Research audit: causal, streaming and real-time video generation

This file records the search and verification trail behind docs/generative-models/causal-streaming-generation.md.

Scope

• Review date: 2026-08-30 (Asia/Shanghai; refreshed from the original 2026-08-29 audit)
• Primary question: How did video generation move from offline bidirectional diffusion to causal, few-step, long-horizon and deadline-aware streaming systems, and which claims belong to codec, generator, commit or runtime layers?
• Subquestions: causal codec/generator separation, training-history exposure, teacher–student architectural mismatch, few-step versus measured NFE, streaming commit/revision, KV/resident-memory scaling, open-horizon drift, online serving/backpressure, interactive condition changes and world-model negative classification.
• Time window: 2024-01-01 through 2026-08-30, with older work included only when required to explain a mechanism.
• Evidence policy: peer-reviewed paper or official proceedings first; arXiv paper for new work without proceedings; project pages and model cards only for implementation or availability claims. Secondary surveys were not used as evidence for method claims.

Search strategy

Queries combined the following phrases and close variants:

"autoregressive video diffusion"
"causal video diffusion"
"streaming video generation"
"real-time interactive video generation"
"rolling KV cache" video diffusion
"self forcing" video
"causal forcing" video
"long-horizon" autoregressive video diffusion
"streaming commit" video generation lookahead revision
"first frame latency" video diffusion deadline jitter
"backpressure" real-time video generation
"open horizon" video generation constant memory
"separable causal diffusion" video
"flow cache" autoregressive video
"motion stream" interactive video generation

The search was deliberately broader than the final chapter. API result counts below are raw responses before deduplication and eligibility screening, not PRISMA-style included-study counts.

Source	Search use	Raw response observed	Notes
arXiv search/API	exact phrases, title and abstract discovery	65 for the main exact-phrase query	Primary full records and submission histories used for preprints
OpenAlex	citation-neighborhood and synonym discovery	8,217 for the broad query	Very noisy; used only to find candidates, never as the source of a technical claim
Crossref	venue and DOI cross-check	110,507 for the broad query	Very noisy; useful for proceedings metadata
DBLP	publication-title and venue discovery	58 for the focused publication query	Used to find 2026 causal/streaming papers, then verified at the primary paper
MLSys proceedings	official venue verification	direct paper lookup	Used for StreamDiffusionV2
CVF / NeurIPS / OpenReview	official venue verification where available	direct paper lookup	Used when the paper had a formal proceedings version
Semantic Scholar API	attempted metadata/citation lookup	HTTP 429	Rate-limited; no claims or counts from this API were used

Inclusion criteria

A work was included when it changed or carefully evaluated at least one of the following:

1. causal/frame-wise/chunk-wise factorization of video generation;
2. self-generated-history training or teacher–student causal distillation;
3. few-step sampling needed for streaming latency;
4. bounded, compressed, recurrent, hierarchical or retrieved long-term memory;
5. time-to-first-frame, frame deadlines, jitter or multi-GPU streaming serving;
6. interactive prompt/action changes during rollout;
7. explicit streaming commit, lookahead, revision, deadline or recovery contracts;
8. long-horizon geometry, motion or source-preservation objectives specific to streaming AR models.

Exclusion criteria

• image-only streaming diffusion without a temporal video model;
• generic video acceleration that did not address causal or streaming generation;
• product announcements without a technical paper or system card;
• duplicate arXiv versions after the latest revision had been checked;
• papers found only in aggregators when the primary paper could not be verified;
• a speed claim lacking enough context to state at least its hardware or model setting.
• a product or project label of "real-time" without a paper/system card or reproducible timing boundary;
• causal codec work whose only contribution was representation compression; it remains owned by the tokenizer audit unless it also changed the generator/serving contract.

