Repository coverage, depth and visual audit — 2026-08-30

This is the working gap matrix for the long-term goal: every subfield should have an understandable visual summary, a current technical route, defensible milestones, detailed paper review and explicit evidence limits.

Scoring and priority

• Depth 0: absent.
• Depth 1: outline or paper-name list.
• Depth 2: coherent introductory tutorial.
• Depth 3: review chapter with mechanisms, comparisons, evaluation and failure analysis.
• Depth 4: review chapter plus reproducible evidence trail, current primary sources, quantitative protocols and verified visuals.
• Reproduction is a separate axis: Depth measures the chapter/review artifact, not whether a large checkpoint was executed. Proposed-only experiments are labelled in each row and research log; actual smoke/matched/independent runs must never be inferred from Depth 4. Navigation pages and bibliography infrastructure use N/A when the subfield checklist is not applicable.
• P0: factual/classification error, unsupported claim, structural absence or field coverage clearly stale by 2026-08-29.
• P1: route is sound but misses mechanisms, milestones, quantitative evidence or a major recent branch.
• P2: presentation, visual coverage, metadata normalization or secondary breadth.

Counts are a repository snapshot after adding the causal-streaming, video post-training, native audio-video, fine-grained control, video degradation restoration, video-tokenizer, multi-view/4D, Video DiT/backbone and open-set video-personalization chapters, with the mechanism, foundation, task-map, application, getting-started, interpolation, inpainting, text-to-video, image-to-video, video-to-video, video-prediction, action-conditioned prediction, interactive-world, digital-human, story/multi-shot, unconditional-video, open-model, video-reasoning, JEPA, diffusion/flow, physics, variational-stochastic-future, timeline, reading-route and bibliography rows refreshed on 2026-08-30. “Visual” counts Mermaid blocks plus embedded images; a visual can still fail the scientific-quality gate. The former mixed VAE/tokenizer page is now split: variational-generation.md owns ELBO, learned prior, posterior collapse and multi-future evidence, while video-tokenizers.md owns representation, quantization, causal/chunk boundaries, exact budgets, bitstream maturity and fixed-consumer replacement tests. video-dit-backbones.md separately owns post-codec patch/token geometry, mixer/mask, position/fusion, active capacity and execution scaling; it does not absorb objective derivations or streaming SLO. Open-set personalization now owns reference-outside-timeline subject state, adaptation, binding, leakage and identity–motion Pareto evidence; I2V, V2V, generic control, digital-human and story pages retain their adjacent contracts. The unified bibliography now has dedicated variational stochastic-video and open-set personalization sections, while chapter-local branches remain outside the core registry unless they meet its cross-chapter inclusion policy.

A. Generation mechanisms and foundation models

File	Lines	Refs	Visuals	Current depth	Main gap	Priority
docs/generative-models.md	466	32	4	4	Five-axis cross-classification now routes representation, objective, backbone and deployment to disjoint specialist contracts; cascade/cache no longer leak into the backbone taxonomy; next gate is controlled implementation comparison	P2
recurrent-prediction.md	370	23	3	4	Stochastic recurrent latents now route ELBO/learned-prior semantics to variational generation and compression-only representations to the tokenizer contract; next gate is matched open-loop/closed-loop reproduction	P2
variational-generation.md	591	31	4	4	Refreshed from the 2015 VRNN interface through 2026 LPWM/direct-video continuations, with strict train-posterior/deployment-prior visibility, sequential ELBO, collapse/gap diagnostics, sample-aggregation boundaries and a named thresholded LatentFork-1; next gate is matched implementation and independent prior-only reproduction	P2
video-tokenizers.md	416	25	3	4	New representation/codec review spanning VQ/LFQ/FSQ/BSQ, exact finite-clip and bitstream ledgers, causal prefix/chunk contracts, adaptive/structured/aligned frontiers, BS0–BS3 and TokenizerFork-1; next gate is matched open-checkpoint and independent bitstream reproduction	P2
video-dit-backbones.md	576	35	3	4	New primary-source review from U-Net/DiT ancestry through full/factorized/window/sparse/linear/hybrid attention, position/fusion, noise-time experts, distributed execution, cache and quantization; includes a verified original compute-contract visual plus accessible block/falsification diagrams and proposed BackboneFork-1/ServeFork-1; no model run was executed	P2
adversarial-generation.md	358	32	3	4	Rewritten across full video GANs, tokenizer/decoder critics and explicit adversarial video distillation, including negative classifications; next gate is fixed-teacher/NFE/critic reproduction	P2
autoregressive-generation.md	647	20	2	4	Rewritten across strict token, frame, set and chunk factorization, continuous heads, cache complexity and forcing; next gate is a matched decoder/cache implementation study	P2
masked-generation.md	399	18	3	4	Rewritten across absorbing-state/discrete diffusion links, confidence calibration and token/frame/tube/block schedules; next gate is controlled schedule/calibration reproduction	P2
diffusion-models.md	409	24	2	4	DDPM→score→reverse SDE→PF-ODE remains chapter-owned; duplicated tokenizer taxonomy is reduced to an integration boundary, and rate claims now require bitstream evidence; next gate is numerical notebook reproduction	P2
flow-consistency-models.md	636	20	4	4	Rewritten across FM/RF/reflow, CM/sCM/rCM, Shortcut/MeanFlow/FACM/AlphaFlow and DMD/DMD2; a verified five-layer map now separates training objectives, continuous processes, training-free solvers, learned few-step routes and the orthogonal streaming axis; next gate is matched-budget reproduction	P2
causal-streaming-generation.md	658	28	5	4	Four-layer codec→generator→commit→SLO contract, four clocks, rolling-noise/commit visual, bounded-memory lifecycle, formal-status/artifact audit and proposed StreamFork-1 are explicit; global dense cost and data-time KV cache are now separated from other mixers/inter-step reuse; no checkpoint/runtime run was executed	P2
video-post-training-alignment.md	385	38	2	4	New focused review separating CPT/SFT, RM, pair construction, DPO/RWR, online feedback, policy-gradient RL, test-time guidance and reward-guided distillation, with credit, cost, hacking and evaluator-independence gates; next gate is matched-checkpoint reproduction	P2
foundation-models.md	360	42	3	4	System view now separates tokenizer, objective, backbone/execution and deployment manifests; detailed acceptance adds parameter/FLOP-matched architecture and fixed-checkpoint serving rows while retaining 4D/streaming/bitstream contracts	P2

