ "s009" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "outcome_of", "note": "This locally stalled compression interpretation arose from DDPM." } ] }, { "id": "f02", "type": "failed_or_stalled_attempt", "title": "Single-view score-distillation text-to-3D", "summary": "Opened text-to-3D without 3D training sets but stalled on slow optimization and view consistency.", "year": 2022, "yearEnd": 2024, "yearPrecision": "range", "domain": "3D", "tags": [ "failed-or-stalled-attempt", "3d" ], "bandId": "band-outcomes", "status": { "state": "reopened_by_a_new_complement", "scope": "Single-view score-distillation text-to-3D is assessed as reopened by a new complement only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "partial", "sourceIds": [ "s047", "s048", "s051" ], "note": "Report branch classification: 3D." }, "evidence": [ { "type": "primary_source", "grade": "partial", "claim": "Opened text-to-3D without 3D training sets but stalled on slow optimization and view consistency.", "sourceIds": [ "s047", "s048", "s051" ], "note": "Evidence confidence assigned by the supplied report: partial." } ], "atlasLinks": [ { "atlasNodeId": "neural3d", "relation": "related_to", "note": "Single-view score distillation is an early 2D-prior route into neural 3D." } ] }, { "id": "f03", "type": "failed_or_stalled_attempt", "title": "Diffusion docking superiority claim", "summary": "Later fair-comparison work challenged the strength of the conventional docking baselines.", "year": 2022, "yearEnd": 2024, "yearPrecision": "range", "domain": "science/docking", "tags": [ "failed-or-stalled-attempt", "science/docking" ], "bandId": "band-outcomes", "status": { "state": "evidence_mixed", "scope": "Diffusion docking superiority claim is assessed as evidence mixed only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s052", "s056" ], "note": "Report branch classification: science/docking." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Later fair-comparison work challenged the strength of the conventional docking baselines.", "sourceIds": [ "s052", "s056" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "proteinlm", "relation": "related_to", "note": "The contested docking result belongs to the scientific generative-design neighborhood." } ] }, { "id": "s01", "type": "competing_or_substitute_approach", "title": "Generative adversarial networks", "summary": "Dominant pre-diffusion high-fidelity image paradigm; still valuable where one-pass latency matters.", "year": 2014, "yearEnd": 2026, "yearPrecision": "range", "domain": "cross-cutting", "tags": [ "competing-or-substitute-approach", "cross-cutting" ], "bandId": "band-outcomes", "status": { "state": "mature_but_useful", "scope": "Generative adversarial networks is assessed as mature but useful only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s006", "s009", "s013" ], "note": "Report branch classification: cross-cutting." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Dominant pre-diffusion high-fidelity image paradigm; still valuable where one-pass latency matters.", "sourceIds": [ "s006", "s009", "s013" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "gan", "relation": "same_as", "note": "Direct mapping to GANs as a competing paradigm." } ] }, { "id": "s02", "type": "competing_or_substitute_approach", "title": "Autoregressive generation", "summary": "Competes in language, audio, visual tokens and robot actions.", "year": 2016, "yearEnd": 2026, "yearPrecision": "range", "domain": "cross-cutting", "tags": [ "competing-or-substitute-approach", "cross-cutting" ], "bandId": "band-outcomes", "status": { "state": "actively_improving", "scope": "Autoregressive generation is assessed as actively improving only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s009", "s028", "s063" ], "note": "Report branch classification: cross-cutting." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Competes in language, audio, visual tokens and robot actions.", "sourceIds": [ "s009", "s028", "s063" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "gpt2", "relation": "related_to", "note": "Autoregressive generation is represented by the atlas language-model lineage." } ] }, { "id": "s03", "type": "competing_or_substitute_approach", "title": "Flow matching and rectified flow", "summary": "A mathematically adjacent continuous-transport paradigm and increasingly important partial substitute.", "year": 2022, "yearEnd": 2026, "yearPrecision": "range", "domain": "cross-cutting", "tags": [ "competing-or-substitute-approach", "cross-cutting" ], "bandId": "band-outcomes", "status": { "state": "rapidly_expanding", "scope": "Flow matching and rectified flow is assessed as rapidly expanding only within this bounded diffusion-model opportunity map, not as a global verdict on the field.", "evidenceGrade": "direct", "sourceIds": [ "s025", "s026", "s027", "s062" ], "note": "Report branch classification: cross-cutting." }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "A mathematically adjacent continuous-transport paradigm and increasingly important partial substitute.", "sourceIds": [ "s025", "s026", "s027", "s062" ], "note": "Evidence confidence assigned by the supplied report: direct." } ], "atlasLinks": [ { "atlasNodeId": "flowmatching", "relation": "same_as", "note": "Direct mapping to flow matching and rectified flow." } ] }, { "id": "opp01", "type": "open_opportunity", "title": "Measurement-consistent reconstruction with calibrated abstention", "summary": "Exact/near-exact data consistency plus uncertainty that detects prior-dominated reconstruction.", "yearPrecision": "undated", "domain": "inverse", "tags": [ "open-opportunity", "inverse" ], "bandId": "band-frontier", "status": { "state": "not_yet_assessed", "scope": "This is a falsifiable research hypothesis; inclusion is not a novelty, patentability, feasibility, or value claim.", "evidenceGrade": "unassessed", "sourceIds": [], "note": "Adjacent work is cited in the opportunity card, but the proposed residual opportunity remains unassessed." }, "evidence": [ { "type": "novelty_search", "grade": "hypothesis", "claim": "Exact/near-exact data consistency plus uncertainty that detects prior-dominated reconstruction.", "sourceIds": [ "s032", "s069", "s075", "s076" ], "note": "The sources establish adjacent work and unmet constraints; they do not establish novelty or expected success." } ], "atlasLinks": [ { "atlasNodeId": "diffusion", "relation": "opportunity_for", "note": "Hypothesis for reliable inverse reconstruction using diffusion/flow priors." } ] }, { "id": "opp02", "type": "open_opportunity", "title": "Persistent object-centric 3D world state", "summary": "Separate persistent world state from generative video rendering to improve long rollouts.", "yearPrecision": "undated", "domain": "world/video", "tags": [ "open-opportunity", "world/video" ], "bandId": "band-frontier", "status": { "state": "not_yet_assessed", "scope": "This is a falsifiable research hypothesis; inclusion is not a novelty, patentability, feasibility, or value claim.", "evidenceGrade": "unassessed", "sourceIds": [], "note": "Adjacent work is cited in the opportunity card, but the proposed residual opportunity remains unassessed." }, "evidence": [ { "type": "novelty_search", "grade": "hypothesis", "claim": "Separate persistent world state from generative video rendering to improve long rollouts.", "sourceIds": [ "s037", "s064", "s065", "s079", "pat02" ], "note": "The sources establish adjacent work and unmet constraints; they do not establish novelty or expected success." } ], "atlasLinks": [ { "atlasNodeId": "genie", "relation": "opportunity_for", "note": "Hypothesis for persistent state in interactive world models." } ] }, { "id": "opp03", "type": "open_opportunity", "title": "Path-consistent safety for multimodal robot policies", "summary": "Constrain trajectories without pushing a learned policy off its data manifold.", "yearPrecision": "undated", "domain": "robotics", "tags": [ "open-opportunity", "robotics" ], "bandId": "band-frontier", "status": { "state": "not_yet_assessed", "scope": "This is a falsifiable research hypothesis; inclusion is not a novelty, patentability, feasibility, or value claim.", "evidenceGrade": "unassessed", "sourceIds": [], "note": "Adjacent work is cited in the opportunity card, but the proposed residual opportunity remains unassessed." }, "evidence": [ { "type": "novelty_search", "grade": "hypothesis", "claim": "Constrain trajectories without pushing a learned policy off its data manifold.", "sourceIds": [ "s070", "s071", "s078" ], "note": "The sources establish adjacent work and unmet constraints; they do not establish novelty or expected success." } ], "atlasLinks": [ { "atlasNodeId": "aicontrol", "relation": "opportunity_for", "note": "Hypothesis for safety interventions in generative robot policies." } ] }, { "id": "opp04", "type": "open_opportunity", "title": "Joint material and executable synthesis-protocol generation", "summary": "Generate structure and synthesis procedure, then learn from laboratory outcomes.", "yearPrecision": "undated", "domain": "science", "tags": [ "open-opportunity", "science" ], "bandId": "band-frontier", "status": { "state": "not_yet_assessed", "scope": "This is a falsifiable research hypothesis; inclusion is not a novelty, patentability, feasibility, or value claim.", "evidenceGrade": "unassessed", "sourceIds": [], "note": "Adjacent work is cited in the opportunity card, but the proposed residual opportunity remains unassessed." }, "evidence": [ { "type": "novelty_search", "grade": "hypothesis", "claim": "Generate structure and synthesis procedure, then learn from laboratory outcomes.", "sourceIds": [ "s054", "s077" ], "note": "The sources establish adjacent work and unmet constraints; they do not establish novelty or expected success." } ], "atlasLinks": [ { "atlasNodeId": "materials", "relation": "opportunity_for", "note": "Hypothesis for closed-loop structure and synthesis-protocol generation." } ] }, { "id": "opp05", "type": "open_opportunity", "title": "Solver-certified manufacturable B-Rep generation", "summary": "Generate native CAD under explicit topology/tolerance/process constraints.", "yearPrecision": "undated", "domain": "3D/CAD", "tags": [ "open-opportunity", "3d/cad" ], "bandId": "band-frontier", "status": { "state": "not_yet_assessed", "scope": "This is a falsifiable research hypothesis; inclusion is not a novelty, patentability, feasibility, or value claim.", "evidenceGrade": "unassessed", "sourceIds": [], "note": "Adjacent work is cited in the opportunity card, but the proposed residual opportunity remains unassessed." }, "evidence": [ { "type": "novelty_search", "grade": "hypothesis", "claim": "Generate native CAD under explicit topology/tolerance/process constraints.", "sourceIds": [ "s073", "s074", "pat01" ], "note": "The sources establish adjacent work and unmet constraints; they do not establish novelty or expected success." } ], "atlasLinks": [ { "atlasNodeId": "neural3d", "relation": "opportunity_for", "note": "Hypothesis for solver-certified native CAD geometry." } ] }, { "id": "opp06", "type": "open_opportunity", "title": "OOD-aware medical inverse reconstruction", "summary": "Detect domain/operator shifts and warn when generated detail is weakly measurement-supported.", "yearPrecision": "undated", "domain": "inverse/medical", "tags": [ "open-opportunity", "inverse/medical" ], "bandId": "band-frontier", "status": { "state": "not_yet_assessed", "scope": "This is a falsifiable research hypothesis; inclusion is not a novelty, patentability, feasibility, or value claim.", "evidenceGrade": "unassessed", "sourceIds": [], "note": "Adjacent work is cited in the opportunity card, but the proposed residual opportunity remains unassessed." }, "evidence": [ { "type": "novelty_search", "grade": "hypothesis", "claim": "Detect domain/operator shifts and warn when generated detail is weakly measurement-supported.", "sourceIds": [ "s033", "s069", "s075", "s076" ], "note": "The sources establish adjacent work and unmet constraints; they do not establish novelty or expected success." } ], "atlasLinks": [ { "atlasNodeId": "medicine", "relation": "opportunity_for", "note": "Hypothesis for prospective OOD alarms in medical reconstruction." } ] }, { "id": "opp07", "type": "open_opportunity", "title": "Causal audiovisual event-state generation", "summary": "Shared event/physical state jointly controls video and synchronized sound.", "yearPrecision": "undated", "domain": "video/audio/world", "tags": [ "open-opportunity", "video/audio/world" ], "bandId": "band-frontier", "status": { "state": "not_yet_assessed", "scope": "This is a falsifiable research hypothesis; inclusion is not a novelty, patentability, feasibility, or value claim.", "evidenceGrade": "unassessed", "sourceIds": [], "note": "Adjacent work is cited in the opportunity card, but the proposed residual opportunity remains unassessed." }, "evidence": [ { "type": "novelty_search", "grade": "hypothesis", "claim": "Shared event/physical state jointly controls video and synchronized sound.", "sourceIds": [ "s037", "s044" ], "note": "The sources establish adjacent work and unmet constraints; they do not establish novelty or expected success." } ], "atlasLinks": [ { "atlasNodeId": "gap_causal", "relation": "related_to", "note": "Hypothesis uses explicit causal event state for audiovisual generation." } ] }, { "id": "opp08", "type": "open_opportunity", "title": "Audited synthetic data", "summary": "Optimize downstream utility while screening memorization, coverage and real-domain gap.", "yearPrecision": "undated", "domain": "synthetic data", "tags": [ "open-opportunity", "synthetic data" ], "bandId": "band-frontier", "status": { "state": "not_yet_assessed", "scope": "This is a falsifiable research hypothesis; inclusion is not a novelty, patentability, feasibility, or value claim.", "evidenceGrade": "unassessed", "sourceIds": [], "note": "Adjacent work is cited in the opportunity card, but the proposed residual opportunity remains unassessed." }, "evidence": [ { "type": "novelty_search", "grade": "hypothesis", "claim": "Optimize downstream utility while screening memorization, coverage and real-domain gap.", "sourceIds": [ "s066", "s067", "s068" ], "note": "The sources establish adjacent work and unmet constraints; they do not establish novelty or expected success." } ], "atlasLinks": [ { "atlasNodeId": "datacentric", "relation": "opportunity_for", "note": "Hypothesis for audited, utility-tested synthetic data." } ] } ], "relationships": [ { "id": "r001", "type": "documented_historical_influence", "sourceNodeId": "p01", "targetNodeId": "c01", "summary": "C01 explicitly says its Markov transport idea is used in nonequilibrium physics/AIS.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "C01 explicitly says its Markov transport idea is used in nonequilibrium physics/AIS.", "sourceIds": [ "s001", "s002", "s007" ], "note": "Direction: precursor → model. Conflicting evidence or qualification: none. Review note: Explicit citation." }, "atlasLinks": [] }, { "id": "r002", "type": "documented_historical_influence", "sourceNodeId": "p02", "targetNodeId": "c01", "summary": "C01 explicitly invokes Langevin/Fokker–Planck and forward/backward diffusion results.