Primary-source ledger

Foundations and first causal/few-step line

Work	Primary source	Status used in chapter	Reason included
Diffusion Forcing	https://proceedings.neurips.cc/paper_files/paper/2024/hash/2aee1c4159e48407d68fe16ae8e6e49e-Abstract-Conference.html	NeurIPS 2024	Independent per-token noise; not self-rollout, few-step or a commit/SLO guarantee
CausVid	https://openaccess.thecvf.com/content/CVPR2025/html/Yin_From_Slow_Bidirectional_to_Fast_Autoregressive_Video_Diffusion_Models_CVPR_2025_paper.html	CVPR 2025	Bidirectional teacher to 4-step causal student; history is mainly GT/noised GT, not on-policy matching
Self Forcing	https://proceedings.neurips.cc/paper_files/paper/2025/hash/f4823f831af67a3ef15e41a85434422a-Abstract-Conference.html	NeurIPS 2025	On-policy autoregressive rollout on self-generated history; detached history leaves a context-gradient gap
LongLive	https://proceedings.iclr.cc/paper_files/paper/2026/hash/91a1610c6ed9e02d33f826b46f472b92-Abstract-Conference.html	ICLR 2026; first preprint 2025	Prompt recache, frame sink, train-long-test-long and author-reported single-H100 speed
Rolling Forcing	https://openreview.net/forum?id=IAyzXjbfwo	ICLR 2026; first preprint 2025	Joint progressive-noise window, attention sink and long-window distillation; window is not strictly frame-serial
StreamDiffusionV2	https://proceedings.mlsys.org/paper_files/paper/2026/hash/698cfaf72a208aef2e78bcac55b74328-Abstract-Conference.html	MLSys 2026	SLO-aware online system, TTFF, deadline scheduling and multi-GPU pipeline
Separable Causal Diffusion	https://openaccess.thecvf.com/content/CVPR2026/html/Bai_Causality_in_Video_Diffusers_is_Separable_from_Denoising_CVPR_2026_paper.html	CVPR 2026	Once-per-frame causal encoder plus multi-step frame-wise diffusion renderer
FlowCache	https://proceedings.iclr.cc/paper_files/paper/2026/hash/85dc8f85ff978b9c606d3b2f5b0da69a-Abstract-Conference.html	ICLR 2026	Training-free per-chunk feature-cache decisions plus bounded resident KV compression
MotionStream	https://proceedings.iclr.cc/paper_files/paper/2026/hash/0cece806cd3d1dfad4a893f016ad3d7d-Abstract-Conference.html	ICLR 2026	Online motion/camera controls with Self Forcing, sinks and rolling KV

Distillation and train–test alignment

Work	Primary source	Status used in chapter	Distinct claim verified
Causal Forcing	https://arxiv.org/abs/2602.02214	ICML 2026 official accepted-paper list; PMLR page not located at freeze date	Frame-level injectivity problem and AR-teacher ODE initialization
Causal Forcing++	https://arxiv.org/abs/2605.15141	arXiv 2026	Causal consistency initialization for frame-wise 1–2-step generation
Causal-rCM	https://arxiv.org/abs/2606.25473	technical report 2026	Teacher-forcing consistency plus self-forcing DMD recipe
Video-Mirai	https://arxiv.org/abs/2606.03971	arXiv 2026	Training-only foresight supervision for causal representations
Self Gradient Forcing	https://arxiv.org/abs/2607.20368	arXiv 2026	Two-pass restoration of future-to-context KV gradients

Long-horizon memory and efficiency

Work	Primary source	Status used in chapter	Memory design
FAST-AR	https://arxiv.org/abs/2602.01801	ICML 2026 accepted; technical claims from arXiv	Temporal cache compression plus ANN sparse cross/self-attention
Quant VideoGen	https://arxiv.org/abs/2602.02958	ICML 2026 accepted; technical claims from arXiv	Training-free 2-bit KV-cache quantization
Light Forcing	https://arxiv.org/abs/2602.04789	ICML 2026 accepted; technical claims from arXiv	Chunk-aware hierarchical sparse attention
Rolling Sink	https://arxiv.org/abs/2602.07775	arXiv 2026	Training-free cache maintenance beyond the training horizon
Sparse Forcing	https://arxiv.org/abs/2604.21221	arXiv 2026	Trainable persistent block-sparse attention
Forcing-KV	https://arxiv.org/abs/2605.09681	arXiv 2026	Head-specialized static and dynamic KV pruning
ARL2	https://arxiv.org/abs/2605.16579	arXiv 2026	Intra-frame softmax plus fixed recurrent cross-frame state
VideoMLA	https://arxiv.org/abs/2605.30351	arXiv 2026	Shared low-rank content cache and decoupled 3D-RoPE key
LongLive-RAG	https://arxiv.org/abs/2606.02553	arXiv 2026	Content-addressable retrieval over self-generated latent history
FadeMem	https://arxiv.org/abs/2606.10671	arXiv 2026	Dense-near, sparse-far hierarchical consolidation under fixed budget
VideoSSM	https://arxiv.org/abs/2512.04519	arXiv 2025	Global state-space memory plus local context window