Mechanism-level structural correction

The repository should stop treating these as one mutually exclusive taxonomy:

1. representation: pixel, continuous latent, discrete token, hybrid;
2. factorization: full-sequence, recurrent, autoregressive, masked, chunk/rolling;
3. training objective: likelihood/ELBO, adversarial, diffusion/score, flow matching, consistency/DMD, preference/RL;
4. backbone: CNN/RNN, U-Net, Transformer/DiT or recurrent/SSM; then mixer/mask, position/fusion and FFN/MoE routing inside that implementation;
5. deployment: offline, causal streaming, real-time interactive.

Modern systems routinely choose one item from several axes. Diffusion Forcing, CausVid, Self Forcing, MarDini and MaskFlow make this especially visible.

B. Task chapters

File	Lines	Refs	Visuals	Current depth	Main gap or correction	Priority
text-to-video.md	455	36	4	4	Rewritten around a strict task boundary, orthogonal route matrix, runtime condition contract, optional post-training, native AV, release surfaces and acceptance evidence; next gate is matched open-checkpoint reproduction	P2
image-to-video.md	467	34	3	4	Pixel/latent/soft anchors, condition tensors and preservation–motion evidence remain here; reference-only open-set subject state now routes explicitly to personalization; next gate is matched checkpoint/control reproduction	P2
video-to-video.md	447	47	3	4	The source-time-axis contract, eight route families and source preservation remain here; per-subject adaptation on a new timeline now routes to personalization; next gate is matched-checkpoint editing reproduction	P2
digital-human.md	368	45	3	4	Seven human-specific tasks retain audio/pose synchronization and authorization, while generic person/animal/object subject binding routes to personalization; next gate is matched open-checkpoint reproduction	P2
frame-interpolation.md	480	32	4	4	Rewritten as correspondence/reconstruction versus generative routes, with hard endpoint constraints, milestone-date rules, protocol traps, failure analysis and a reproducible evidence log; next gate is matched-data/runtime reproduction	P2
video-prediction.md	468	30	3	4	Rewritten around train-only versus deployment-GT boundaries, deterministic/stochastic/latent/diffusion routes, forcing distinctions, calibrated rollout and an evidence ladder; next gate is matched long-rollout and downstream-utility reproduction	P2
action-conditioned-prediction.md	447	33	3	4	Rewritten around action clocks, intervention versus logged correlation, pixel/probabilistic/task/representation/generative routes, model exploitation, calibrated futures and real closed-loop utility; next gate is the preregistered ActionFork-2D reproduction	P2
video-restoration.md	361	30	3	4	New inverse-problem review separating blur/downsample/noise/compression recovery from masked completion, spanning alignment, propagation, Transformer, generative prior and causal/one-step routes through formal 2026 work; includes degradation manifests, four-axis evaluation, hallucination audit and RestorationFork-1; next gate is matched open-checkpoint reproduction	P2
video-inpainting.md	546	38	3	4	Rewritten from valid-pixel propagation through missing-region synthesis and global temporal verification, with mask-outside preservation, scene-cut reset, benchmark traps and a reproducible evidence log; now explicitly excludes full-frame degradation restoration; next gate is controlled mask/protocol reproduction	P2
story-multishot.md	394	34	3	4	Shot planning, cross-shot state, memory transactions and rollback remain here; single-shot subject extraction/binding routes to personalization, and PoCo now uses its official title; next gate is matched long-story reproduction	P2
interactive-world-generation.md	463	39	3	4	Rewritten around action schedules, dual clocks, speculative versus authoritative memory, formal 2026 frontiers, model/product/artifact release surfaces, external interlock and post-hoc audit; next gate is common-hardware closed-loop reproduction	P2
unconditional-video-generation.md	456	35	3	4	Rewritten around strict deployment-time \(p(X)\) boundaries, GAN/token/diffusion/flow routes, fidelity–coverage–tail–memorization evidence, protocol traps and a reproducible experiment; Sora/Cosmos moved to adjacent conditional systems; next gate is matched-data reproduction	P2
native-audio-video-generation.md	479	33	2	4	New task-contract review separating V2A/A2V, shared-condition, staged, inference-coupled and native joint generation, then auditing codecs, clocks, dual-stream/single-stream fusion, streaming memory, search and five independent acceptance gates; next gate is a fixed-output open-checkpoint comparison	P2
controllable-video-generation.md	524	45	2	4	Camera, trajectory, pose/structure and generic identity-signal injection remain here; open-set adaptation, subject slots, binding and leakage route to personalization; next gate is a matched-backbone controller reproduction	P2
multiview-4d-generation.md	538	32	3	4	New camera×world-time review separating video paths, multi-view grids and renderable states, with reconstruction/generative/streaming routes, 2021–2026 milestones, six evidence gates and proposed GridFork-1; no model run was executed	P2
personalized-video-generation.md	663	35	3	4	New reference-outside-timeline review covers seven combinable route families, 2022–2026 milestones, five detailed red-team reviews, multi-subject binding/leakage falsifiers, rights/revocation, release surfaces and proposed PersonaBind-1; no checkpoint run was executed	P2
video-virtual-try-on.md (planned)	0	0	0	0	Structural absence: garment correspondence, body/identity/background conservation, occlusion/turning, temporal seams, dataset rights and video-specific try-on evaluation are deferred from inpainting but not yet owned by a chapter	P0