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "C01 explicitly invokes Langevin/Fokker–Planck and forward/backward diffusion results.", "sourceIds": [ "s007" ], "note": "Direction: stochastic process → model. Conflicting evidence or qualification: none. Review note: Explicit discussion." }, "atlasLinks": [] }, { "id": "r003", "type": "adapts_formalism_from", "sourceNodeId": "p05", "targetNodeId": "c01", "summary": "C01's lower bound resembles contemporary variational objectives while motivation/model form differ.", "status": "partially_supported", "evidence": { "type": "primary_source", "grade": "partial", "claim": "C01's lower bound resembles contemporary variational objectives while motivation/model form differ.", "sourceIds": [ "s005", "s007" ], "note": "Direction: objective family → training. Conflicting evidence or qualification: C01 rejects variational Bayes as main motivation. Review note: Do not draw VAE as sole ancestor." }, "atlasLinks": [] }, { "id": "r004", "type": "derives_from", "sourceNodeId": "p03", "targetNodeId": "c02", "summary": "NCSN estimates data scores using score-matching objectives.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "NCSN estimates data scores using score-matching objectives.", "sourceIds": [ "s003", "s008" ], "note": "Direction: score matching → NCSN. Conflicting evidence or qualification: none. Review note: Direct method lineage." }, "atlasLinks": [] }, { "id": "r005", "type": "later_mathematical_equivalence", "sourceNodeId": "p04", "targetNodeId": "c03", "summary": "DDPM epsilon prediction is a weighted denoising-score objective.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "DDPM epsilon prediction is a weighted denoising-score objective.", "sourceIds": [ "s004", "s009" ], "note": "Direction: DSM → DDPM interpretation. Conflicting evidence or qualification: none. Review note: DDPM explicitly cites Vincent." }, "atlasLinks": [] }, { "id": "r006", "type": "derives_from", "sourceNodeId": "p02", "targetNodeId": "c02", "summary": "NCSN samples with annealed Langevin dynamics.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "NCSN samples with annealed Langevin dynamics.", "sourceIds": [ "s008" ], "note": "Direction: Langevin → sampler. Conflicting evidence or qualification: none. Review note: Explicit." }, "atlasLinks": [] }, { "id": "r007", "type": "derives_from", "sourceNodeId": "c01", "targetNodeId": "c03", "summary": "DDPM explicitly builds on diffusion probabilistic models and reuses the variational formulation.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "DDPM explicitly builds on diffusion probabilistic models and reuses the variational formulation.", "sourceIds": [ "s007", "s009" ], "note": "Direction: 2015 DPM → DDPM. Conflicting evidence or qualification: none. Review note: Explicit citation." }, "atlasLinks": [] }, { "id": "r008", "type": "adapts_formalism_from", "sourceNodeId": "c02", "targetNodeId": "c03", "summary": "DDPM connects its parameterization to multiscale denoising score matching and NCSN-like sampling.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "DDPM connects its parameterization to multiscale denoising score matching and NCSN-like sampling.", "sourceIds": [ "s008", "s009" ], "note": "Direction: score model → interpretation. Conflicting evidence or qualification: DDPM derives sampler coefficients differently. Review note: Cross-fertilization, not simple inheritance." }, "atlasLinks": [] }, { "id": "r009", "type": "adapts_formalism_from", "sourceNodeId": "p05", "targetNodeId": "c03", "summary": "DDPM trains the reverse chain as a variational latent-variable model.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "DDPM trains the reverse chain as a variational latent-variable model.", "sourceIds": [ "s009" ], "note": "Direction: VI → training. Conflicting evidence or qualification: none. Review note: Formal influence." }, "atlasLinks": [] }, { "id": "r010", "type": "derives_from", "sourceNodeId": "c02", "targetNodeId": "c04", "summary": "Score-SDE generalizes multiscale score generation to continuous time.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Score-SDE generalizes multiscale score generation to continuous time.", "sourceIds": [ "s008", "s010" ], "note": "Direction: score models → SDE. Conflicting evidence or qualification: none. Review note: Explicit unification." }, "atlasLinks": [] }, { "id": "r011", "type": "later_mathematical_equivalence", "sourceNodeId": "c03", "targetNodeId": "c04", "summary": "Variance-preserving diffusion fits inside the score-SDE family with reverse SDE/probability-flow descriptions.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Variance-preserving diffusion fits inside the score-SDE family with reverse SDE/probability-flow descriptions.", "sourceIds": [ "s009", "s010" ], "note": "Direction: discrete diffusion → continuous reinterpretation. Conflicting evidence or qualification: none. Review note: Later mathematics." }, "atlasLinks": [] }, { "id": "r012", "type": "enables", "sourceNodeId": "c01", "targetNodeId": "cap01", "summary": "Many small reverse transitions turn density learning into local prediction.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Many small reverse transitions turn density learning into local prediction.", "sourceIds": [ "s007" ], "note": "Direction: model → capability. Conflicting evidence or qualification: none. Review note: Original mechanism." }, "atlasLinks": [] }, { "id": "r013", "type": "improves", "sourceNodeId": "c03", "targetNodeId": "cap01", "summary": "Noise prediction simplifies learning and delivered high-quality images.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Noise prediction simplifies learning and delivered high-quality images.", "sourceIds": [ "s009" ], "note": "Direction: DDPM → capability. Conflicting evidence or qualification: none. Review note: Practical unlock." }, "atlasLinks": [] }, { "id": "r014", "type": "enables", "sourceNodeId": "c02", "targetNodeId": "cap02", "summary": "Scores over noise levels support coarse-to-fine generation.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Scores over noise levels support coarse-to-fine generation.", "sourceIds": [ "s008" ], "note": "Direction: model → capability. Conflicting evidence or qualification: none. Review note: Direct." }, "atlasLinks": [] }, { "id": "r015", "type": "improves", "sourceNodeId": "c04", "targetNodeId": "cap02", "summary": "Continuous time generalizes schedules and numerical samplers.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Continuous time generalizes schedules and numerical samplers.", "sourceIds": [ "s010" ], "note": "Direction: SDE → capability. Conflicting evidence or qualification: none. Review note: Direct." }, "atlasLinks": [] }, { "id": "r016", "type": "enables", "sourceNodeId": "c01", "targetNodeId": "cap04", "summary": "C01 explicitly multiplies a learned distribution by another factor for posterior tasks.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "C01 explicitly multiplies a learned distribution by another factor for posterior tasks.", "sourceIds": [ "s007" ], "note": "Direction: model → inverse-prior capability. Conflicting evidence or qualification: none. Review note: Early proof-of-concept." }, "atlasLinks": [] }, { "id": "r017", "type": "improves", "sourceNodeId": "c04", "targetNodeId": "cap04", "summary": "SDE machinery generalizes inverse-problem sampling.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "SDE machinery generalizes inverse-problem sampling.", "sourceIds": [ "s010" ], "note": "Direction: SDE → inverse capability. Conflicting evidence or qualification: none. Review note: Direct." }, "atlasLinks": [] }, { "id": "r018", "type": "enables", "sourceNodeId": "c03", "targetNodeId": "r01", "summary": "Improved DDPM changes the objective and learns reverse variances.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Improved DDPM changes the objective and learns reverse variances.", "sourceIds": [ "s011" ], "note": "Direction: base model → refinement. Conflicting evidence or qualification: none. Review note: Extension." }, "atlasLinks": [] }, { "id": "r019", "type": "mitigates_constraint", "sourceNodeId": "r01", "targetNodeId": "k01", "summary": "Learned variances reduce required passes substantially in reported experiments.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Learned variances reduce required passes substantially in reported experiments.", "sourceIds": [ "s011" ], "note": "Direction: refinement → latency. Conflicting evidence or qualification: setting-dependent. Review note: Mitigation, not elimination." }, "atlasLinks": [] }, { "id": "r020", "type": "enables", "sourceNodeId": "c03", "targetNodeId": "r02", "summary": "DDIM/DPM-Solver use pretrained diffusion/score models with alternative numerical paths.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "DDIM/DPM-Solver use pretrained diffusion/score models with alternative numerical paths.", "sourceIds": [ "s012", "s022" ], "note": "Direction: model → sampler. Conflicting evidence or qualification: none. Review note: Direct." }, "atlasLinks": [] }, { "id": "r021", "type": "mitigates_constraint", "sourceNodeId": "r02", "targetNodeId": "k01", "summary": "Sampling falls from hundreds/thousands of evaluations toward tens.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Sampling falls from hundreds/thousands of evaluations toward tens.", "sourceIds": [ "s012", "s022" ], "note": "Direction: solver → latency. Conflicting evidence or qualification: quality/model dependent. Review note: Strong local mitigation." }, "atlasLinks": [] }, { "id": "r022", "type": "enables", "sourceNodeId": "c03", "targetNodeId": "r03", "summary": "Guidance modifies DDPM-like score predictions.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Guidance modifies DDPM-like score predictions.", "sourceIds": [ "s013", "s014" ], "note": "Direction: base model → guidance. Conflicting evidence or qualification: none. Review note: Extension." }, "atlasLinks": [] }, { "id": "r023", "type": "enables", "sourceNodeId": "r03", "targetNodeId": "cap03", "summary": "Guidance supplies tunable condition adherence; CFG removes a separate classifier.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Guidance supplies tunable condition adherence; CFG removes a separate classifier.", "sourceIds": [ "s013", "s014" ], "note": "Direction: guidance → conditional capability. Conflicting evidence or qualification: diversity can fall. Review note: Core conditioning channel." }, "atlasLinks": [] }, { "id": "r024", "type": "applied_to", "sourceNodeId": "r03", "targetNodeId": "a01", "summary": "Guidance improves class/text-conditioned image generation.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Guidance improves class/text-conditioned image generation.", "sourceIds": [ "s013", "s017" ], "note": "Direction: guidance → images. Conflicting evidence or qualification: protocols vary. Review note: Application edge." }, "atlasLinks": [] }, { "id": "r025", "type": "applied_to", "sourceNodeId": "r04", "targetNodeId": "a01", "summary": "Cascades provide high-resolution image generation.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Cascades provide high-resolution image generation.", "sourceIds": [ "s016" ], "note": "Direction: cascade → image. Conflicting evidence or qualification: compute confound. Review note: Direct." }, "atlasLinks": [] }, { "id": "r026", "type": "applied_to", "sourceNodeId": "r04", "targetNodeId": "a03", "summary": "Spatial/temporal super-resolution cascades scale video.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Spatial/temporal super-resolution cascades scale video.", "sourceIds": [ "s036" ], "note": "Direction: cascade → video. Conflicting evidence or qualification: system comparisons differ. Review note: Direct." }, "atlasLinks": [] }, { "id": "r027", "type": "enables", "sourceNodeId": "c03", "targetNodeId": "r05", "summary": "Latent diffusion retains denoising while changing representation.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Latent diffusion retains denoising while changing representation.", "sourceIds": [ "s015" ], "note": "Direction: DDPM family → LDM. Conflicting evidence or qualification: none. Review note: Direct." }, "atlasLinks": [] }, { "id": "r028", "type": "combines_with", "sourceNodeId": "m03", "targetNodeId": "r05", "summary": "LDM depends on a pretrained perceptual autoencoder.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "LDM depends on a pretrained perceptual autoencoder.", "sourceIds": [ "s015" ], "note": "Direction: latent complement → LDM. Conflicting evidence or qualification: compression can lose details. Review note: Critical complement." }, "atlasLinks": [] }, { "id": "r029", "type": "enables", "sourceNodeId": "r05", "targetNodeId": "cap05", "summary": "Latent denoising reduces spatial compute enough for high-res cross-attention synthesis.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Latent denoising reduces spatial compute enough for high-res cross-attention synthesis.", "sourceIds": [ "s015" ], "note": "Direction: LDM → scalable synthesis. Conflicting evidence or qualification: compression-factor dependent. Review note: Material practical unlock." }, "atlasLinks": [] }, { "id": "r030", "type": "combines_with", "sourceNodeId": "m01", "targetNodeId": "c03", "summary": "DDPM uses a U-Net-like multiscale denoiser.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "DDPM uses a U-Net-like multiscale denoiser.", "sourceIds": [ "s009" ], "note": "Direction: architecture → model. Conflicting evidence or qualification: none. Review note: Complement, not origin." }, "atlasLinks": [] }, { "id": "r031", "type": "combines_with", "sourceNodeId": "m02", "targetNodeId": "cap03", "summary": "Language encoders/paired data make semantic open-vocabulary conditions useful.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Language encoders/paired data make semantic open-vocabulary conditions useful.", "sourceIds": [ "s016", "s017" ], "note": "Direction: encoder/data → conditioning. Conflicting evidence or qualification: data-scale confounding. Review note: Strong ablation support in Imagen." }, "atlasLinks": [] }, { "id": "r032", "type": "combines_with", "sourceNodeId": "m04", "targetNodeId": "r06", "summary": "DiT/rectified-flow systems pair Transformers with accelerator scale.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "DiT/rectified-flow systems pair Transformers with accelerator scale.", "sourceIds": [ "s020", "s027" ], "note": "Direction: architecture/compute → refinement. Conflicting evidence or qualification: model/data differ. Review note: Scaling complement." }, "atlasLinks": [] }, { "id": "r033", "type": "improves", "sourceNodeId": "r06", "targetNodeId": "cap05", "summary": "Transformer scaling improves image metrics and text handling.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Transformer scaling improves image metrics and text handling.", "sourceIds": [ "s020", "s027" ], "note": "Direction: DiT → scalable synthesis. Conflicting evidence or qualification: not architecture-only comparisons. Review note: Active path." }, "atlasLinks": [] }, { "id": "r034", "type": "mitigates_constraint", "sourceNodeId": "r07", "targetNodeId": "k01", "summary": "Distillation/consistency reduce inference to few/one model evaluations.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Distillation/consistency reduce inference to few/one model evaluations.", "sourceIds": [ "s023", "s024" ], "note": "Direction: refinement → latency. Conflicting evidence or qualification: quality tradeoff. Review note: Reopens low-latency use." }, "atlasLinks": [] }, { "id": "r035", "type": "derives_from", "sourceNodeId": "c03", "targetNodeId": "r08", "summary": "D3PM generalizes DDPM corruption/reversal to structured discrete states.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "D3PM generalizes DDPM corruption/reversal to structured discrete states.", "sourceIds": [ "s009", "s028" ], "note": "Direction: continuous → discrete. Conflicting evidence or qualification: transitions differ materially. Review note: Direct extension." }, "atlasLinks": [] }, { "id": "r036", "type": "competes_with", "sourceNodeId": "r08", "targetNodeId": "s02", "summary": "Masked/discrete diffusion generates tokens iteratively instead of left-to-right.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Masked/discrete diffusion generates tokens iteratively instead of left-to-right.", "sourceIds": [ "s028", "s029" ], "note": "Direction: diffusion LM ↔ AR. Conflicting evidence or qualification: AR remains dominant in many workloads. Review note: Active competition." }, "atlasLinks": [] }, { "id": "r037", "type": "blocked_by", "sourceNodeId": "r08", "targetNodeId": "k01", "summary": "Iterative token refinement introduces latency and lacks straightforward standard AR KV-caching behavior.", "status": "partially_supported", "evidence": { "type": "primary_source", "grade": "partial", "claim": "Iterative token refinement introduces latency and lacks straightforward standard AR KV-caching behavior.", "sourceIds": [ "s029" ], "note": "Direction: discrete diffusion → latency. Conflicting evidence or qualification: acceleration work is rapid. Review note: Reassess frequently." }, "atlasLinks": [] }, { "id": "r038", "type": "competes_with", "sourceNodeId": "s03", "targetNodeId": "c03", "summary": "Flow matching/rectified flow learn continuous transport without requiring the classical stochastic reverse-diffusion objective.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Flow matching/rectified flow learn continuous transport without requiring the classical stochastic reverse-diffusion objective.", "sourceIds": [ "s025", "s026", "s027" ], "note": "Direction: flow ↔ DDPM. Conflicting evidence or qualification: terminology overlaps. Review note: Partial substitute." }, "atlasLinks": [] }, { "id": "r039", "type": "displaced_in_context_by", "sourceNodeId": "s03", "targetNodeId": "a01", "summary": "Some frontier text-to-image systems replace established diffusion objectives with rectified flow.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Some frontier text-to-image systems replace established diffusion objectives with rectified flow.", "sourceIds": [ "s027" ], "note": "Direction: classical diffusion → RF in context. Conflicting evidence or qualification: diffusion remains widespread. Review note: Local displacement only." }, "atlasLinks": [] }, { "id": "r040", "type": "applied_to", "sourceNodeId": "cap05", "targetNodeId": "a01", "summary": "Scalable conditional synthesis powers text-to-image and editing.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Scalable conditional synthesis powers text-to-image and editing.", "sourceIds": [ "s015", "s016", "s017" ], "note": "Direction: capability → application. Conflicting evidence or qualification: none. Review note: Canonical branch." }, "atlasLinks": [] }, { "id": "r041", "type": "applied_to", "sourceNodeId": "cap03", "targetNodeId": "a01", "summary": "Conditional guidance supports prompt and structural control.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Conditional guidance supports prompt and structural control.", "sourceIds": [ "s014", "s018", "s019" ], "note": "Direction: capability → application. Conflicting evidence or qualification: imperfect composition. Review note: Direct." }, "atlasLinks": [] }, { "id": "r042", "type": "applied_to", "sourceNodeId": "cap04", "targetNodeId": "a02", "summary": "Diffusion/score priors combine with measurement operators for reconstruction.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Diffusion/score priors combine with measurement operators for reconstruction.", "sourceIds": [ "s030", "s031", "s032", "s033" ], "note": "Direction: capability → inverse. Conflicting evidence or qualification: approximate posterior methods differ. Review note: Strong branch." }, "atlasLinks": [] }, { "id": "r043", "type": "blocked_by", "sourceNodeId": "a02", "targetNodeId": "k05", "summary": "Approximate posterior guidance can fail exact measurement/domain constraints.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Approximate posterior guidance can fail exact measurement/domain constraints.", "sourceIds": [ "s032", "s069" ], "note": "Direction: application → constraint. Conflicting evidence or qualification: some linear cases easier. Review note: Operator-dependent." }, "atlasLinks": [] }, { "id": "r044", "type": "blocked_by", "sourceNodeId": "a02", "targetNodeId": "k06", "summary": "Weak/shifted evidence can yield plausible unsupported reconstruction.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Weak/shifted evidence can yield plausible unsupported reconstruction.", "sourceIds": [ "s033", "s069", "s075" ], "note": "Direction: application → hallucination. Conflicting evidence or qualification: not diffusion-exclusive. Review note: Deployment bottleneck." }, "atlasLinks": [] }, { "id": "r045", "type": "applied_to", "sourceNodeId": "cap05", "targetNodeId": "a03", "summary": "High-resolution latent conditional generation extends to space-time representations.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "High-resolution latent conditional generation extends to space-time representations.", "sourceIds": [ "s035", "s036", "s037" ], "note": "Direction: capability → video. Conflicting evidence or qualification: temporal scaling cost. Review note: Direct." }, "atlasLinks": [] }, { "id": "r046", "type": "applied_to", "sourceNodeId": "r06", "targetNodeId": "a03", "summary": "Large video generators use transformerized space-time denoisers/velocity models.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Large video generators use transformerized space-time denoisers/velocity models.", "sourceIds": [ "s037", "s038" ], "note": "Direction: transformer refinement → video. Conflicting evidence or qualification: some use flow objectives. Review note: Architecture survives objective shift." }, "atlasLinks": [] }, { "id": "r047", "type": "blocked_by", "sourceNodeId": "a03", "targetNodeId": "k04", "summary": "Long videos remain limited by identity/object persistence.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Long videos remain limited by identity/object persistence.", "sourceIds": [ "s037" ], "note": "Direction: video → consistency. Conflicting evidence or qualification: rapidly improving. Review note: No saturation inference." }, "atlasLinks": [] }, { "id": "r048", "type": "blocked_by", "sourceNodeId": "a03", "targetNodeId": "k05", "summary": "Video can violate physical state transitions and causality.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Video can violate physical state transitions and causality.", "sourceIds": [ "s037", "s079" ], "note": "Direction: video → physics. Conflicting evidence or qualification: benchmarks immature. Review note: Major frontier." }, "atlasLinks": [] }, { "id": "r049", "type": "applied_to", "sourceNodeId": "cap01", "targetNodeId": "a04", "summary": "DiffWave/WaveGrad apply iterative denoising to waveform synthesis.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "DiffWave/WaveGrad apply iterative denoising to waveform synthesis.", "sourceIds": [ "s040", "s041" ], "note": "Direction: capability → audio. Conflicting evidence or qualification: none. Review note: Early branch." }, "atlasLinks": [] }, { "id": "r050", "type": "applied_to", "sourceNodeId": "r05", "targetNodeId": "a04", "summary": "AudioLDM applies latent diffusion to audio representations.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "AudioLDM applies latent diffusion to audio representations.", "sourceIds": [ "s043" ], "note": "Direction: latent diffusion → audio. Conflicting evidence or qualification: embedding quality matters. Review note: Practical scaling." }, "atlasLinks": [] }, { "id": "r051", "type": "blocked_by", "sourceNodeId": "a04", "targetNodeId": "k01", "summary": "Repeated denoising can be slower than low-step/AR alternatives.", "status": "partially_supported", "evidence": { "type": "primary_source", "grade": "partial", "claim": "Repeated denoising can be slower than low-step/AR alternatives.", "sourceIds": [ "s040", "s041" ], "note": "Direction: audio → latency. Conflicting evidence or qualification: few-step variants exist. Review note: Mature, not dead." }, "atlasLinks": [] }, { "id": "r052", "type": "applied_to", "sourceNodeId": "cap06", "targetNodeId": "a05", "summary": "Diffusion directly generates point clouds and other 3D representations.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Diffusion directly generates point clouds and other 3D representations.", "sourceIds": [ "s046", "s049", "s050" ], "note": "Direction: structured generation → 3D. Conflicting evidence or qualification: representation choice matters. Review note: Direct." }, "atlasLinks": [] }, { "id": "r053", "type": "combines_with", "sourceNodeId": "m05", "targetNodeId": "a05", "summary": "Differentiable rendering and multiview priors adapt image models to 3D.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Differentiable rendering and multiview priors adapt image models to 3D.", "sourceIds": [ "s047", "s051" ], "note": "Direction: geometry complement → 3D. Conflicting evidence or qualification: 2D prior does not guarantee 3D. Review note: Critical complement." }, "atlasLinks": [] }, { "id": "r054", "type": "reopened_by", "sourceNodeId": "f02", "targetNodeId": "m05", "summary": "Multiview priors/direct 3D representations mitigate single-view score-distillation failure modes.", "status": "partially_supported", "evidence": { "type": "primary_source", "grade": "partial", "claim": "Multiview priors/direct 3D representations mitigate single-view score-distillation failure modes.", "sourceIds": [ "s050", "s051" ], "note": "Direction: stalled channel → complement. Conflicting evidence or qualification: validity still unresolved. Review note: Reopened." }, "atlasLinks": [] }, { "id": "r055", "type": "blocked_by", "sourceNodeId": "a05", "targetNodeId": "k05", "summary": "Attractive renders do not imply topology/multiview/manufacturing validity.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Attractive renders do not imply topology/multiview/manufacturing validity.", "sourceIds": [ "s051", "s073", "s074" ], "note": "Direction: 3D → hard constraints. Conflicting evidence or qualification: CAD differs from free-form 3D. Review note: Representation-specific." }, "atlasLinks": [] }, { "id": "r056", "type": "applied_to", "sourceNodeId": "cap06", "targetNodeId": "a06", "summary": "Equivariant/manifold diffusion generates poses, proteins and crystals.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Equivariant/manifold diffusion generates poses, proteins and crystals.", "sourceIds": [ "s052", "s053", "s055" ], "note": "Direction: structured generation → science. Conflicting evidence or qualification: domain tests required. Review note: Direct." }, "atlasLinks": [] }, { "id": "r057", "type": "combines_with", "sourceNodeId": "m05", "targetNodeId": "a06", "summary": "Equivariance and geometric coordinates enforce relevant symmetries.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Equivariance and geometric coordinates enforce relevant symmetries.", "sourceIds": [ "s052", "s053", "s055" ], "note": "Direction: geometry → science. Conflicting evidence or qualification: implementations vary. Review note: Critical complement." }, "atlasLinks": [] }, { "id": "r058", "type": "combines_with", "sourceNodeId": "m06", "targetNodeId": "a06", "summary": "Predictors, structure filters and experiments validate generated candidates.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Predictors, structure filters and experiments validate generated candidates.", "sourceIds": [ "s053", "s054" ], "note": "Direction: validator → science. Conflicting evidence or qualification: validator quality can dominate. Review note: Generation alone insufficient." }, "atlasLinks": [] }, { "id": "r059", "type": "blocked_by", "sourceNodeId": "a06", "targetNodeId": "f03", "summary": "DiffDock's headline ranking changes under stronger conventional-docking comparisons.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "DiffDock's headline ranking changes under stronger conventional-docking comparisons.", "sourceIds": [ "s052", "s056" ], "note": "Direction: docking branch → contested result. Conflicting evidence or qualification: protocols differ. Review note: Do not generalize to all science diffusion." }, "atlasLinks": [] }, { "id": "r060", "type": "blocked_by", "sourceNodeId": "a06", "targetNodeId": "k05", "summary": "Geometric plausibility alone does not prove biochemical function, synthesizability or material stability.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Geometric plausibility alone does not prove biochemical function, synthesizability or material stability.", "sourceIds": [ "s053", "s054", "s055" ], "note": "Direction: science → validity. Conflicting evidence or qualification: wet-lab protein results are counterevidence. Review note: Constraint partly relaxed." }, "atlasLinks": [] }, { "id": "r061", "type": "applied_to", "sourceNodeId": "cap06", "targetNodeId": "a07", "summary": "Diffuser/Diffusion Policy generate trajectories or action chunks by denoising.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Diffuser/Diffusion Policy generate trajectories or action chunks by denoising.", "sourceIds": [ "s058", "s060" ], "note": "Direction: capability → robotics. Conflicting evidence or qualification: closed-loop constraints harder. Review note: Direct." }, "atlasLinks": [] }, { "id": "r062", "type": "combines_with", "sourceNodeId": "m06", "targetNodeId": "a07", "summary": "Robot demonstrations, 3D observations and simulators are core supervision.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Robot demonstrations, 3D observations and simulators are core supervision.", "sourceIds": [ "s059", "s060", "s061" ], "note": "Direction: complement → robotics. Conflicting evidence or qualification: coverage matters. Review note: Critical." }, "atlasLinks": [] }, { "id": "r063", "type": "blocked_by", "sourceNodeId": "a07", "targetNodeId": "k01", "summary": "Multiple denoising steps per action chunk create control latency.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Multiple denoising steps per action chunk create control latency.", "sourceIds": [ "s060", "s063" ], "note": "Direction: robotics → latency. Conflicting evidence or qualification: flow/fast samplers help. Review note: Important substitution pressure." }, "atlasLinks": [] }, { "id": "r064", "type": "displaced_in_context_by", "sourceNodeId": "s03", "targetNodeId": "a07", "summary": "π0 uses flow matching instead of classical diffusion for continuous action generation.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "π0 uses flow matching instead of classical diffusion for continuous action generation.", "sourceIds": [ "s062" ], "note": "Direction: diffusion action head → flow in context. Conflicting evidence or qualification: Diffusion Policy remains active. Review note: Local displacement." }, "atlasLinks": [] }, { "id": "r065", "type": "competes_with", "sourceNodeId": "s02", "targetNodeId": "a07", "summary": "FAST reports competitive AR action tokenization with faster training in its setup.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "FAST reports competitive AR action tokenization with faster training in its setup.", "sourceIds": [ "s063" ], "note": "Direction: AR ↔ diffusion/VLA. Conflicting evidence or qualification: setups not universal. Review note: Concrete pressure." }, "atlasLinks": [] }, { "id": "r066", "type": "applied_to", "sourceNodeId": "cap05", "targetNodeId": "a08", "summary": "Conditional generation supplies synthetic training images/video.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Conditional generation supplies synthetic training images/video.", "sourceIds": [ "s064", "s066", "s067" ], "note": "Direction: capability → synthetic data. Conflicting evidence or qualification: real-domain gap. Review note: Direct." }, "atlasLinks": [] }, { "id": "r067", "type": "applied_to", "sourceNodeId": "cap06", "targetNodeId": "a08", "summary": "Action-conditioned sequence generation supplies visual dynamics/world simulation.