C. Reasoning, world models, physics and evaluation

File	Lines	Refs	Visuals	Current depth	Main gap	Priority
video-reasoning.md	811	58	3	4	Added UniVR, the causal-generative gap and RuleMaze as a scoped adjacent control; normalized MME-CoF/VChain/Thinking-with-Video/VIPER venues, separated output, denoising and interaction clocks, and renamed its local R1–R4 scale to avoid collision with the canonical L0–L7 evidence ladder; next gate is causal-state and closed-loop reproduction	P2
world-models.md	542	33	3	4	Rewritten around cascaded versus joint world-action routes, receding-horizon execution, pseudo-action recovery, verifier scope, persistent memory, uncertainty and decision utility; its local WM0–WM6 reporting stack maps explicitly to the canonical global L0–L7 ladder, and both Mermaid diagrams plus the dual-route figure have accessible text alternatives; next gate is independent closed-loop reproduction	P2
physical-consistency.md	594	41	5	4	Rewritten with a six-level concept taxonomy, one canonical L0–L7 evidence ladder, 2024–26 method/benchmark registries and an explicit condition→state→rollout→measurement→falsification contract; training signals are separated from sealed evaluators, and the next gate is independent GAUGE/robotics replication	P2
evaluation.md	906	62	3	4	The five task-specific success definitions now include open-set personalization with subject/reference/prompt/seed/budget units, binding/leakage failures and a separate report template; restoration, streaming, 4D, variational and backbone contracts remain distinct	P2
jepa.md	340	19	4	4	Rewritten around the exact teacher–student contract, collapse, dense/semantic trade-off, action-conditioned latent MPC, V2.1 attribution, multi-future uncertainty, release surfaces and claim-specific tests; next gate is matched latent-control reproduction	P2

D. Navigation, applications and resources

File	Lines	Refs	Visuals	Current depth	Main gap	Priority
getting-started.md	279	link-only	4 embedded images	N/A	Engineering and research routes now include backbone/token-attention-cost and stochastic-future deployment-contract checkpoints before system claims; navigation pages are not scored as subfield reviews	P2
taxonomy.md	304	12	3	4	Condition source × source-content relation × interaction horizon now includes open-set subject references as a separate task and distinguishes them from I2V anchors, V2V sources, generic control and multi-shot state	P2
applications.md	331	12	3	4	Capability→workflow→acceptance now adds virtual production/AR/VR dynamic assets and separates camera path, multi-view grid and renderable 4D state	P2
timeline.md	935	0 local anchors	2 Mermaid + 75 images	4	A text-only 2022–2026 personalization lineage now separates per-subject adaptation, amortized open-set binding, benchmark infrastructure and multi-reference frontiers while preserving all 75 image pairs	P2
reading-list.md	613	current external links	2	4	The shared tokenizer/stochastic-future/Video-DiT trunk now feeds eight task branches; branch H provides a complete personalization paper route, falsifiers and deliverables without changing the seven-axis model manifest	P2
resources/datasets.md	795	38	2	4	Rewritten as a current scoping review with release-surface taxonomy, an exact ten-stage data-engine path plus governance loop, rights/provenance, action/physics branches, manifest and fixed-compute validation; release-unit claims passed independent audit, and the next gate is acquisition/decode-yield replication	P2
resources/open-models.md	451	35	4	4	Dated release-surface audit now also prevents DiT names from implying topology/routing; Hunyuan SSTA, Wan2.2 total/active and SANA2 checkpoint boundaries are explicit; next gate is clean-environment reproduction	P2
bibliography.md / registry	188 / registry	132 registered core items	0	N/A	The new eight-item section N covers open-set single/multi-subject personalization; registry, metadata, 66 unique GitHub snapshots, BibTeX and generated index remain identity-consistent	P2

P0 factual and evidence corrections

These should be fixed before broad stylistic expansion:

1. ~~Reclassify IFRNet as an efficient convolutional encoder–decoder/feature-refinement method, not a Transformer.~~ Corrected on 2026-08-30 in the interpolation chapter.
2. ~~Stop using general Video GAN, MoCoGAN, SVG, SVD or Video Diffusion Models as direct task-specific milestones without an explicit “technical ancestor” label.~~ Corrected on 2026-08-30 across the refreshed task chapters; direct milestones and technical ancestors are now distinguished.
3. ~~Remove or replace SVD as evidence for an LLM storyboard/multi-shot pipeline.~~ Corrected on 2026-08-30; the story chapter now separates SVD as an unrelated ancestor and audits current multi-shot routes.
4. ~~Remove Sora and Cosmos from the representative unconditional-generation route.~~ Corrected on 2026-08-30; both are now adjacent conditional systems, not direct milestones.
5. ~~Standardize dates as “first preprint year / formal publication year”; FramePack, Hallo, SadTalker, ShotAdapter and LDF-VFI require the same rule.~~ Corrected on 2026-08-30; LDF-VFI is recorded as 2026-01 first preprint → CVPR 2026 formal publication, and the timeline now applies the distinction systematically.
6. ~~Mark the digital-human overview's Identity 0.92 and Temporal Consistency 0.90 as illustrative or regenerate it without fabricated-looking scores.~~ Corrected on 2026-08-30 by keeping the legacy file uncited and using a score-free contract visual; on 2026-08-31 the original path was additionally redrawn in Chinese, the numeric scores were removed, and the remaining evaluation bars were explicitly marked as illustrative rather than experimental results.
7. ~~Update C2PA from 2.2 to the official 2.4 specification and preserve the version date.~~ Corrected on 2026-08-30 in the evaluation and application evidence boundaries.
8. ~~Split model technical significance from current product availability; a discontinued product does not erase its research milestone, but should not be presented as currently available.~~ Corrected on 2026-08-30 with independent timeline evidence axes and dated availability states.
9. ~~Stop routing full-frame super-resolution, deblurring, denoising and decompression to a mask-completion chapter, and stop reusing local L-level names for incompatible evidence scales.~~ Corrected on 2026-08-30 with a dedicated degradation-restoration contract, separate inpainting boundary, canonical global L0–L7 ladder and explicitly mapped WM/R local scales.
10. ~~Complete the tokenizer/VAE ownership split: move representation, quantization, causal-codec and bitstream accounting out of variational-generation.md, keep ELBO/learned-prior/multi-future semantics there, then recount both chapters and re-run link/reference/depth checks.~~ Corrected on 2026-08-30: the two independently readable pages now own disjoint task contracts, both have closed references and rendered accessible diagrams, and shared pages reserve bpp/bitrate for entropy-coded bitstreams.
11. ~~Stop conflating review depth with checkpoint reproduction, omit navigation pages from the inventory, and leave multi-view/4D as an unnamed control sub-branch.~~ Corrected on 2026-08-30: Depth now measures the review artifact, run status remains explicit, getting-started.md/bibliography are N/A, and camera×time 4D has its own chapter and evidence contract. The variational row reached Depth 4 only after its current-frontier, protocol and visual refresh; checkpoint reproduction remains a separate unmet axis.
12. ~~Treat best-of-\(K\) or posterior-assisted output as calibrated deployment evidence, mix direct stochastic-future prediction with RSSM/world-model acceptance, and collapse first-public and formal-publication years.~~ Corrected on 2026-08-30: prior-only deployment samples, random-set/proper scores and posterior-oracle advantage are now separate; action/reward/return remain world-model evidence; PlaNet, DreamerV2/V3, DDLP, LPWM and Video Diffusion Models use explicit first-public/formal status.