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Action-conditioned sequence generation supplies visual dynamics/world simulation.", "sourceIds": [ "s064", "s065" ], "note": "Direction: capability → world models. Conflicting evidence or qualification: generalization uncertain. Review note: Direct." }, "atlasLinks": [] }, { "id": "r068", "type": "blocked_by", "sourceNodeId": "a08", "targetNodeId": "k05", "summary": "World models need physics/causal fidelity beyond attractive frames.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "World models need physics/causal fidelity beyond attractive frames.", "sourceIds": [ "s079" ], "note": "Direction: world model → physics. Conflicting evidence or qualification: evaluation evolving. Review note: Core bottleneck." }, "atlasLinks": [] }, { "id": "r069", "type": "blocked_by", "sourceNodeId": "a08", "targetNodeId": "k06", "summary": "Synthetic/world models may overfit source environments and fail under novel conditions.", "status": "partially_supported", "evidence": { "type": "primary_source", "grade": "partial", "claim": "Synthetic/world models may overfit source environments and fail under novel conditions.", "sourceIds": [ "s065", "s066" ], "note": "Direction: application → shift. Conflicting evidence or qualification: mixed real/synthetic may help. Review note: Evidence mixed." }, "atlasLinks": [] }, { "id": "r070", "type": "blocked_by", "sourceNodeId": "a01", "targetNodeId": "k03", "summary": "Image diffusion models can memorize and emit training examples.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Image diffusion models can memorize and emit training examples.", "sourceIds": [ "s068" ], "note": "Direction: image → privacy. Conflicting evidence or qualification: duplication/data choices matter. Review note: Direct extraction evidence." }, "atlasLinks": [] }, { "id": "r071", "type": "blocked_by", "sourceNodeId": "a03", "targetNodeId": "k02", "summary": "Space-time token count and long horizons multiply compute.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Space-time token count and long horizons multiply compute.", "sourceIds": [ "s036", "s038", "s039" ], "note": "Direction: video → compute. Conflicting evidence or qualification: latent/patch methods mitigate. Review note: Persistent." }, "atlasLinks": [] }, { "id": "r072", "type": "blocked_by", "sourceNodeId": "a01", "targetNodeId": "k07", "summary": "Strong rectified-flow substitutes reduce returns to diffusion-only objective work.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Strong rectified-flow substitutes reduce returns to diffusion-only objective work.", "sourceIds": [ "s027" ], "note": "Direction: image → competition. Conflicting evidence or qualification: diffusion still active. Review note: No global displacement." }, "atlasLinks": [] }, { "id": "r073", "type": "blocked_by", "sourceNodeId": "a01", "targetNodeId": "k08", "summary": "FID, prompt alignment and human preference need not rank models identically.", "status": "partially_supported", "evidence": { "type": "primary_source", "grade": "partial", "claim": "FID, prompt alignment and human preference need not rank models identically.", "sourceIds": [ "s009", "s020", "s027" ], "note": "Direction: image → evaluation. Conflicting evidence or qualification: human evaluation has variance. Review note: Saturation claims fragile." }, "atlasLinks": [] }, { "id": "r074", "type": "blocked_by", "sourceNodeId": "a06", "targetNodeId": "k08", "summary": "Scientific validity can reverse rankings implied by generative metrics.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "Scientific validity can reverse rankings implied by generative metrics.", "sourceIds": [ "s053", "s056" ], "note": "Direction: science → evaluation. Conflicting evidence or qualification: experiments expensive. Review note: Domain evaluation mandatory." }, "atlasLinks": [] }, { "id": "r075", "type": "applied_to", "sourceNodeId": "c03", "targetNodeId": "f01", "summary": "DDPM's progressive compression interpretation was acknowledged as impractical with the proposed coding machinery.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "DDPM's progressive compression interpretation was acknowledged as impractical with the proposed coding machinery.", "sourceIds": [ "s009" ], "note": "Direction: DDPM → stalled branch. Conflicting evidence or qualification: later compression work possible. Review note: Explicit negative result." }, "atlasLinks": [] }, { "id": "r076", "type": "competes_with", "sourceNodeId": "s01", "targetNodeId": "c03", "summary": "DDPM/classifier-guided diffusion were benchmarked against GANs.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "DDPM/classifier-guided diffusion were benchmarked against GANs.", "sourceIds": [ "s009", "s013" ], "note": "Direction: GAN ↔ diffusion. Conflicting evidence or qualification: GAN latency advantage. Review note: Competition, not ancestry." }, "atlasLinks": [] }, { "id": "r077", "type": "retrospective_analogy", "sourceNodeId": "s02", "targetNodeId": "c03", "summary": "DDPM shows a special masking diffusion can be interpreted as AR decoding.", "status": "demonstrated", "evidence": { "type": "primary_source", "grade": "direct", "claim": "DDPM shows a special masking diffusion can be interpreted as AR decoding.", "sourceIds": [ "s009" ], "note": "Direction: AR ↔ diffusion analogy. Conflicting evidence or qualification: not historical influence. Review note: Explicitly retrospective." }, "atlasLinks": [] }, { "id": "r078", "type": "adapts_formalism_from", "sourceNodeId": "c04", "targetNodeId": "s03", "summary": "Flow matching includes diffusion-like probability paths but changes the learned transport objective.", "status": "partially_supported", "evidence": { "type": "primary_source", "grade": "partial", "claim": "Flow matching includes diffusion-like probability paths but changes the learned transport objective.", "sourceIds": [ "s025", "s026" ], "note": "Direction: SDE landscape → flow neighbor. Conflicting evidence or qualification: terminology varies. Review note: Formal adjacency." }, "atlasLinks": [] }, { "id": "r079", "type": "candidate_application", "sourceNodeId": "k01", "targetNodeId": "opp03", "summary": "Low-latency control motivates safety mechanisms that avoid wholesale resampling.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "Low-latency control motivates safety mechanisms that avoid wholesale resampling.", "sourceIds": [ "s070", "s071" ], "note": "Direction: constraint → hypothesis. Conflicting evidence or qualification: direct adjacent work exists. Review note: Crowded." }, "atlasLinks": [] }, { "id": "r080", "type": "candidate_application", "sourceNodeId": "a02", "targetNodeId": "opp01", "summary": "Inverse problems allow direct testing of consistency plus calibrated uncertainty.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "Inverse problems allow direct testing of consistency plus calibrated uncertainty.", "sourceIds": [ "s032", "s069", "s075" ], "note": "Direction: application → hypothesis. Conflicting evidence or qualification: many posterior samplers exist. Review note: Narrow novelty only." }, "atlasLinks": [] }, { "id": "r081", "type": "candidate_application", "sourceNodeId": "k06", "targetNodeId": "opp01", "summary": "Hallucination motivates abstention when prior dominates evidence.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "Hallucination motivates abstention when prior dominates evidence.", "sourceIds": [ "s075", "s076" ], "note": "Direction: constraint → hypothesis. Conflicting evidence or qualification: some hallucination is information-theoretic. Review note: Falsifiable." }, "atlasLinks": [] }, { "id": "r082", "type": "candidate_application", "sourceNodeId": "a08", "targetNodeId": "opp02", "summary": "Long-horizon world models need persistent state beyond frame realism.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "Long-horizon world models need persistent state beyond frame realism.", "sourceIds": [ "s064", "s065", "s079" ], "note": "Direction: application → hypothesis. Conflicting evidence or qualification: object-centric models adjacent. Review note: Combination novelty unproven." }, "atlasLinks": [] }, { "id": "r083", "type": "combines_with", "sourceNodeId": "m05", "targetNodeId": "opp02", "summary": "Explicit 3D scene/object state is the proposed persistence complement.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "Explicit 3D scene/object state is the proposed persistence complement.", "sourceIds": [ "s051" ], "note": "Direction: complement → hypothesis. Conflicting evidence or qualification: interface difficult. Review note: Candidate confluence." }, "atlasLinks": [] }, { "id": "r084", "type": "candidate_application", "sourceNodeId": "a07", "targetNodeId": "opp03", "summary": "Robot generative policies could use path-consistent safety interventions.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "Robot generative policies could use path-consistent safety interventions.", "sourceIds": [ "s070", "s071", "s078" ], "note": "Direction: robotics → hypothesis. Conflicting evidence or qualification: direct 2026 work. Review note: Only narrow residual opportunity." }, "atlasLinks": [] }, { "id": "r085", "type": "candidate_application", "sourceNodeId": "a06", "targetNodeId": "opp04", "summary": "Materials generation can extend from structures to executable synthesis programs.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "Materials generation can extend from structures to executable synthesis programs.", "sourceIds": [ "s054", "s077" ], "note": "Direction: science → hypothesis. Conflicting evidence or qualification: autonomous labs already exist. Review note: Novel combination unproven." }, "atlasLinks": [] }, { "id": "r086", "type": "combines_with", "sourceNodeId": "m06", "targetNodeId": "opp04", "summary": "Laboratory robotics/measurement is needed for closed-loop updates.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "Laboratory robotics/measurement is needed for closed-loop updates.", "sourceIds": [ "s077" ], "note": "Direction: complement → hypothesis. Conflicting evidence or qualification: equipment heterogeneity. Review note: Necessary complement." }, "atlasLinks": [] }, { "id": "r087", "type": "candidate_application", "sourceNodeId": "a05", "targetNodeId": "opp05", "summary": "3D generation can target native CAD/B-Rep validated by kernels.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "3D generation can target native CAD/B-Rep validated by kernels.", "sourceIds": [ "s073", "s074", "pat01" ], "note": "Direction: 3D → hypothesis. Conflicting evidence or qualification: strong direct prior art. Review note: Narrow certification opportunity." }, "atlasLinks": [] }, { "id": "r088", "type": "candidate_application", "sourceNodeId": "k05", "targetNodeId": "opp05", "summary": "Hard topology/tolerance constraints motivate solver-in-loop generation.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "Hard topology/tolerance constraints motivate solver-in-loop generation.", "sourceIds": [ "s074", "pat01" ], "note": "Direction: constraint → hypothesis. Conflicting evidence or qualification: classical generative design already constrained. Review note: Novelty implementation-specific." }, "atlasLinks": [] }, { "id": "r089", "type": "candidate_application", "sourceNodeId": "a02", "targetNodeId": "opp06", "summary": "Medical reconstruction is a high-value test case for prospective OOD alarms.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "Medical reconstruction is a high-value test case for prospective OOD alarms.", "sourceIds": [ "s033", "s069", "s076" ], "note": "Direction: inverse → hypothesis. Conflicting evidence or qualification: active literature. Review note: Prospective validation differentiates." }, "atlasLinks": [] }, { "id": "r090", "type": "candidate_application", "sourceNodeId": "a03", "targetNodeId": "opp07", "summary": "Shared event state may improve audiovisual synchronization and causal timing.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "Shared event state may improve audiovisual synchronization and causal timing.", "sourceIds": [ "s037", "s044" ], "note": "Direction: video → hypothesis. Conflicting evidence or qualification: multimodal systems already exist. Review note: Benchmark needed." }, "atlasLinks": [] }, { "id": "r091", "type": "combines_with", "sourceNodeId": "a04", "targetNodeId": "opp07", "summary": "Audio generative models can consume the same event-state representation as video.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "Audio generative models can consume the same event-state representation as video.", "sourceIds": [ "s044" ], "note": "Direction: audio → hypothesis. Conflicting evidence or qualification: decoder need not be diffusion. Review note: Mechanism-neutral complement." }, "atlasLinks": [] }, { "id": "r092", "type": "candidate_application", "sourceNodeId": "a08", "targetNodeId": "opp08", "summary": "Synthetic data is valuable only when real-domain utility and privacy are jointly controlled.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "Synthetic data is valuable only when real-domain utility and privacy are jointly controlled.", "sourceIds": [ "s066", "s067", "s068" ], "note": "Direction: application → hypothesis. Conflicting evidence or qualification: crowded area. Review note: Multi-objective test." }, "atlasLinks": [] }, { "id": "r093", "type": "candidate_application", "sourceNodeId": "k03", "targetNodeId": "opp08", "summary": "Memorization motivates sample-level membership/influence auditing.", "status": "hypothesis", "evidence": { "type": "novelty_search", "grade": "hypothesis", "claim": "Memorization motivates sample-level membership/influence auditing.", "sourceIds": [ "s068" ], "note": "Direction: constraint → hypothesis. Conflicting evidence or qualification: audits can miss leakage. Review note: High-value but hard." }, "atlasLinks": [] }, { "id": "r094", "type": "competes_with", "sourceNodeId": "p06", "targetNodeId": "c03", "summary": "Energy-based models, neural generators, VAEs, flows, GANs, and autoregressive models formed the competing generative landscape around DDPM.", "status": "partially_supported", "evidence": { "type": "editorial_analysis", "grade": "contextual", "claim": "The broader pre-DDPM generative-model landscape is contextual and competitive rather than a single causal parent of diffusion.", "sourceIds": [ "s006", "s007", "s009" ], "note": "Editorial connectivity edge derived from the report’s P06 node description; it is not one of the report’s 93 proposed relationship-table rows and does not assert direct historical influence." }, "atlasLinks": [] } ], "sources": [ { "id": "s001", "type": "paper", "title": "Nonequilibrium Equality for Free Energy Differences", "authors": [ "Christopher Jarzynski" ], "year": 1997, "url": "https://doi.org/10.1103/PhysRevLett.78.2690", "doi": "10.1103/PhysRevLett.78.2690" }, { "id": "s002", "type": "paper", "title": "Annealed Importance Sampling", "authors": [ "Radford M. Neal" ], "year": 2001, "url": "https://doi.org/10.1023/A:1008923215028", "doi": "10.1023/A:1008923215028" }, { "id": "s003", "type": "paper", "title": "Estimation of Non-Normalized Statistical Models by Score Matching", "authors": [ "Aapo Hyvarinen" ], "year": 2005, "url": "https://jmlr.org/papers/v6/hyvarinen05a.html" }, { "id": "s004", "type": "paper", "title": "A Connection Between Score Matching and Denoising Autoencoders", "authors": [ "Pascal Vincent" ], "year": 2011, "url": "https://doi.org/10.1162/NECO_a_00142", "doi": "10.1162/NECO_a_00142" }, { "id": "s005", "type": "paper", "title": "Auto-Encoding Variational Bayes", "authors": [ "Diederik P. Kingma", "Max Welling" ], "year": 2013, "url": "https://arxiv.org/abs/1312.6114", "doi": "10.48550/arXiv.1312.6114", "arxivId": "1312.6114" }, { "id": "s006", "type": "paper", "title": "Generative Adversarial Nets", "authors": [ "Ian Goodfellow et al." ], "year": 2014, "url": "https://arxiv.org/abs/1406.2661", "doi": "10.48550/arXiv.1406.2661", "arxivId": "1406.2661" }, { "id": "s007", "type": "paper", "title": "Deep Unsupervised Learning using Nonequilibrium Thermodynamics", "authors": [ "Jascha Sohl-Dickstein", "Eric A. Weiss", "Niru Maheswaranathan", "Surya Ganguli" ], "year": 2015, "url": "https://arxiv.org/abs/1503.03585", "doi": "10.48550/arXiv.1503.03585", "arxivId": "1503.03585" }, { "id": "s008", "type": "paper", "title": "Generative Modeling by Estimating Gradients of the Data Distribution", "authors": [ "Yang Song", "Stefano Ermon" ], "year": 2019, "url": "https://arxiv.org/abs/1907.05600", "doi": "10.48550/arXiv.1907.05600", "arxivId": "1907.05600" }, { "id": "s009", "type": "paper", "title": "Denoising Diffusion Probabilistic Models", "authors": [ "Jonathan Ho", "Ajay Jain", "Pieter Abbeel" ], "year": 2020, "url": "https://arxiv.org/abs/2006.11239", "doi": "10.48550/arXiv.2006.11239", "arxivId": "2006.11239" }, { "id": "s010", "type": "paper", "title": "Score-Based Generative Modeling through Stochastic Differential Equations", "authors": [ "Yang Song", "Jascha Sohl-Dickstein", "Diederik P. Kingma", "Abhishek Kumar", "Stefano Ermon", "Ben Poole" ], "year": 2021, "url": "https://arxiv.org/abs/2011.13456", "doi": "10.48550/arXiv.2011.13456", "arxivId": "2011.13456" }, { "id": "s011", "type": "paper", "title": "Improved Denoising Diffusion Probabilistic Models", "authors": [ "Alex Nichol", "Prafulla Dhariwal" ], "year": 2021, "url": "https://arxiv.org/abs/2102.09672", "doi": "10.48550/arXiv.2102.09672", "arxivId": "2102.09672" }, { "id": "s012", "type": "paper", "title": "Denoising Diffusion Implicit Models", "authors": [ "Jiaming Song", "Chenlin Meng", "Stefano Ermon" ], "year": 2020, "url": "https://arxiv.org/abs/2010.02502", "doi": "10.48550/arXiv.2010.02502", "arxivId": "2010.02502" }, { "id": "s013", "type": "paper", "title": "Diffusion Models Beat GANs on Image Synthesis", "authors": [ "Prafulla Dhariwal", "Alex Nichol" ], "year": 2021, "url": "https://arxiv.org/abs/2105.05233", "doi": "10.48550/arXiv.2105.05233", "arxivId": "2105.05233" }, { "id": "s014", "type": "paper", "title": "Classifier-Free Diffusion Guidance", "authors": [ "Jonathan Ho", "Tim Salimans" ], "year": 2022, "url": "https://arxiv.org/abs/2207.12598", "doi": "10.48550/arXiv.2207.12598", "arxivId": "2207.12598" }, { "id": "s015", "type": "paper", "title": "High-Resolution Image Synthesis with Latent Diffusion Models", "authors": [ "Robin Rombach", "Andreas Blattmann", "Dominik Lorenz", "Patrick Esser", "Bjorn Ommer" ], "year": 2022, "url": "https://arxiv.org/abs/2112.10752", "doi": "10.48550/arXiv.2112.10752", "arxivId": "2112.10752" }, { "id": "s016", "type": "paper", "title": "Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding", "authors": [ "Chitwan Saharia et al." ], "year": 2022, "url": "https://arxiv.org/abs/2205.11487", "doi": "10.48550/arXiv.2205.11487", "arxivId": "2205.11487" }, { "id": "s017", "type": "paper", "title": "GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models", "authors": [ "Alex Nichol et al." ], "year": 2021, "url": "https://arxiv.org/abs/2112.10741", "doi": "10.48550/arXiv.2112.10741", "arxivId": "2112.10741" }, { "id": "s018", "type": "paper", "title": "Adding Conditional Control to Text-to-Image Diffusion Models", "authors": [ "Lvmin Zhang", "Anyi Rao", "Maneesh Agrawala" ], "year": 2023, "url": "https://arxiv.org/abs/2302.05543", "doi": "10.48550/arXiv.2302.05543", "arxivId": "2302.05543" }, { "id": "s019", "type": "paper", "title": "SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations", "authors": [ "Chenlin Meng et al." ], "year": 2021, "url": "https://arxiv.org/abs/2108.01073", "doi": "10.48550/arXiv.2108.01073", "arxivId": "2108.01073" }, { "id": "s020", "type": "paper", "title": "Scalable Diffusion Models with Transformers", "authors": [ "William Peebles", "Saining Xie" ], "year": 2022, "url": "https://arxiv.org/abs/2212.09748", "doi": "10.48550/arXiv.2212.09748", "arxivId": "2212.09748" }, { "id": "s021", "type": "paper", "title": "Elucidating the Design Space of Diffusion-Based Generative Models", "authors": [ "Tero Karras et al." ], "year": 2022, "url": "https://arxiv.org/abs/2206.00364", "doi": "10.48550/arXiv.2206.00364", "arxivId": "2206.00364" }, { "id": "s022", "type": "paper", "title": "DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling in Around 10 Steps", "authors": [ "Cheng Lu et al." ], "year": 2022, "url": "https://arxiv.org/abs/2206.00927", "doi": "10.48550/arXiv.2206.00927", "arxivId": "2206.00927" }, { "id": "s023", "type": "paper", "title": "Progressive Distillation for Fast Sampling of Diffusion Models", "authors": [ "Tim Salimans", "Jonathan Ho" ], "year": 2022, "url": "https://arxiv.org/abs/2202.00512", "doi": "10.48550/arXiv.2202.00512", "arxivId": "2202.00512" }, { "id": "s024", "type": "paper", "title": "Consistency Models", "authors": [ "Yang Song et al." ], "year": 2023, "url": "https://arxiv.org/abs/2303.01469", "doi": "10.48550/arXiv.2303.01469", "arxivId": "2303.01469" }, { "id": "s025", "type": "paper", "title": "Flow Matching for Generative Modeling", "authors": [ "Yaron Lipman", "Ricky T. Q. Chen", "Heli Ben-Hamu", "Maximilian Nickel", "Matt Le" ], "year": 2022, "url": "https://arxiv.org/abs/2210.02747", "doi": "10.48550/arXiv.2210.02747", "arxivId": "2210.02747" }, { "id": "s026", "type": "paper", "title": "Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow", "authors": [ "Xingchao Liu", "Chengyue Gong", "Qiang Liu" ], "year": 2022, "url": "https://arxiv.org/abs/2209.03003", "doi": "10.48550/arXiv.2209.03003", "arxivId": "2209.03003" }, { "id": "s027", "type": "paper", "title": "Scaling Rectified Flow Transformers for High-Resolution Image Synthesis", "authors": [ "Patrick Esser et al." ], "year": 2024, "url": "https://arxiv.org/abs/2403.03206", "doi": "10.48550/arXiv.2403.03206", "arxivId": "2403.03206" }, { "id": "s028", "type": "paper", "title": "Structured Denoising Diffusion Models in Discrete State-Spaces", "authors": [ "Jacob Austin", "Daniel D. Johnson", "Jonathan Ho", "Daniel Tarlow", "Rianne van den Berg" ], "year": 2021, "url": "https://arxiv.org/abs/2107.03006", "doi": "10.48550/arXiv.2107.03006", "arxivId": "2107.03006" }, { "id": "s029", "type": "paper", "title": "Large Language Diffusion Models", "authors": [ "Shen Nie et al." ], "year": 2025, "url": "https://arxiv.org/abs/2502.09992", "doi": "10.48550/arXiv.2502.09992", "arxivId": "2502.09992" }, { "id": "s030", "type": "paper", "title": "RePaint: Inpainting using Denoising Diffusion Probabilistic Models", "authors": [ "Andreas Lugmayr et al." ], "year": 2022, "url": "https://arxiv.org/abs/2201.09865", "doi": "10.48550/arXiv.2201.09865", "arxivId": "2201.09865" }, { "id": "s031", "type": "paper", "title": "Denoising Diffusion Restoration Models", "authors": [ "Bahjat Kawar", "Michael Elad", "Stefano Ermon", "Jiaming Song" ], "year": 2022, "url": "https://arxiv.org/abs/2201.11793", "doi": "10.48550/arXiv.2201.11793", "arxivId": "2201.11793" }, { "id": "s032", "type": "paper", "title": "Diffusion Posterior Sampling for General Noisy Inverse Problems", "authors": [ "Hyungjin Chung et al." ], "year": 2022, "url": "https://arxiv.org/abs/2209.14687", "doi": "10.48550/arXiv.2209.14687", "arxivId": "2209.14687" }, { "id": "s033", "type": "paper", "title": "Solving Inverse Problems in Medical Imaging with Score-Based Generative Models", "authors": [ "Yang Song", "Liyue Shen", "Lei Xing", "Stefano Ermon" ], "year": 2021, "url": "https://arxiv.org/abs/2111.08005", "doi": "10.48550/arXiv.2111.08005", "arxivId": "2111.08005" }, { "id": "s035", "type": "paper", "title": "Video Diffusion Models", "authors": [ "Jonathan Ho et al." ], "year": 2022, "url": "https://arxiv.org/abs/2204.03458", "doi": "10.48550/arXiv.2204.03458", "arxivId": "2204.03458" }, { "id": "s036", "type": "paper", "title": "Imagen Video: High Definition Video Generation with Diffusion Models", "authors": [ "Jonathan Ho et al." ], "year": 2022, "url": "https://arxiv.org/abs/2210.02303", "doi": "10.48550/arXiv.2210.02303", "arxivId": "2210.02303" }, { "id": "s037", "type": "report", "title": "Video Generation Models as World Simulators", "authors": [ "OpenAI" ], "year": 2024, "url": "https://openai.com/index/video-generation-models-as-world-simulators/" }, { "id": "s038", "type": "paper", "title": "Wan: Open and Advanced Large-Scale Video Generative Models", "authors": [ "Wan Team et al." ], "year": 2025, "url": "https://arxiv.org/abs/2503.20314", "doi": "10.48550/arXiv.2503.20314", "arxivId": "2503.20314" }, { "id": "s039", "type": "paper", "title": "Cosmos World Foundation Model Platform for Physical AI", "authors": [ "NVIDIA et al." ], "year": 2025, "url": "https://arxiv.org/abs/2501.03575", "doi": "10.48550/arXiv.2501.03575", "arxivId": "2501.03575" }, { "id": "s040", "type": "paper", "title": "DiffWave: A Versatile Diffusion Model for Audio Synthesis", "authors": [ "Zhifeng Kong", "Wei Ping", "Jiaji Huang", "Kexin Zhao", "Bryan Catanzaro" ], "year": 2020, "url": "https://arxiv.org/abs/2009.09761", "doi": "10.48550/arXiv.2009.09761", "arxivId": "2009.09761" }, { "id": "s041", "type": "paper", "title": "WaveGrad: Estimating Gradients for Waveform Generation", "authors": [ "Nanxin Chen et al." ], "year": 2020, "url": "https://arxiv.org/abs/2009.00713", "doi": "10.48550/arXiv.2009.00713", "arxivId": "2009.00713" }, { "id": "s042", "type": "paper", "title": "Grad-TTS: A Diffusion Probabilistic Model for Text-to-Speech", "authors": [ "Vadim Popov et al." ], "year": 2021, "url": "https://arxiv.org/abs/2105.06337", "doi": "10.48550/arXiv.2105.06337", "arxivId": "2105.06337" }, { "id": "s043", "type": "paper", "title": "AudioLDM: Text-to-Audio Generation with Latent Diffusion Models", "authors": [ "Haohe Liu et al." ], "year": 2023, "url": "https://arxiv.org/abs/2301.12503", "doi": "10.48550/arXiv.2301.12503", "arxivId": "2301.12503" }, { "id": "s044", "type": "paper", "title": "AudioLDM 2: Learning Holistic Audio Generation with Self-Supervised Pretraining", "authors": [ "Haohe Liu et al." ], "year": 2023, "url": "https://arxiv.org/abs/2308.05734", "doi": "10.48550/arXiv.2308.05734", "arxivId": "2308.05734" }, { "id": "s045", "type": "paper", "title": "MusicLDM: Enhancing Novelty in Text-to-Music Generation Using Beat-Synchronous Mixup Strategies", "authors": [ "Ke Chen et al." ], "year": 2023, "url": "https://arxiv.org/abs/2308.01546", "doi": "10.48550/arXiv.2308.01546", "arxivId": "2308.01546" }, { "id": "s046", "type": "paper", "title": "Diffusion Probabilistic Models for 3D Point Cloud Generation", "authors": [ "Shitong Luo", "Wei Hu" ], "year": 2021, "url": "https://arxiv.org/abs/2103.01458", "doi": "10.48550/arXiv.2103.01458", "arxivId": "2103.01458" }, { "id": "s047", "type": "paper", "title": "DreamFusion: Text-to-3D using 2D Diffusion", "authors": [ "Ben Poole", "Ajay Jain", "Jonathan T. Barron", "Ben Mildenhall" ], "year": 2022, "url": "https://arxiv.org/abs/2209.14988", "doi": "10.48550/arXiv.2209.14988", "arxivId": "2209.14988" }, { "id": "s048", "type": "paper", "title": "Magic3D: High-Resolution Text-to-3D Content Creation", "authors": [ "Chen-Hsuan Lin et al." ], "year": 2022, "url": "https://arxiv.org/abs/2211.10440", "doi": "10.48550/arXiv.2211.10440", "arxivId": "2211.10440" }, { "id": "s049", "type": "paper", "title": "Point-E: A System for Generating 3D Point Clouds from Complex Prompts", "authors": [ "Alex Nichol et al." ], "year": 2022, "url": "https://arxiv.org/abs/2212.08751", "doi": "10.48550/arXiv.2212.08751", "arxivId": "2212.08751" }, { "id": "s050", "type": "paper", "title": "Shap-E: Generating Conditional 3D Implicit Functions", "authors": [ "Heewoo Jun", "Alex Nichol" ], "year": 2023, "url": "https://arxiv.org/abs/2305.02463", "doi": "10.48550/arXiv.2305.02463", "arxivId": "2305.02463" }, { "id": "s051", "type": "paper", "title": "MVDream: Multi-view Diffusion for 3D Generation", "authors": [ "Yichun Shi et al." ], "year": 2023, "url": "https://arxiv.org/abs/2308.16512", "doi": "10.48550/arXiv.2308.16512", "arxivId": "2308.16512" }, { "id": "s052", "type": "paper", "title": "DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking", "authors": [ "Gabriele Corso et al." ], "year": 2022, "url": "https://arxiv.org/abs/2210.01776", "doi": "10.48550/arXiv.2210.01776", "arxivId": "2210.01776" }, { "id": "s053", "type": "paper", "title": "De novo design of protein structure and function with RFdiffusion", "authors": [ "Joseph L. Watson et al." ], "year": 2023, "url": "https://doi.org/10.1038/s41586-023-06415-8", "doi": "10.1038/s41586-023-06415-8" }, { "id": "s054", "type": "paper", "title": "MatterGen: a generative model for inorganic materials design", "authors": [ "Cheng-Hao Zeni et al." ], "year": 2023, "url": "https://arxiv.org/abs/2312.03687", "doi": "10.48550/arXiv.2312.03687", "arxivId": "2312.03687" }, { "id": "s055", "type": "paper", "title": "DiffCSP / crystal structure prediction by equivariant diffusion", "authors": [ "Rui Jiao et al." ], "year": 2023, "url": "https://arxiv.org/abs/2309.04475", "doi": "10.48550/arXiv.2309.04475", "arxivId": "2309.04475", "notes": "Title/author string should be checked against final publication record before public ingestion." }, { "id": "s056", "type": "paper", "title": "Fair comparisons between AI docking and conventional docking methods", "authors": [ "Ananth N. Jain et al." ], "year": 2024, "url": "https://arxiv.org/abs/2412.02889", "doi": "10.48550/arXiv.2412.02889", "arxivId": "2412.02889", "notes": "Final bibliographic capitalization/author list requires manual transcription." }, { "id": "s058", "type": "paper", "title": "Planning with Diffusion for Flexible Behavior Synthesis", "authors": [ "Michael Janner", "Yilun Du", "Joshua B. Tenenbaum", "Sergey Levine" ], "year": 2022, "url": "https://arxiv.org/abs/2205.09991", "doi": "10.48550/arXiv.2205.09991", "arxivId": "2205.09991" }, { "id": "s059", "type": "paper", "title": "Diffusion Policies as an Expressive Policy Class for Offline Reinforcement Learning", "authors": [ "Zhendong Wang", "Jonathan J. Hunt", "Mingyuan Zhou" ], "year": 2022, "url": "https://arxiv.org/abs/2208.06193", "doi": "10.48550/arXiv.2208.06193", "arxivId": "2208.06193" }, { "id": "s060", "type": "paper", "title": "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion", "authors": [ "Cheng Chi et al." ], "year": 2023, "url": "https://arxiv.org/abs/2303.04137", "doi": "10.48550/arXiv.2303.04137", "arxivId": "2303.04137" }, { "id": "s061", "type": "paper", "title": "3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations", "authors": [ "Ze et al." ], "year": 2024, "url": "https://arxiv.org/abs/2403.03954", "doi": "10.48550/arXiv.2403.03954", "arxivId": "2403.03954", "notes": "Full author list should be manually transcribed from primary record." }, { "id": "s062", "type": "paper", "title": "pi0: A Vision-Language-Action Flow Model for General Robot Control", "authors": [ "Kevin Black et al." ], "year": 2024, "url": "https://arxiv.org/abs/2410.24164", "doi": "10.48550/arXiv.2410.24164", "arxivId": "2410.24164" }, { "id": "s063", "type": "paper", "title": "FAST: Efficient Action Tokenization for Vision-Language-Action Models", "authors": [ "Physical Intelligence researchers; manual full-author transcription required" ], "year": 2025, "url": "https://arxiv.org/abs/2501.09747", "doi": "10.48550/arXiv.2501.09747", "arxivId": "2501.09747", "notes": "Metadata