Visual implementation queue

Each visual must have an explicit learning objective, editable/accessible alternative and a post-generation visual check.

Priority	Visual	Learning objective
Done	Orthogonal mechanism cube/map	Show representation, factorization, objective, backbone and deployment as composable axes
Done	Video token generation routes	Separate continuous latent, discrete AR and masked/discrete routes while keeping representation, factorization, objective and deployment claims orthogonal
Done	Tokenizer four-ledger finite-clip contract	Derive exact grid/element counts for a 9-frame example, then separate nominal token capacity from probability-model/entropy-coded bitstream
Done	Causal tokenizer first-frame/chunk contract	Show first-frame anchoring, past-only receptive fields, cache carry/reset, optional overlap/crop, committed frames and the non-implication chain to real-time SLO
Done	Per-token noise to rolling commit frontier	Separate next-unit, full-sequence and rolling schedules; show that noise levels do not imply causal access, self-history, few-step or SLO
Done	Bounded generator-memory lifecycle	Show write, aging, anchor/persistent/compressed/state/retrieval/drop routes; separate GPU working set from growing external storage and information loss
Done	Streaming commit/backpressure/recovery state machine	Trace speculative output through decode/gate/hash/immutable commit/display, with declared degradation, backpressure, cache reset and future-leak probes
Done	Modern foundation system stack	Track governance→representation→generator→post-train→decode→deployment and separate full GAN, codec critic and adversarial distillation roles
Done	DDPM–SDE–PF-ODE–FM/RF–CM/DMD map	Prevent the most common diffusion/flow/consistency terminology errors
Done	Offline bidirectional vs causal streaming	Separate full-clip denoising from chunk-causal output, bounded memory and new-condition loop
Done	Data/provenance pipeline	Source→shot split→dedup→filter→caption→license/provenance→versioned split; generated figure plus accessible Mermaid alternative verified
Done	Physics evidence loop	Condition→state→generator/simulator→measurement→constraint/reward→falsification
Done	WAM dual route	Compare cascaded planner+generator with joint action-video model and verifier
Done	Three-axis task map	Condition source, source-preservation strength and open-loop→closed-loop
Done	T2V prompt-to-evidence contract	Separate pure text, source-pixel hybrid and closed-loop tasks; connect runtime conditions, optional post-training, generation and acceptance gates
Done	I2V anchor-and-control contract	Separate pixel/latent/soft anchors and trace image, motion, camera and audio conditions into the denoiser and preservation/motion gates
Done	Video-prediction deployment contract	Separate train-only future supervision from deployment rollout and connect claim strength to calibration, intervention and closed-loop evidence
Done	Interactive-world closed-loop stack	Separate causal action flow, memory read/update, external pre-execution interlock and post-hoc evidence audit
Done	VFI dual branch	Flow/depth/occlusion deterministic branch vs diffusion/DiT generative branch
Done	Video inpainting pipeline	Valid-pixel propagation→missing-region synthesis→global consistency→outside-mask protection
Done	Video degradation restoration contract	Degradation observation→alignment/propagation/generative routes→fidelity/temporal/detail gates, with inpainting as a separate missing-support branch
Done	Video reasoning three clocks	Output-time, denoising-time and interaction-time reasoning with intervention points
Done	Evaluation evidence ladder	L0 visual plausibility through L7 real-world closed-loop utility
Done	Multi-shot memory workflow	Script/character bible→shot plan→references→generation→memory update→conflict/rollback
Done	Unconditional-generation evidence chain	No external condition→marginal model→sampling→fixed output contract→quality/coverage/tail/memorization evidence
Done	Digital-human condition and sync contract	Authorized inputs→timeline alignment→condition fusion→generation→separate sync/identity/motion/media gates
Done	Open-model release surface	Official identity→version→artifact split→license intersection→hardware→smoke test→reproducibility manifest
Done	V2V method selector	Local/global, appearance/motion, new-view and multi-turn requirements to method families
Done	Action-conditioned intervention-to-control loop	Observed history→action schedule→counterfactual rollouts→fresh-observation planning loop
Done	JEPA MPC loop	Encoder→action-conditioned predictor→cost→planner→action→new observation
Done	Video post-training parallel-route evidence map	Separate data/SFT, pairwise preference, reward/RL, test-time optimization and distillation without implying a compulsory pipeline
Done	Native audio-video generation contract	Separate V2A, A2V, shared-condition, staged and learned-joint factorization before evaluating synchronization
Done	Fine-grained control contract	Trace camera, trajectory, structure and identity signals through coordinates/conflicts, injection and four independent acceptance gates
Done	Camera-view × world-time 4D contract	Distinguish a video row, multi-view image column, camera-controlled diagonal and full view-time grid, then separate the grid from a renderable dynamic state
Done	Video DiT compute contract	Derive token growth, distinguish full/factorized/window-sparse/linear/hybrid topology, separate token reduction, noise-time experts and parallel/cache, then require output/FLOPs/latency/VRAM/communication/quality gates
Done	Video DiT block ownership map	Route codec latent→patch/position→mixer/fusion→FFN/experts→objective head while keeping tokenizer, backbone and sampler chapters disjoint; accessible Mermaid and sequential text supplied
Done	BackboneFork-1 evidence map	Freeze codec/data/objective/sampler/output, run parameter- and FLOP-matched forks, then retain a claim only after quality, long-range, grid and cost falsifiers pass
Done	Variational train-versus-deployment future contract	Separate the train-only future posterior from the history-only deployment prior, then test event intervention, prior–posterior match, mode support and calibrated frequency without treating best-of-100 as calibration
Done	Open-set personalization binding map	Separates reference sets, subject slots and four route choices; successful generation and blend/drop/copy/freeze failures passed original-size and grayscale inspection
P0	Video virtual try-on conservation contract	Trace garment/body/pose/mask inputs through correspondence, temporal generation and detail recovery, then independently protect garment pattern/material, person identity/body, background and cross-frame continuity