completeness: partial." }, { "id": "s064", "type": "paper", "title": "Diffusion Models Are Real-Time Game Engines", "authors": [ "Dani Valevski et al." ], "year": 2024, "url": "https://arxiv.org/abs/2408.14837", "doi": "10.48550/arXiv.2408.14837", "arxivId": "2408.14837" }, { "id": "s065", "type": "paper", "title": "Diffusion for World Modeling: Visual Details Matter in Atari", "authors": [ "Eloi Alonso et al." ], "year": 2024, "url": "https://arxiv.org/abs/2405.12399", "doi": "10.48550/arXiv.2405.12399", "arxivId": "2405.12399" }, { "id": "s066", "type": "paper", "title": "StableRep: Synthetic Images from Text-to-Image Models Make Strong Visual Representation Learners", "authors": [ "Yonglong Tian et al." ], "year": 2023, "url": "https://arxiv.org/abs/2306.00984", "doi": "10.48550/arXiv.2306.00984", "arxivId": "2306.00984" }, { "id": "s067", "type": "paper", "title": "Synthetic Data from Diffusion Models Improves ImageNet Classification", "authors": [ "Shekoofeh Azizi", "Simon Kornblith", "Chitwan Saharia", "Mohammad Norouzi", "David J. Fleet" ], "year": 2023, "url": "https://arxiv.org/abs/2304.08466", "doi": "10.48550/arXiv.2304.08466", "arxivId": "2304.08466" }, { "id": "s068", "type": "paper", "title": "Extracting Training Data from Diffusion Models", "authors": [ "Nicholas Carlini", "Jamie Hayes", "Milad Nasr", "Matthew Jagielski", "Vikash Sehwag", "Florian Tramer", "Borja Balle", "Daphne Ippolito", "Eric Wallace" ], "year": 2023, "url": "https://arxiv.org/abs/2301.13188", "doi": "10.48550/arXiv.2301.13188", "arxivId": "2301.13188" }, { "id": "s069", "type": "paper", "title": "Benchmarking Plug-and-Play Diffusion Priors for Inverse Problems", "authors": [ "manual transcription required from primary record" ], "year": 2025, "url": "https://arxiv.org/abs/2503.11043", "doi": "10.48550/arXiv.2503.11043", "arxivId": "2503.11043", "notes": "Metadata completeness: partial." }, { "id": "s070", "type": "paper", "title": "Path-Consistent Safety Filtering for Diffusion Policies", "authors": [ "manual transcription required from primary record" ], "year": 2025, "url": "https://arxiv.org/abs/2511.06385", "doi": "10.48550/arXiv.2511.06385", "arxivId": "2511.06385", "notes": "Metadata completeness: partial." }, { "id": "s071", "type": "paper", "title": "Safe Vision Language Action Models via Barrier Enhanced Flow Matching", "authors": [ "manual transcription required from primary record" ], "year": 2026, "url": "https://arxiv.org/abs/2607.29569", "doi": "10.48550/arXiv.2607.29569", "arxivId": "2607.29569", "notes": "Metadata completeness: partial." }, { "id": "s073", "type": "paper", "title": "Image-to-CAD Generation via Sequence-Based Diffusion", "authors": [ "manual transcription required from primary record" ], "year": 2026, "url": "https://arxiv.org/abs/2605.13293", "doi": "10.48550/arXiv.2605.13293", "arxivId": "2605.13293", "notes": "Metadata completeness: partial." }, { "id": "s074", "type": "paper", "title": "BrepForge: Factorized B-rep Synthesis via Wireframe Composition and Boundary-Conditioned Surface Instantiation", "authors": [ "manual transcription required from primary record" ], "year": 2026, "url": "https://arxiv.org/abs/2605.19411", "doi": "10.48550/arXiv.2605.19411", "arxivId": "2605.19411", "notes": "Metadata completeness: partial." }, { "id": "s075", "type": "paper", "title": "On Hallucinations in Inverse Problems", "authors": [ "manual transcription required from primary record" ], "year": 2026, "url": "https://arxiv.org/abs/2605.13146", "doi": "10.48550/arXiv.2605.13146", "arxivId": "2605.13146", "notes": "Metadata completeness: partial." }, { "id": "s076", "type": "paper", "title": "Reliability-Guided Diffusion for Trustworthy Ultra-Low-Field Imaging", "authors": [ "manual transcription required from primary record" ], "year": 2026, "url": "https://arxiv.org/abs/2603.11325", "doi": "10.48550/arXiv.2603.11325", "arxivId": "2603.11325", "notes": "Metadata completeness: partial." }, { "id": "s077", "type": "paper", "title": "An autonomous laboratory for the accelerated synthesis of novel materials", "authors": [ "Nathan J. Szymanski et al." ], "year": 2023, "url": "https://doi.org/10.1038/s41586-023-06734-w", "doi": "10.1038/s41586-023-06734-w" }, { "id": "s078", "type": "paper", "title": "Provably Safe Generative Sampling with Constricting Flows", "authors": [ "manual transcription required from primary record" ], "year": 2026, "url": "https://arxiv.org/abs/2602.21429", "doi": "10.48550/arXiv.2602.21429", "arxivId": "2602.21429", "notes": "Metadata completeness: partial." }, { "id": "s079", "type": "paper", "title": "WBench: A Comprehensive Multi-turn Benchmark for Interactive World Models", "authors": [ "manual transcription required from primary record" ], "year": 2026, "url": "https://arxiv.org/abs/2605.25874", "doi": "10.48550/arXiv.2605.25874", "arxivId": "2605.25874", "notes": "Metadata completeness: partial." }, { "id": "pat01", "type": "patent", "title": "Computer aided generative design with feature thickness control to facilitate manufacturing and structural performance", "authors": [ "Autodesk, Inc." ], "year": 2026, "url": "https://patents.google.com/patent/US12608510B2/en" }, { "id": "pat02", "type": "patent", "title": "Four-dimensional object and scene model synthesis using machine learning", "authors": [ "author/assignee metadata requires manual verification" ], "year": 2025, "url": "https://patents.google.com/patent/US20250182404A1/en" } ], "applicationBranches": [ { "id": "branch-a01", "title": "Image synthesis and editing", "summary": "Mature but useful: diffusion remains foundational and operationally important, while rectified-flow objectives partially displace classical diffusion at the frontier. Remaining bottlenecks include exact composition, identity consistency, provenance, memorization, and deployment cost.", "anchorNodeId": "a01", "nodeIds": [ "c03", "cap03", "cap05", "r03", "r04", "r05", "r06", "m02", "m03", "m04", "a01", "k02", "k03", "k04", "k07", "k08", "s01", "s03" ], "relationshipIds": [ "r022", "r023", "r024", "r025", "r027", "r028", "r029", "r031", "r032", "r033", "r038", "r039", "r040", "r041", "r070", "r072", "r073", "r076" ], "constraintIds": [ "constraint-k02", "constraint-k03", "constraint-k04", "constraint-k07", "constraint-k08" ], "status": { "state": "mature_but_useful", "scope": "Branch-level assessment for Image synthesis and editing; it does not imply uniform maturity across every task, dataset, or deployment.", "evidenceGrade": "direct", "sourceIds": [ "s009", "s013", "s014", "s015", "s016", "s017", "s018", "s019", "s020", "s027", "s068" ] }, "sourceIds": [ "s009", "s013", "s014", "s015", "s016", "s017", "s018", "s019", "s020", "s027", "s068" ] }, { "id": "branch-a02", "title": "Restoration and inverse problems", "summary": "Constraint-bound: diffusion and score models are strong generic priors, but measurement consistency, hallucination, calibration, and operator/domain shift limit high-stakes deployment.", "anchorNodeId": "a02", "nodeIds": [ "c01", "c04", "cap04", "m06", "a02", "k05", "k06", "k08", "opp01", "opp06" ], "relationshipIds": [ "r016", "r017", "r042", "r043", "r044", "r080", "r081", "r089" ], "constraintIds": [ "constraint-k05", "constraint-k06", "constraint-k08" ], "status": { "state": "constraint_bound", "scope": "Branch-level assessment for Restoration and inverse problems; it does not imply uniform maturity across every task, dataset, or deployment.", "evidenceGrade": "direct", "sourceIds": [ "s007", "s010", "s030", "s031", "s032", "s033", "s069", "s075", "s076" ] }, "sourceIds": [ "s007", "s010", "s030", "s031", "s032", "s033", "s069", "s075", "s076" ] }, { "id": "branch-a03", "title": "Video generation", "summary": "Rapidly expanding: diffusion-derived architectures remain central while flow formulations overlap increasingly; persistent identity, physical validity, long-horizon memory, and compute are the harder frontier.", "anchorNodeId": "a03", "nodeIds": [ "cap05", "r04", "r06", "m02", "m04", "m06", "a03", "k02", "k04", "k05", "k07", "s03", "opp02", "opp07" ], "relationshipIds": [ "r026", "r032", "r033", "r045", "r046", "r047", "r048", "r071", "r090" ], "constraintIds": [ "constraint-k02", "constraint-k04", "constraint-k05", "constraint-k07" ], "status": { "state": "rapidly_expanding", "scope": "Branch-level assessment for Video generation; it does not imply uniform maturity across every task, dataset, or deployment.", "evidenceGrade": "direct", "sourceIds": [ "s035", "s036", "s037", "s038", "s039", "s064", "s065", "s079" ] }, "sourceIds": [ "s035", "s036", "s037", "s038", "s039", "s064", "s065", "s079" ] }, { "id": "branch-a04", "title": "Audio, speech and music", "summary": "Mature but useful: diffusion is one strong option among autoregressive, flow, and hybrid systems; its remaining advantage is most plausible where stochastic refinement, editing, or global conditioning matters more than minimum latency.", "anchorNodeId": "a04", "nodeIds": [ "cap01", "r05", "m02", "m03", "a04", "k01", "s02", "opp07" ], "relationshipIds": [ "r028", "r049", "r050", "r051", "r091" ], "constraintIds": [ "constraint-k01" ], "status": { "state": "mature_but_useful", "scope": "Branch-level assessment for Audio, speech and music; it does not imply uniform maturity across every task, dataset, or deployment.", "evidenceGrade": "direct", "sourceIds": [ "s040", "s041", "s042", "s043", "s044", "s045" ] }, "sourceIds": [ "s040", "s041", "s042", "s043", "s044", "s045" ] }, { "id": "branch-a05", "title": "3D, geometry and scene generation", "summary": "Actively improving after a local stall: early single-view score distillation was slow and inconsistent, while multiview priors, direct 3D representations, and structured geometry reopened the branch.", "anchorNodeId": "a05", "nodeIds": [ "cap06", "m05", "a05", "f02", "k05", "opp05" ], "relationshipIds": [ "r052", "r053", "r054", "r055", "r087", "r088" ], "constraintIds": [ "constraint-k05" ], "status": { "state": "actively_improving", "scope": "Branch-level assessment for 3D, geometry and scene generation; it does not imply uniform maturity across every task, dataset, or deployment.", "evidenceGrade": "direct", "sourceIds": [ "s046", "s047", "s048", "s049", "s050", "s051", "s073", "s074", "pat01" ] }, "sourceIds": [ "s046", "s047", "s048", "s049", "s050", "s051", "s073", "s074", "pat01" ] }, { "id": "branch-a06", "title": "Molecules, proteins, materials and scientific design", "summary": "Rapidly expanding: geometry, equivariance, and experimental validators turn generation into scientific search, but results must be judged with domain and laboratory validation rather than generative metrics alone.", "anchorNodeId": "a06", "nodeIds": [ "cap04", "cap06", "m05", "m06", "a06", "f03", "k05", "k08", "opp04" ], "relationshipIds": [ "r056", "r057", "r058", "r059", "r060", "r074", "r085", "r086" ], "constraintIds": [ "constraint-k05", "constraint-k08" ], "status": { "state": "rapidly_expanding", "scope": "Branch-level assessment for Molecules, proteins, materials and scientific design; it does not imply uniform maturity across every task, dataset, or deployment.", "evidenceGrade": "direct", "sourceIds": [ "s052", "s053", "s054", "s055", "s056", "s077" ] }, "sourceIds": [ "s052", "s053", "s054", "s055", "s056", "s077" ] }, { "id": "branch-a07", "title": "Robotics, policies, trajectories and planning", "summary": "Reopened by new complements: action-chunk diffusion remains influential, while 3D observations, large VLM/VLA backbones, flow matching, and autoregressive action tokens create new capability and substitution paths.", "anchorNodeId": "a07", "nodeIds": [ "cap06", "m04", "m06", "a07", "k01", "k05", "k06", "s02", "s03", "opp03" ], "relationshipIds": [ "r061", "r062", "r063", "r064", "r065", "r079", "r084" ], "constraintIds": [ "constraint-k01", "constraint-k05", "constraint-k06" ], "status": { "state": "reopened_by_a_new_complement", "scope": "Branch-level assessment for Robotics, policies, trajectories and planning; it does not imply uniform maturity across every task, dataset, or deployment.", "evidenceGrade": "direct", "sourceIds": [ "s058", "s059", "s060", "s061", "s062", "s063", "s070", "s071", "s078" ] }, "sourceIds": [ "s058", "s059", "s060", "s061", "s062", "s063", "s070", "s071", "s078" ] }, { "id": "branch-a08", "title": "Simulation, synthetic data and world models", "summary": "Rapidly expanding: diffusion can generate training data or learned visual dynamics, but world-model usefulness requires action fidelity, persistent memory, physical validity, domain transfer, and task-grounded evaluation.", "anchorNodeId": "a08", "nodeIds": [ "cap05", "cap06", "m04", "m06", "a08", "k02", "k03", "k05", "k06", "k08", "s02", "s03", "opp02", "opp08" ], "relationshipIds": [ "r066", "r067", "r068", "r069", "r082", "r092", "r093" ], "constraintIds": [ "constraint-k02", "constraint-k03", "constraint-k05", "constraint-k06", "constraint-k08" ], "status": { "state": "rapidly_expanding", "scope": "Branch-level assessment for Simulation, synthetic data and world models; it does not imply uniform maturity across every task, dataset, or deployment.", "evidenceGrade": "direct", "sourceIds": [ "s064", "s065", "s066", "s067", "s068", "s079" ] }, "sourceIds": [ "s064", "s065", "s066", "s067", "s068", "s079" ] } ], "constraints": [ { "id": "constraint-k01", "title": "Iterative sampling latency", "summary": "Repeated network evaluations impede real-time, video, edge and control use. Remaining test: Matched quality-versus-wall-clock-versus-energy comparisons across samplers and objectives.", "category": "sampling_latency", "nodeId": "k01", "affectsNodeIds": [ "r01", "r02", "r07", "r08", "a04", "a07", "opp03" ], "mitigatedByNodeIds": [ "r01", "r02", "r07" ], "relationshipIds": [ "r019", "r021", "r034", "r037", "r051", "r063", "r079" ], "status": { "state": "partially_mitigated", "scope": "Constraint assessment is limited to the applications and comparisons documented in the supplied report. Matched quality-versus-wall-clock-versus-energy comparisons across samplers and objectives.", "evidenceGrade": "direct", "sourceIds": [ "s009", "s012", "s022", "s023", "s024" ] }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Repeated network evaluations impede real-time, video, edge and control use.", "sourceIds": [ "s009", "s012", "s022", "s023", "s024" ], "note": "Evidence confidence assigned by the supplied report: direct." } ] }, { "id": "constraint-k02", "title": "Training/inference compute, memory and energy", "summary": "High resolution and temporal context make training/inference expensive. Remaining test: Standardized accelerator-hours, memory, and joules per matched-quality output; the report does not establish a separate energy-saturation claim.", "category": "compute_energy", "nodeId": "k02", "affectsNodeIds": [ "cap05", "r06", "a01", "a03", "a08" ], "mitigatedByNodeIds": [ "r05", "r06" ], "relationshipIds": [ "r071" ], "status": { "state": "active", "scope": "Constraint assessment is limited to the applications and comparisons documented in the supplied report. Standardized accelerator-hours, memory, and joules per matched-quality output; the report does not establish a separate energy-saturation claim.", "evidenceGrade": "direct", "sourceIds": [ "s015", "s020", "s027", "s038" ] }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "High resolution and temporal context make training/inference expensive.", "sourceIds": [ "s015", "s020", "s027", "s038" ], "note": "Evidence confidence assigned by the supplied report: direct." } ] }, { "id": "constraint-k03", "title": "Data provenance, memorization and privacy", "summary": "Large diffusion models can reproduce examples from training sets. Remaining test: Extraction and membership auditing after deduplication and privacy mitigation; licensing requires separate legal review.", "category": "privacy_memorization", "nodeId": "k03", "affectsNodeIds": [ "a01", "a08", "opp08" ], "mitigatedByNodeIds": [], "relationshipIds": [ "r070", "r093" ], "status": { "state": "active", "scope": "Constraint assessment is limited to the applications and comparisons documented in the supplied report. Extraction and membership auditing after deduplication and privacy mitigation; licensing requires separate legal review.", "evidenceGrade": "direct", "sourceIds": [ "s068" ] }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Large diffusion models can reproduce examples from training sets.", "sourceIds": [ "s068" ], "note": "Evidence confidence assigned by the supplied report: direct." } ] }, { "id": "constraint-k04", "title": "Controllability and identity/spatiotemporal consistency", "summary": "Prompt adherence improved faster than exact spatial relations and long-term identity persistence. Remaining test: A long-horizon controlled identity and relational-consistency benchmark.", "category": "spatial_temporal_identity_consistency", "nodeId": "k04", "affectsNodeIds": [ "a01", "a03", "a05" ], "mitigatedByNodeIds": [ "r03", "m05" ], "relationshipIds": [ "r047" ], "status": { "state": "partially_mitigated", "scope": "Constraint assessment is limited to the applications and comparisons documented in the supplied report. A long-horizon controlled identity and relational-consistency benchmark.", "evidenceGrade": "direct", "sourceIds": [ "s018", "s037", "s051" ] }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Prompt adherence improved faster than exact spatial relations and long-term identity persistence.", "sourceIds": [ "s018", "s037", "s051" ], "note": "Evidence confidence assigned by the supplied report: direct." } ] }, { "id": "constraint-k05", "title": "Physical, causal and hard-constraint validity", "summary": "Plausible generations can violate measurements, geometry, chemistry, kinematics or physics. Remaining test: Exact, independently verified constraint-satisfaction rates at matched utility.", "category": "physical_causal_consistency", "nodeId": "k05", "affectsNodeIds": [ "a02", "a03", "a05", "a06", "a07", "a08", "opp02", "opp03", "opp05", "opp07" ], "mitigatedByNodeIds": [ "m05", "m06" ], "relationshipIds": [ "r043", "r048", "r055", "r060", "r068", "r088" ], "status": { "state": "active", "scope": "Constraint assessment is limited to the applications and comparisons documented in the supplied report. Exact, independently verified constraint-satisfaction rates at matched utility.", "evidenceGrade": "direct", "sourceIds": [ "s032", "s052", "s053", "s079" ] }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Plausible generations can violate measurements, geometry, chemistry, kinematics or physics.", "sourceIds": [ "s032", "s052", "s053", "s079" ], "note": "Evidence confidence assigned by the supplied report: direct." } ] }, { "id": "constraint-k06", "title": "Distribution shift, hallucination and calibration", "summary": "Strong priors can invent plausible detail when observational evidence is insufficient. Remaining test: Prospective OOD tests requiring calibrated alarm or abstention.", "category": "out_of_distribution_robustness", "nodeId": "k06", "affectsNodeIds": [ "a02", "a07", "a08", "opp01", "opp03", "opp06", "opp08" ], "mitigatedByNodeIds": [], "relationshipIds": [ "r044", "r069", "r081" ], "status": { "state": "active", "scope": "Constraint assessment is limited to the applications and comparisons documented in the supplied report. Prospective OOD tests requiring calibrated alarm or abstention.", "evidenceGrade": "direct", "sourceIds": [ "s033", "s069", "s075", "s076" ] }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Strong priors can invent plausible detail when observational evidence is insufficient.", "sourceIds": [ "s033", "s069", "s075", "s076" ], "note": "Evidence confidence assigned by the supplied report: direct." } ] }, { "id": "constraint-k07", "title": "Competition from GAN, autoregressive and flow paradigms", "summary": "Diffusion is one generative formulation among increasingly strong substitutes. Remaining test: Same-data, same-backbone, same-compute diffusion-versus-flow-versus-autoregressive ablations.", "category": "competition", "nodeId": "k07", "affectsNodeIds": [ "a01", "a03", "a04", "a07", "a08" ], "mitigatedByNodeIds": [], "relationshipIds": [ "r072" ], "status": { "state": "uncertain", "scope": "Constraint assessment is limited to the applications and comparisons documented in the supplied report. Same-data, same-backbone, same-compute diffusion-versus-flow-versus-autoregressive ablations.", "evidenceGrade": "direct", "sourceIds": [ "s006", "s025", "s026", "s027", "s062", "s063" ] }, "evidence": [ { "type": "primary_source", "grade": "direct", "claim": "Diffusion is one generative formulation among increasingly strong substitutes.", "sourceIds": [ "s006", "s025", "s026", "s027", "s062", "s063" ], "note": "Evidence confidence assigned by the supplied report: direct." } ] }, { "id": "constraint-k08", "title": "Evaluation and likelihood-perception mismatch", "summary": "Likelihood, FID, preference and downstream utility measure different things. Remaining test: Independent, task-grounded evaluations whose metrics match deployment value.", "category": "evaluation", "nodeId": "k08", "affectsNodeIds": [ "a01", "a02", "a03", "a06", "a08", "opp08" ], "mitigatedByNodeIds": [ "m06" ], "relationshipIds": [ "r073", "r074" ], "status": { "state": "active", "scope": "Constraint assessment is limited to the applications and comparisons documented in the supplied report. Independent, task-grounded evaluations whose metrics match deployment value.", "evidenceGrade": "partial", "sourceIds": [ "s009", "s010", "s020", "s069", "s079" ] }, "evidence": [ { "type": "primary_source", "grade": "partial", "claim": "Likelihood, FID, preference and downstream utility measure different things.", "sourceIds": [ "s009", "s010", "s020", "s069", "s079" ], "note": "Evidence confidence assigned by the supplied report: partial." } ] } ], "openOpportunities": [ { "id": "card-opp01", "nodeId": "opp01", "title": "Measurement-consistent generative reconstruction with calibrated abstention", "summary": "A high-value but crowded hypothesis: jointly test near-exact data consistency and calibrated abstention under prospective operator or domain shift.", "falsifiableQuestion": "Can a diffusion or flow posterior sampler achieve near-exact measurement consistency while producing uncertainty estimates that reliably identify pixels, voxels, or features dominated by the learned prior rather than the measurement?", "proposedMechanism": "Use CAP04 as a generative prior, combine it with an explicit measurement operator and likelihood, and estimate epistemic or aleatoric fragility from ensembles or posterior trajectories.", "unmetNeed": "Reliable MRI, CT, astronomy, microscopy, seismic, and other ill-posed reconstruction without presenting prior-dominated detail as measured fact.", "adjacentWorkSummary": "DPS and plug-and-play diffusion priors establish a substantial adjacent literature; InverseBench shows heterogeneous scientific settings, and the 2026 hallucination analysis sharpens the need for recognition and abstention. Broad inverse-diffusion uncertainty is not claimed as novel.", "noveltySearch": { "status": "partial", "asOf": "2026-08-19", "scope": "Targeted searches for diffusion posterior sampling, measurement-consistent diffusion, inverse-problem uncertainty, hallucination, plug-and-play priors, and medical variants.", "result": "The broad space is crowded. The narrower joint test of constraint satisfaction and calibrated abstention under prospective shift was not established by the reviewed primary set; this is not proof of novelty.", "sourceIds": [ "s032", "s069", "s075", "s076" ] }, "blockerConstraintIds": [ "constraint-k06" ], "requiredComplementNodeIds": [ "m06" ], "minimalExperiment": "Train or adapt on one accelerated-MRI scanner/domain A; reconstruct A and unseen B with injected operator shifts; measure data residual, reconstruction error, calibration error, and ability to abstain on hallucinated regions.", "baselines": [ "Diffusion Posterior Sampling (DPS)", "DDRM where the operator permits it", "A strong deterministic reconstruction model", "A domain-specific iterative reconstruction method", "Simple residual and ensemble uncertainty baselines" ], "disconfirmingResult": "The proposed uncertainty signal does not rank unsupported reconstruction errors better than simple residual or ensemble baselines.", "resources": [ "Approximately 4–8 modern GPUs for adaptation and evaluation", "Existing medical-imaging datasets", "No new imaging hardware for the first experiment" ], "crowdedness": "high", "tractability": "high", "failureReasons": [ "Posterior approximations may be miscalibrated precisely where the learned prior dominates." ], "evidenceGrade": "hypothesis", "status": "hypothesis", "sourceIds": [ "s032", "s069", "s075", "s076" ] }, { "id": "card-opp02", "nodeId": "opp02", "title": "Persistent object-centric 3D state for generative world models", "summary": "A crowded, medium-tractability hypothesis: separate persistent state from visual rendering and test long-horizon identity, geometry, physics, and policy utility at matched compute.", "falsifiableQuestion": "At fixed video-generator capacity and training data, does explicit persistent 3D or object state reduce disappearance, duplicate creation, geometric drift, and physics violations over 30–120-second action-conditioned rollouts?", "proposedMechanism": "Combine CAP05 rendering and CAP06 structured dynamics with M05 geometry so the system maintains world state independently of pixels rather than rediscovering identity in every frame.", "unmetNeed": "More reliable robotics simulators, interactive games, and planning environments with persistent objects and long-horizon physical state.", "adjacentWorkSummary": "GameNGen and DIAMOND establish diffusion world-model adjacency; physics and consistency benchmarks establish the gap. Direct 3D/4D-consistency systems and patent PAT02 mean that “3D-consistent generative video” is not itself an open white space.", "noveltySearch": { "status": "partial", "asOf": "2026-08-19", "scope": "Targeted searches for 3D world models, persistent scene state, object-centric video world models, 4D scene diffusion, physics video generation, interactive generative video, and a targeted patent check.", "result": "The field is crowded and converging. Only the matched-compute persistent-state test is retained as a hypothesis; no broad novelty is asserted.", "sourceIds": [ "s037", "s064", "s065", "s079", "pat02" ] }, "blockerConstraintIds": [ "constraint-k04", "constraint-k05" ], "requiredComplementNodeIds": [ "m05", "m06" ], "minimalExperiment": "Instrument a game environment with ground-truth object IDs and geometry, then compare a pixel-only world model with a same-capacity state-augmented model.", "baselines": [ "A DIAMOND or GameNGen-style visual world model", "A non-generative structured dynamics baseline" ], "disconfirmingResult": "State augmentation fails to improve long-horizon identity or physics metrics, or downstream policy performance, at matched compute.", "resources": [ "A few dozen high-end GPUs for a game-scale experiment", "Simulator trajectories rather than real robotics initially" ], "crowdedness": "high", "tractability": "medium", "failureReasons": [ "Learned state extraction may be harder than direct video generation.", "A state bottleneck may discard relevant visual detail." ], "evidenceGrade": "hypothesis", "status": "hypothesis", "sourceIds": [ "s037", "s064", "s065", "s079", "pat02" ] }, { "id": "card-opp03", "nodeId": "opp03", "title": "Path-consistent safety interventions for multimodal generative robot policies", "summary": "Very crowded prior art narrows this to an evaluation and mechanism question about constraint guarantees, task competence, multimodality, and distribution shift.", "falsifiableQuestion": "Can a safety mechanism guarantee state and action constraints while keeping interventions close enough to the learned behavior manifold that task completion and multimodal recovery do not collapse under distribution shift?", "proposedMechanism": "Represent multimodal action distributions with diffusion or flow policies and constrain them with a barrier or predictive-safety filter designed to preserve path consistency.", "unmetNeed": "Safe contact-rich and mobile manipulation in dynamic environments without forcing the learned policy into out-of-distribution trajectories.", "adjacentWorkSummary": "Path-consistent diffusion-policy filtering, barrier-enhanced flow matching, and constricting flows directly occupy broad safe generative control. The residual opportunity is narrow; broad novelty is explicitly disconfirmed.", "noveltySearch": { "status": "partial", "asOf": "2026-08-19", "scope": "Targeted primary-record review of diffusion-policy safety filters, barrier-enhanced flow matching for VLAs, and constrained generative sampling.", "result": "Broad “safe diffusion or flow policy” is not open white space as of the cutoff. Only a narrow comparative question about preserving multimodality, calibration, and competence remains.", "sourceIds": [ "s070", "s071", "s078" ] }, "blockerConstraintIds": [ "constraint-k05", "constraint-k06" ], "requiredComplementNodeIds": [ "m06" ], "minimalExperiment": "Use dynamic-obstacle manipulation with at least two valid avoidance modes and controlled novelty; measure violation rate, task success, action-distribution shift, and recovery diversity.", "baselines": [ "Diffusion Policy with a conventional control-barrier-function filter", "A path-consistent safety filter", "A pi0 or flow-style policy when available" ], "disconfirmingResult": "A simple reactive filter achieves equal safety and task success without a meaningful distribution-shift penalty.", "resources": [ "One to four robot arms or high-quality simulation", "Approximately 4–8 GPUs" ], "crowdedness": "high", "tractability": "high", "failureReasons": [ "Formal guarantees may remain impossible for the learned closed-loop system with approximate dynamics." ], "evidenceGrade": "hypothesis", "status": "hypothesis", "sourceIds": [ "s070", "s071", "s078" ] }, { "id": "card-opp04", "nodeId": "opp04", "title": "Joint material structure and executable synthesis-protocol generation", "summary": "A medium-high-crowdedness, medium-low-tractability hypothesis: generate candidate structures and executable synthesis protocols jointly, then learn from laboratory outcomes.", "falsifiableQuestion": "Does jointly modeling candidate structure and property with executable synthesis actions reduce the fraction of high-scoring generated materials that cannot be synthesized, compared with structure-only generation plus recipe retrieval?", "proposedMechanism": "Use scientific diffusion for continuous structures, discrete diffusion or autoregressive planning for precursor and process steps, and an autonomous laboratory for outcome feedback.", "unmetNeed": "Close the gap between in-silico material novelty and successful laboratory realization.", "adjacentWorkSummary": "MatterGen, autonomous laboratories, active learning, and synthesis planning are substantial adjacent work. The proposed residual is their coupled structure/protocol distribution with explicit learning from failed synthesis; novelty remains unassessed.", "noveltySearch": { "status": "partial", "asOf": "2026-08-19", "scope": "Targeted searches for inverse materials design, generative synthesis planning, autonomous laboratories, active learning, closed-loop discovery, reaction or protocol generation, and diffusion-material variants.", "result": "Heavy adjacent work was found. An exhaustive chemistry and materials patent search was not completed, so novelty is unassessed rather than asserted.", "sourceIds": [ "s054", "s077" ] }, "blockerConstraintIds": [ "constraint-k05", "constraint-k08" ], "requiredComplementNodeIds": [ "m06" ], "minimalExperiment": "Within one inorganic synthesis family, compare structure-only MatterGen-like proposals plus recipe retrieval against a joint model; score synthesis success, target property, and experiments-to-hit.", "baselines": [ "An A-Lab-like active-learning loop plus an existing structure generator" ], "disconfirmingResult": "Joint structure/protocol modeling does not improve synthesis success or sample efficiency.", "resources": [ "Approximately 4–16 GPUs", "Automated synthesis and characterization equipment", "Machine-readable synthesis protocols and failed-outcome data" ], "crowdedness": "medium", "tractability": "low", "failureReasons": [ "Laboratory outcome noise may overwhelm model benefit.", "Sparse, heterogeneous negative synthesis data may be insufficient." ], "evidenceGrade": "hypothesis", "status": "hypothesis", "sourceIds": [ "s054", "s077" ] }, { "id": "card-opp05", "nodeId": "opp05", "title": "Solver-certified manufacturable B-Rep and CAD generation", "summary": "A high-crowdedness hypothesis focused narrowly on kernel-executable certification and tolerance or process validity, not the already-occupied broad concept of manufacturable generative design.", "falsifiableQuestion": "Can a generative model produce native B-Rep or CAD designs that a geometry kernel and manufacturing solver certify at substantially higher yield than unconstrained generators, without collapsing design diversity?", "proposedMechanism": "Combine CAP06 structured generation and M05 exact geometry with an industrial CAD kernel and explicit topology, minimum-feature, tolerance, and process constraints.", "unmetNeed": "Move generative 3D from plausible visual assets to editable and manufacturable engineering geometry.", "adjacentWorkSummary": "BRepGen-style work, Img2CADSeq, BrepForge, and longstanding Autodesk generative-design patents make the broad concept dense. The residual test makes executable certification the training and evaluation target.", "noveltySearch": { "status": "partial", "asOf": "2026-08-19", "scope": "Targeted academic search for B-Rep and image-to-CAD generation plus targeted patent reconnaissance for generative design and manufacturing constraints.", "result": "Broad manufacturable generative CAD is not novel. Only the implementation-specific solver-in-loop certification target remains a hypothesis; the patent review is not a freedom-to-operate search.", "sourceIds": [ "s073", "s074", "pat01" ] }, "blockerConstraintIds": [ "constraint-k05" ], "requiredComplementNodeIds": [ "m05" ], "minimalExperiment": "Use a constrained bracket or fitting family with known CAD-validity tests; compare an unconstrained B-Rep generator, rejection sampling, projection, and solver-in-loop training.", "baselines": [ "A BrepForge or BRepGen-style structured generator with post-hoc repair", "Rejection sampling", "Constraint projection" ], "disconfirmingResult": "Solver-in-loop generation offers no validity or diversity advantage over inexpensive post-hoc repair.", "resources": [ "Approximately 4–8 GPUs", "A large CAD dataset", "A commercial or open geometry kernel", "No physical manufacturing for the initial experiment" ], "crowdedness": "high", "tractability": "high", "failureReasons": [ "Certification objectives may collapse shape diversity.", "Kernel failures and mixed discrete/continuous geometry may be too discontinuous for training." ], "evidenceGrade": "hypothesis", "status": "hypothesis", "sourceIds": [ "s073", "s074", "pat01" ] }, { "id": "card-opp06", "nodeId": "opp06", "title": "Prospective OOD and hallucination alarms for medical and scientific inverse diffusion", "summary": "A medium-high-crowdedness, high-tractability-if-data-exists hypothesis: issue an alarm before ground truth when scanner, anatomy, operator, or disease shifts make reconstruction fragile.", "falsifiableQuestion": "Can a system identify, before a clinician or scientist sees the reconstruction, when quality is likely to deteriorate because scanner, anatomy, acquisition operator, or disease distribution differs from training?", "proposedMechanism": "Use CAP04 posterior samples, measurement residuals, and domain embeddings to estimate fragility and trigger calibrated alarm or abstention.", "unmetNeed": "Prevent strong priors from converting uncertainty into realistic but unsupported detail in high-stakes reconstruction.", "adjacentWorkSummary": "Medical score-based reconstruction, scientific inverse benchmarks, hallucination analysis, and reliability-guided diffusion are direct neighbors. Novelty is limited to a preregistered prospective alarm test under meaningful shifts.", "noveltySearch": { "status": "partial", "asOf": "2026-08-19", "scope": "Targeted searches for OOD diffusion MRI reconstruction, inverse-problem uncertainty, medical reconstruction hallucination, and steerable conditional diffusion under OOD shift.", "result": "Direct neighboring work exists. Only the prospective, preregistered alarm or abstention test is retained as a residual opportunity.", "sourceIds": [ "s033", "s069", "s075", "s076" ] }, "blockerConstraintIds": [ "constraint-k06" ], "requiredComplementNodeIds": [ "m06" ], "minimalExperiment": "Train on one institution and acquisition protocol, then prospectively evaluate two unseen sites or protocols; issue the alarm before using ground truth.", "baselines": [ "A measurement-residual threshold", "Deterministic uncertainty quantification", "A deep ensemble", "Diffusion posterior variance" ], "disconfirmingResult": "Diffusion-derived uncertainty performs no better than ordinary residual or domain-shift detectors.", "resources": [ "Approximately 4–8 GPUs", "Multi-site retrospective data", "Clinical governance before any deployment" ], "crowdedness": "high", "tractability": "high", "failureReasons": [ "Generative uncertainty can itself be miscalibrated.", "Multi-site data access may be the binding constraint." ], "evidenceGrade": "hypothesis", "status": "hypothesis", "sourceIds": [ "s033", "s069", "s075", "s076" ] }, { "id": "card-opp07", "nodeId": "opp07", "title": "Causal audiovisual event-state generation", "summary": "A medium-crowdedness, medium-tractability hypothesis: make object contacts, material interactions, and event times explicit, then test synchronization and causal interventions.", "falsifiableQuestion": "Does a shared latent event state for objects, contacts, material interactions, and event times improve audiovisual synchronization and causal correctness beyond direct joint or separate perceptual generation?", "proposedMechanism": "Couple stochastic audio and video decoders through an explicit shared physical or event state before rendering.", "unmetNeed": "Better film and game simulation, robotics simulation, and audiovisual synthetic data with synchronized causal events.", "adjacentWorkSummary": "Diffusion-based synchronized video-to-audio and newer joint media systems make this adjacent rather than empty. The narrower target is explicit event-level state evaluated under interventions; the supplied registry lacks a fully transcribed Diff-Foley source record, so it is not added as a new citation here.", "noveltySearch": { "status": "partial", "asOf": "2026-08-19", "scope": "Targeted searches for video-to-audio diffusion, audiovisual generation, synchronization, sound-event generation, physics-conditioned sound, and joint audio-video diffusion.", "result": "Substantial perceptual-generation work exists. Explicit causal event state under controlled intervention remains a hypothesis, not a novelty claim.", "sourceIds": [ "s037", "s044" ] }, "blockerConstraintIds": [ "constraint-k05", "constraint-k08" ], "requiredComplementNodeIds": [ "m06" ], "minimalExperiment": "Generate controlled collisions, footsteps, and object interactions from a simulator with known event timing and physical parameters; compare a shared-state model with direct audiovisual generation.", "baselines": [ "A Diff-Foley-style video-to-audio system", "A modern joint audiovisual generator" ], "disconfirmingResult": "Explicit state does not improve event timing, source localization, or out-of-distribution causal interventions.", "resources": [ "Approximately 8–32 GPUs", "A large audiovisual dataset", "A modest simulator corpus with event ground truth" ], "crowdedness": "medium", "tractability": "medium", "failureReasons": [ "Perceptual encoders may already capture enough event information, making explicit state redundant.", "Event labels and physical ground truth are scarce." ], "evidenceGrade": "hypothesis", "status": "hypothesis", "sourceIds": [ "s037", "s044" ] }, { "id": "card-opp08", "nodeId": "opp08", "title": "Audited synthetic data with per-sample privacy, provenance, and utility control", "summary": "A crowded but tractable hypothesis: select generated examples with joint privacy, coverage, and downstream-utility audits rather than assuming that more synthetic data helps.", "falsifiableQuestion": "Can a diffusion or flow synthetic-data pipeline select generated examples that improve real-domain downstream performance while meeting memorization or privacy thresholds and covering underrepresented real-data modes?", "proposedMechanism": "Condition generators on requested concepts, screen samples with influence or membership tests, and estimate real-domain gap and coverage with embeddings plus downstream evaluation.", "unmetNeed": "Useful shareable training data for domains with expensive labels or redistribution limits without silently leaking examples or widening domain gaps.", "adjacentWorkSummary": "StableRep and diffusion augmentation demonstrate downstream utility, while extraction work demonstrates memorization. Privacy-preserving synthetic data is crowded; the residual question is whether sample-level multi-objective auditing predicts real-domain value.", "noveltySearch": { "status": "partial", "asOf": "2026-08-19", "scope": "Targeted searches for diffusion synthetic training data, privacy-preserving synthetic data, membership inference, memorization-aware generation, and synthetic-data attribution.", "result": "The literature is crowded. The multi-objective sample-selection test is retained as a falsifiable question, not a broad novelty claim.", "sourceIds": [ "s066", "s067", "s068" ] }, "blockerConstraintIds": [ "constraint-k03", "constraint-k06", "constraint-k08" ], "requiredComplementNodeIds": [ "m06" ], "minimalExperiment": "With a fixed real-data budget for ImageNet-like or medical classification, compare real-only training, unfiltered synthetic augmentation, privacy-only filtering, and joint privacy/coverage/utility selection.", "baselines": [ "StableRep or Azizi-style synthetic augmentation", "Simple deduplication", "Confidence-only filtering", "Diversity-only filtering" ], "disconfirmingResult": "Joint auditing fails to outperform simple confidence or diversity filtering, or its utility loss negates the privacy gain.", "resources": [ "Approximately 8–32 GPUs depending on generator scale", "No special equipment" ], "crowdedness": "high", "tractability": "high", "failureReasons": [ "Attribution and membership metrics may not track true privacy.", "Audit metrics may not predict downstream usefulness." ], "evidenceGrade": "hypothesis", "status": "hypothesis", "sourceIds": [ "s066", "s067", "s068" ] } ], "unresolvedClaims": [ { "id": "claim-u01", "claim": "Exact historical influence of denoising autoencoders on the 2015 diffusion-model conception.", "reason": "influence_not_verified", "status": "open", "relatedNodeIds": [ "p04", "c01", "c03" ], "relatedRelationshipIds": [ "r005" ], "sourceIds": [ "s004", "s007", "s009" ], "notes": "DDPM explicitly invokes Vincent’s result, but the reviewed 2015 primary text does not justify upgrading this to direct historical influence." }, { "id": "claim-u02", "claim": "Whether flow matching and rectified flow should taxonomically be called diffusion.", "reason": "scope_unclear", "status": "open", "relatedNodeIds": [ "c03", "c04", "s03" ], "relatedRelationshipIds": [ "r038", "r078" ], "sourceIds": [ "s025", "s026", "s027" ], "notes": "Shared probability-path and SDE connections coexist with distinct objectives and deterministic transport formulations; the alpha treats flow as a formal neighbor and partial substitute." }, { "id": "claim-u03", "claim": "Global image-generation saturation.", "reason": "conflicting_evidence", "status": "resolved", "relatedNodeIds": [ "a01", "r06", "s03", "k07", "k08" ], "relatedRelationshipIds": [ "r033", "r039", "r072", "r073" ], "sourceIds": [ "s020", "s027" ], "notes": "Rejected as unsupported at the report cutoff; scaling results are counterevidence, while only particular engineering dimensions show mature returns." }, { "id": "claim-u04", "claim": "DiffDock establishes general superiority of diffusion for docking.", "reason": "conflicting_evidence", "status": "partially_resolved", "relatedNodeIds": [ "a06", "f03" ], "relatedRelationshipIds": [ "r059" ], "sourceIds": [ "s052", "s056" ], "notes": "Later conventional-docking comparisons materially alter the ranking; matched replication remains a manual-review requirement." }, { "id": "claim-u05", "claim": "Diffusion video models are genuine general-purpose world models.", "reason": "scope_unclear", "status": "open", "relatedNodeIds": [ "a03", "a08", "k04", "k05", "k06" ], "relatedRelationshipIds": [ "r045", "r048", "r067", "r068", "r069" ], "sourceIds": [ "s037", "s064", "s065", "s079" ], "notes": "Interactive demonstrations exist, but long-horizon physical, causal, and domain-transfer evidence remains insufficient." }, { "id": "claim-u06", "claim": "Standardized diffusion energy inefficiency relative to competing paradigms.", "reason": "other", "status": "open", "relatedNodeIds": [ "k02", "k07" ], "relatedRelationshipIds": [ "r071", "r072" ], "sourceIds": [ "s009", "s012", "s022", "s027", "s038", "s039" ], "notes": "The report found compute, memory, model-size, and step-count evidence but insufficient comparable joules-level measurements across model families." }, { "id": "claim-u07", "claim": "Broad novelty of safe diffusion or flow robot policies.", "reason": "novelty_unverified", "status": "resolved", "relatedNodeIds": [ "a07", "opp03" ], "relatedRelationshipIds": [ "r079", "r084" ], "sourceIds": [ "s070", "s071", "s078" ], "notes": "Broad novelty is disconfirmed by direct 2025–2026 work; only the narrower intervention-quality hypothesis remains." }, { "id": "claim-u08", "claim": "Broad novelty of generative manufacturable CAD.", "reason": "novelty_unverified", "status": "resolved", "relatedNodeIds": [ "a05", "opp05" ], "relatedRelationshipIds": [ "r087", "r088" ], "sourceIds": [ "s073", "s074", "pat01" ], "notes": "Direct B-Rep generation work and generative-design patents occupy the broad concept; patent reconnaissance was not a freedom-to-operate search." }, { "id": "claim-u09", "claim": "Bibliographic completeness for selected 2025–2026 primary records.", "reason": "other", "status": "open", "relatedNodeIds": [ "r08", "a02", "a05", "a06", "a07", "opp03", "opp05", "opp06" ], "relatedRelationshipIds": [], "sourceIds": [ "s055", "s056", "s061", "s063", "s069", "s070", "s071", "s073", "s074", "s075", "s076", "s078", "s079" ], "notes": "Stable identifiers and titles were supplied, but several author lists or final publication details require manual transcription and current-version review before public ingestion." } ] }