Primary-source refresh queues

Before editing each field, verify its queue at the primary paper or official proceedings; do not copy this audit's shorthand into final prose without a fresh read.

• Variational stochastic video: VRNN, SV2P, SVG-LP, SAVP, Improved Conditional VRNN, SRVP, GHVAE, CW-VAE, DDLP, LPWM and the 2026 direct long-horizon continuation; freeze what posterior and prior can see, distinguish direct video prediction from RSSM/control, preserve first-public/formal dates and report oracle, random-sample, proper-score and prior-only evidence separately.
• Open-set personalization (completed 2026-08-30): Video Alchemist, Movie Weaver, PersonalVideo, ConsisID and AlcheMinT received table-level red-team review; the chapter extends through formal 2026 multi-reference work while separating architecture support, qualitative evidence and matched quantitative binding.
• Next uncovered task: audit video virtual try-on from dynamic-pose and fast/video-DiT routes into 2026 detail injection. Framewise image-try-on scores cannot substitute for temporal garment/body/background conservation evidence.
• Representation/tokenizers: MAGVIT-v2, CV-VAE, OmniTokenizer, VidTok, Causal VAE, BSQ-ViT, ElasticTok, LARP, CoordTok, VidTwin, Divot, InfoTok, AdapTok, NeRV-Diffusion, VideoRAE, V-RAE and KVAE; always freeze exact shape/dtype/token accounting, mark formal/preprint status and distinguish real entropy-coded bitstreams from nominal rates.
• Mechanisms/foundation: Diffusion Forcing, MarDini, MaskFlow, LTX-Video, VidTok, Pyramid Flow, T2V-Turbo, CausVid, Self/Causal Forcing, Rolling Forcing, LongLive, SCD, FlowCache, MotionStream, StreamDiffusionV2, Causal-rCM, Step-Video, Open-Sora 2, SkyReels-V2, MAGI-1, Wan2.2, HunyuanVideo 1.5, Ovi, LTX-2; VideoDPO, VideoAlign, DynamicsBoost, Dual-IPO, BranchGRPO and noisy-latent rewards. For causal/streaming work, separately verify history exposure, measured NFE, commit semantics, resident/external memory and end-to-end SLO.
• Backbone/scaling (structural chapter completed 2026-08-30): factorized/full/window/sparse/linear/hybrid attention, MMDiT/Expert AdaLN, 3D RoPE, noise-time MoE, context parallelism, inter-step reuse, quantization and current formal/artifact status are now separated in video-dit-backbones.md; next gate is to execute BackboneFork-1 on matched small models and ServeFork-1 on one fixed open checkpoint, without multiplying independent paper speedups.
• Multi-view/4D: D-NeRF, Nerfies, MAV3D, 4D-GS, 4Real, 4DiM, SV4D/2.0, GenXD, CAT4D, 4Real-Video, 4C4D, DGGT, MoRel, 4DSurf, Full-4D, MV-Forcing, Stream4D and 4DStreamCtrl; always separate camera path, view-time grid, renderable state, unseen-region generation, scene-build cost and query/render cost.
• Physics/evaluation: NewtonRewards, Physion-Eval, Physics-IQ Verified, Apple-π, GAUGE, Interpreting Physics in Video World Models, VBench 2.0, WorldReasonBench, WBench, MIND, EntityBench, HuM-Eval/HuM-Bench.
• World/action: WAM survey, World Action Verifier, DreamGen, ViPRA, WorldPack, Infinite-World, ReWorld.
• Tasks: IFRNet, EDEN, BiM-VFI, LDF-VFI; MCVD/FramePack; STTN/FuseFormer/E²FGVI/ProPainter; STIV/ReasonDiff; ShotAdapter/OneStory/MultiShotMaster; VACE/EasyV2V/Editto; Exploration-Driven GIE/Vid2World/Dexterous World Models; MM-Diffusion/Ovi/LTX-2/NAVA/Ripple; MotionCtrl/Motion Prompting/GEN3C/LAMP/BulletTime/FlashMotion/4DStreamCtrl.
• Resources: InternVid, OpenVid-1M, Physion/CLEVRER/PHYRE, NewtonBench-60K, VBVR, AgiBot World, RoboMIND/RH20T, modern audio-video and multi-view/3D sets.

Batch verification — action/V2V/reasoning/JEPA refresh

Final checks on 2026-08-30 covered the four chapters, four dated research records and this audit:

Check	Result
Markdown	markdownlint-cli2 0.23.2 / markdownlint 0.41.1: 9 files, 0 issues
Reference closure	action 33 refs / 67 citations; V2V 47 / 119; reasoning 58 / 155; JEPA 19 / 60; no missing or orphan anchor
Local links and images	31 chapter-local targets resolved; 0 missing
Mermaid	8 blocks total; all contain accessibility metadata and all 8 rendered with Mermaid CLI plus system Chrome to non-empty PNG/SVG audit artifacts
Mermaid visual check	2 action, 2 V2V, 1 reasoning and 3 JEPA charts inspected individually/contact-sheet; no clipping, broken branch or unreadable terminal node
Generated PNGs	4 teaching assets inspected at original detail; dimensions, SHA-256, grayscale contrast, prompt and scientific boundary are recorded in the corresponding dated research file
Patch hygiene	git diff --check returned no error

Temporary Mermaid renderings and the contact sheet remain outside the repository. This batch does not run model checkpoints, does not constitute R1/R2 reproduction, and does not commit or push.

Batch verification — transport, physics, timeline, reading and bibliography refresh

Final checks on 2026-08-30 covered the six changed documentation pages, five dated audit/research records, generated assets and bibliography infrastructure:

Check	Result
Markdown	markdownlint-cli2 0.23.2 / markdownlint 0.41.1: 11 files, 0 issues
Reference closure	diffusion 24 / 24; flow 20 / 20; physics 41 / 41; no missing, orphan or mismatched anchor
Local links and images	All relative targets in the changed Markdown set resolved; 0 missing
Mermaid	12 blocks across diffusion, flow, physics, reading list and timeline; all include accTitle / accDescr, and all 12 rendered with Mermaid CLI plus system Chrome to non-empty SVG audit artifacts
Timeline media preservation	75 existing HTML images retained; all have non-empty alt text, and the referenced image set is unchanged from HEAD
Generated PNGs	Two 1672×941 teaching assets inspected at original size and in grayscale; SHA-256, final prompts, rejected variants and scientific boundaries are recorded in the dated visual research file
Bibliography integrity	Registry, metadata and BibTeX each contain the same 109 unique identities; JSON parse, Python compile, updater --check and --dry-run all passed
Patch hygiene	git diff --check returned no error

This batch verifies documentation structure and generated artifacts, not model checkpoints. GAUGE, few-step video and closed-loop claims remain author-reported until an independent reproduction is added.

Batch verification — post-training, native AV and fine-grained control

Final checks on 2026-08-30 covered the three new chapters, four dated research/audit records, three generated assets and the shared navigation updates:

Check	Result
Markdown	markdownlint-cli2 0.23.2 / markdownlint 0.41.1: 15 changed/new docs and logs, 0 issues; README retains two pre-existing warnings outside the changed lines
Reference closure	applications 12 refs / 23 citations; foundation 42 / 63; mechanism overview 32 / 48; taxonomy 12 / 22; post-training 38 / 71; control 45 / 98; native AV 33 / 75; no missing, orphan, duplicate or numbering gap
Local links and images	279 relative targets checked across 16 changed/new Markdown files, 139 unique resolved paths, 0 missing
External evidence links	Chapter-level checks covered 40 post-training, 34 native-AV and 48 control URLs; site-policy/TLS/timeout cases were separated from deterministic 404 and re-read through the browser layer
Mermaid	16 blocks across 10 changed Markdown files; all contain accTitle / accDescr, and all 16 rendered with Mermaid CLI 11.12.0 plus system Chrome to non-empty SVG audit artifacts
Mermaid visual check	The revised shared falsification loop and all three new chapter diagrams were rendered to PNG and inspected; the control-protocol and conclusion-downgrade nodes remain distinct, with no clipping or broken branch
Generated PNGs	Three 1672×941 teaching assets inspected at original detail and in grayscale; final SHA-256, prompts, rejected variants and scientific boundaries are recorded in the corresponding dated research files
Patch and secret hygiene	Staged diff whitespace check and changed-file credential-pattern scan passed before commit

Temporary Mermaid and grayscale audit files remained outside the repository. This batch validates educational structure and evidence boundaries; it does not train or independently reproduce the cited video models.

Batch verification — degradation restoration and evidence-scale correction

Final checks on 2026-08-30 covered the restoration chapter and research record, task-boundary corrections, WM/R scale normalization, shared navigation, timeline, reading route and generated asset:

Check	Result
Markdown	markdownlint-cli2 0.23.2 / markdownlint 0.41.1: 15 changed/new docs and logs, 0 issues; README retains two pre-existing warnings outside the changed lines
Reference closure	Restoration 30 refs / 56 citations; no missing, orphan, duplicate or numbering gap
External evidence links	All 30 restoration chapter reference URLs returned HTTP 200 during the final audit
Local links and images	306 relative targets checked across 16 changed/new Markdown files; 0 missing
Mermaid	23 blocks across 13 changed documentation pages; all include accTitle / accDescr, and all 23 rendered with Mermaid CLI 11.16.0 plus system Chrome to non-empty SVG artifacts
Mermaid visual check	The six-branch course route, shared falsification loop, R1–R4 reasoning scale, two WM diagrams and both restoration diagrams were inspected individually/contact-sheet; no clipping or broken branch
Generated PNG	The 1672×941 opaque restoration contract was inspected at original detail and in grayscale; labels, task split and acceptance gates are legible; SHA-256 is recorded in the dated research log
Timeline media preservation	75 existing HTML images retained; all still have non-empty alt text
Patch and secret hygiene	git diff --check and changed-file credential-pattern scan passed

Temporary rendering artifacts remained outside the repository. This batch verifies task classification, sources and educational presentation; it does not train or independently reproduce the cited restoration models.

Batch verification — variational/tokenizer ownership split (historical pre-frontier-refresh snapshot)

These checks record the earlier ownership-split batch on 2026-08-30; the later frontier refresh and its current counts are recorded in a separate section below. This snapshot covered the two independently readable chapters, the tokenizer evidence record, all shared navigation/integration pages and the then-current milestone table:

Check	Result
Markdown	markdownlint-cli2 0.23.2 / markdownlint 0.41.1: 22 changed/new non-README Markdown files, 0 issues; README retains two pre-existing MD001/MD028 warnings
Reference closure	Variational 7 refs / 11 citations; tokenizer 25 / 62; no missing, orphan, duplicate or numbering gap
External evidence links	All 32 reference URLs across the two chapters returned HTTP 200 during the final audit
Local links and images	404 relative targets checked across 23 changed/new Markdown files; 0 missing
Mermaid	35 blocks across the changed/new Markdown set; all include accTitle / accDescr, and all 35 rendered with Mermaid CLI 11.16.0 plus system Chrome to non-empty SVG artifacts
Mermaid visual check	The five-axis overview, both variational diagrams, both tokenizer contracts, both reading-route diagrams and the timeline overview were inspected as PNGs; no clipping, broken branch or unreadable terminal node
Timeline media preservation	75 existing HTML images retained; all still have non-empty alt text; VidTok/V-RAE/VideoRAE/KVAE were added as text rows with provisional status
Patch and secret hygiene	git diff --check and changed-diff credential-pattern scan passed

Temporary rendering artifacts remained outside the repository. This batch verifies classification, current source status, notation and educational presentation; it does not retrain tokenizers, rerun generation benchmarks or independently implement the cited entropy-coded bitstreams.

Batch verification — causal, streaming and real-time contract refresh

Final checks on 2026-08-30 covered the causal/streaming chapter, refreshed evidence record, shared mechanism/foundation/evaluation pages, reading route, timeline and navigation:

Check	Result
Markdown	markdownlint-cli2 0.23.2 / markdownlint 0.41.1: 24 changed/new non-README Markdown files, 0 issues; README retains exactly two pre-existing MD001/MD028 warnings
Reference closure	Causal/streaming 28 refs / 71 citations; every changed ref-based chapter has no missing, orphan, duplicate or numbering gap
Formal/artifact status	Diffusion Forcing, CausVid, Self Forcing, Rolling Forcing, LongLive, SCD, FlowCache, MotionStream and StreamDiffusionV2 refreshed from official proceedings/venue pages; code/weights/placeholder surfaces are recorded separately
Local links and images	417 relative targets checked across 25 changed/new Markdown files; 0 missing
External evidence links	38 unique causal-chapter URLs checked; an OpenReview PDF policy 403 was replaced by its official forum page returning 200; later transient GitHub transport 000 retries were separated from earlier successful/browser reads, with no deterministic 404
Mermaid	39 blocks across the changed/new Markdown set; all contain accTitle / accDescr, and all 39 rendered with Mermaid CLI 11.16.0 plus system Chrome to non-empty SVG artifacts
Causal visual check	Four chapter Mermaid charts rendered to PNG and inspected individually plus color/grayscale contact sheets; schedule, memory lifecycle, commit frontier, backpressure and recovery labels remain legible without color
Timeline media preservation	75 HTML images retained; all still have non-empty alt text; the new causal lineage is text-only
Patch and secret hygiene	git diff --check returned no error; strong credential-pattern scan over changed/untracked files returned no candidate

Temporary Mermaid, PNG and grayscale audit files remain outside the repository. The new StreamFork-1 is a proposal rather than a reported run: the literature-review artifact is depth 4, while checkpoint-level commit/runtime evidence remains absent and must be tracked on the separate reproduction axis.

Batch verification — multi-view/4D camera-time contract

Final checks on 2026-08-30 covered the new chapter and research record, generated camera×time asset, shared task/foundation/evaluation pages, course route and timeline:

Check	Result
Markdown	markdownlint-cli2 0.21.0 / markdownlint 0.40.0: 30 changed/new non-README Markdown files, 0 errors; README retains exactly two pre-existing MD001/MD028 warnings
Reference closure	Multi-view/4D 32 refs / 78 citation occurrences / 32 unique cited references; no missing, orphan, duplicate or numbering gap
Primary-source metadata	All 32 titles, first authors and formal venue/year fields were checked against official proceedings or arXiv; first-public dates are separated from later proceedings status
Local links and images	465 relative links and 99 image targets checked across 31 changed/new Markdown files; 0 missing
External evidence links	All 32 chapter reference URLs returned HTTP 200 after correcting two deterministic 404 targets
Mermaid	Both chapter diagrams plus the updated reading-route and timeline diagrams rendered with Mermaid CLI 11.16.0 and system Chrome to non-empty SVG artifacts; accessibility metadata is present
Visual check	Generated 1536×1024 camera×time PNG and both chapter diagrams passed original-size and grayscale inspection; the wide route diagram has an exact sequential prose fallback for narrow screens
Timeline media preservation	75 existing HTML images retained; all still have non-empty alt text
Patch and secret hygiene	git diff --check and the changed-file credential-pattern scan passed

Temporary render and grayscale audit files remain outside the repository. GridFork-1 is a proposed protocol: no model checkpoint was downloaded or executed, so geometry, view-time consistency, latency, memory and throughput remain literature-grounded rather than independently reproduced.

Batch verification — Video DiT backbone and scaling contract

Final checks on 2026-08-30 covered the new backbone chapter and research record, generated compute-contract asset, shared navigation/evaluation pages, reading route, timeline and bibliography infrastructure:

Check	Result
Markdown	markdownlint-cli2 0.23.2 / markdownlint 0.41.1: 15 changed/new non-README Markdown files, 0 issues; README retains exactly two pre-existing MD001/MD028 warnings
Reference closure	Video DiT backbone 35 refs / 102 citation occurrences / 35 unique cited references; no missing, orphan, duplicate or numbering gap
Local links and images	378 relative references, 151 unique resolved targets and 90 image references checked across 16 changed/new Markdown files; 0 missing
External evidence links	All 36 unique external URLs in the new chapter checked: 34 returned 200; OpenAI and OpenReview returned anti-automation 403 but were browser/index verified; no deterministic 404, 410, 5xx or timeout
Mermaid	26 blocks across the changed Markdown set; all contain accessibility titles/descriptions, and all 26 rendered with Mermaid CLI 11.16.0 plus system Chrome to non-empty SVG artifacts
Mermaid visual check	Six affected reading-route, timeline and backbone diagrams were rendered to PNG and inspected as a contact sheet; topology, ownership and falsification branches remain legible with no clipping
Generated PNG	The 1672×941 RGB compute-contract asset was inspected at original detail and in grayscale; labels, warning and topology distinctions are legible; SHA-256 is recorded in the dated research log
Bibliography integrity	Registry contains 117 core items; JSON parse, Python compile and updater --check passed, with registry, snapshots, BibTeX and generated index consistent
Timeline media preservation	The original 75 HTML image src / alt pairs were retained in the same order; every alt is non-empty and every local target exists
Patch and secret hygiene	git diff --check passed; strong credential-pattern scan over all 21 changed/untracked files returned no candidate; no changed file exceeds 5 MiB

Temporary Mermaid and grayscale audit files remain outside the repository. BackboneFork-1, fixed-weight ServeFork-1 and converted-checkpoint ServeFork-1b are proposed protocols rather than reported runs; no model checkpoint was trained, converted or benchmarked in this batch.

Batch verification — variational stochastic future frontier refresh

Final checks on 2026-08-30 covered the refreshed variational chapter, dated research record, nine shared navigation/evaluation pages, generated teaching asset, bibliography infrastructure and this coverage audit:

Check	Result
Markdown	markdownlint-cli2 0.19.0 / markdownlint 0.39.0: 13 documentation, research and audit files, 0 errors
Reference closure	Variational 31 refs / 49 citation occurrences; evaluation 62 / 81; mechanism overview 32 / 48; recurrent 23 / 44; video prediction 30 / 66; taxonomy 12 / 22; world models 33 / 50; no missing, orphan or duplicate anchor
Local links and images	301 local references checked across the 13-file set, 142 unique resolved paths, 0 missing
External evidence links	40 changed/additional URLs returned HTTP 200 directly; two Google-hosted pages blocked command-line fetches but were read through the browser layer; no deterministic 404
Mermaid	19 blocks across nine files; all contain accTitle / accDescr, and all 19 rendered with Mermaid CLI 11.16.0 plus system Chrome to non-empty artifacts
Visual check	The four new or materially changed diagrams were inspected in color and grayscale; the generated 1536×1024 RGB PNG is readable at original size, its accepted 3:2-versus-requested-16:9 deviation is logged, and SHA-256 is ca309b30ebe193282a52a9d3d2c579ad0710552a32a084148d7cf234f6b52531
Bibliography integrity	Registry, metadata, BibTeX and generated index contain the same 124 unique identities; section M contains nine core entries, 60 unique GitHub repositories match the Star snapshot, Latte's verified TMLR venue is retained through a validated metadata override, Python compile and updater --check passed
Timeline media preservation	The same 75 HTML image src / alt pairs remain in the same order as HEAD; every alt is non-empty and every local target exists
Patch and secret hygiene	git diff --check passed; seven currently changed files contain no strong credential-pattern candidate and none exceeds 5 MiB

This batch validates the literature-review artifact, source/status boundaries, visual accessibility and generated bibliography. It did not train, download or run a video-model checkpoint; LatentFork-1, matched prior-only rollouts and independent calibration evidence remain proposed work on the separate reproduction axis.

Batch verification — open-set single/multi-subject video personalization

Final checks on 2026-08-30 covered the new chapter and research record, generated binding-contract asset, ten cross-page integrations, reading route, timeline, evaluation protocol and bibliography infrastructure:

Check	Result
Markdown	markdownlint-cli2 0.23.2 / markdownlint 0.41.1: 13 changed/new non-README Markdown files, 0 issues; README retains exactly two pre-existing MD001/MD028 findings outside the changed lines
Reference closure	Personalization 35 refs / 86 citation occurrences / all 35 uniquely cited; no missing, orphan, duplicate, mismatched or skipped anchor
Local links and images	307 relative references across 14 Markdown files, 143 unique resolved targets, 0 missing
External evidence links	All 35 chapter URLs checked: 33 returned 200, the TMM DOI returned 202, and the ACM DOI automation 403 was verified from DOI/Crossref metadata; no deterministic 404, 410, 5xx or timeout
Mermaid	20 blocks across 11 affected documentation pages; all contain accTitle / accDescr and all rendered with Mermaid CLI 11.16.0 plus system Chrome to non-empty SVG artifacts
Visual check	Both chapter diagrams and the revised eight-branch reading route passed color/grayscale inspection; the 1672×941 RGB generated asset passed original-size/grayscale inspection and retains SHA-256 856761f1d0fc89c76b90aef4407879d59ab54a28a8d8c1ebcc9161e2d294d5b9
Bibliography integrity	Registry, metadata, BibTeX and index contain the same 132 citekeys; section N has eight entries, all 66 unique GitHub URLs match the Stars snapshot, Python compile and updater --check passed
Timeline media preservation	The same 75 HTML image src / alt pairs remain in the same order as HEAD; all alt text is non-empty and all local image targets exist
Patch and secret hygiene	git diff --check passed; strong credential-pattern scan over all 20 changed/untracked files found no candidate; the 1,991,841-byte teaching PNG is largest and no changed file exceeds 5 MiB

Temporary Mermaid and grayscale audit files remain outside the repository. This batch validates a depth-4 review artifact and executable falsification design; it does not train or run a personalization checkpoint. PersonaBind-1, matched identity-disjoint model comparisons and independent long-horizon binding evidence remain proposed work on the separate reproduction axis.

Definition of done for one subfield

A subfield is not complete because a list of recent papers was appended. It must pass all of these gates:

1. task/mechanism definition and nearest-neighbor concepts are separated;
2. technical route is organized by mechanism, not chronology alone;
3. milestones use a stated criterion and include what remained unsolved;
4. current frontier is verified to the review date and new preprints are labelled provisional;
5. at least one visual materially clarifies the route and passes text/layout/scientific checks;
6. evaluation includes inputs, outputs, metrics, protocol, version, uncertainty and failure cases;
7. strong claims are mapped to the minimum evidence required;
8. primary sources, search date and exclusions are recorded in sources/;
9. links, anchors, image paths, Mermaid syntax and repository checks pass;
10. no author-reported performance is presented as independent confirmation.