Geometry and task extensions

Work	Primary source	Status used in chapter	Extension
Stream4D	https://arxiv.org/abs/2608.19556	arXiv 2026, submitted 2026-08-20	Dynamic 4D reconstruction reward and motion prior; treated as provisional
MV-Forcing	https://arxiv.org/abs/2607.05376	arXiv; authors report ECCV 2026, official proceedings page not located	Temporal and view-wise autoregression bridged by 4D geometry
JoyAI-Video-Edit	https://arxiv.org/abs/2608.03974	arXiv 2026	Open-ended causal video editing and source-anchored distillation

Release-surface and timing-boundary audit

Work	Code / weights status checked 2026-08-30	Timing or NFE boundary retained in chapter
Diffusion Forcing	https://github.com/buoyancy99/diffusion-forcing; code/checkpoints and a paper branch are public	No inherent few-step or real-time claim
CausVid	https://github.com/tianweiy/CausVid; training, inference and weights public	640×352, 120 frames/10 s, 4 steps, one H100; 9.4 FPS and TTFF 1.3 s include text encoder, DiT and VAE
Self Forcing	https://github.com/guandeh17/Self-Forcing; code, checkpoints and training public	Wan2.1-1.3B, 832×480, 4 steps, one H100; chunk/frame results are not interchangeable
Rolling Forcing	https://github.com/TencentARC/RollingForcing; code, checkpoint and training public	Reported 0.76 s is steady-state latency, not TTFF; GPU/VAE boundary is insufficient for cross-paper comparison
LongLive	https://github.com/NVlabs/LongLive; code and weights public	1.3B, 832×480, one H100, 20.7 FPS and 240 s demonstration; paper claims map to v1.0, not later v2 infrastructure
Causal Forcing / ++	https://github.com/thu-ml/Causal-Forcing; code/config/checkpoints public	CF++ keeps the first latent frame at 4 steps; 1/2-step modes apply to later frames, and the A800 timing excludes VAE
Causal-rCM	https://github.com/NVlabs/rcm; recipe/code public, checkpoint completeness stated conservatively	Clean-context 4/2/1 steps require 5/3/2 NFE because context encoding adds a forward
SCD	CVF paper/supplement/project page public; no verifiable official code/checkpoint located	832×480, batch 1, one H100; 11.1 FPS/0.29 s is encoder-once plus multi-step renderer, not few-step distillation
FlowCache	https://github.com/mikeallen39/FlowCache; MAGI-1 and SkyReels-V2 code public	2.38× and 6.7× are A800 offline speedups from very long baselines, not a real-time FPS/TTFF result
MotionStream	https://github.com/alex4727/motionstream; repository explicitly says code remains under internal review, no runnable weights	16.7 FPS uses the full VAE; 29.5 FPS depends on a separately trained Tiny VAE
StreamDiffusionV2	https://github.com/chenfengxu714/StreamDiffusionV2; inference/PyPI/checkpoint public; training and parts of scheduler remain TODO	4×H100 pipeline figures are aggregate system evidence, not one-GPU generator latency

Formal acceptance, an open repository, downloadable weights and independent reproduction were treated as four separate evidence fields. A repository placeholder, branch with only inference, or README timing table was not upgraded to a reproduced result.

Candidate works intentionally not promoted to milestones

The search also surfaced FramePack, SkyReels-V2, MAGI-1, LongVie 2 and many 2026 cache/forcing variants. They remain important context, but the chapter's milestone table uses a stricter test: a milestone must change the problem definition, causal factorization, training distribution, memory abstraction or deployment protocol. A higher resolution, longer demo or another variant of an existing sparse-cache recipe is not by itself a milestone.

Primary records checked for this contextual group:

• FramePack: https://proceedings.neurips.cc/paper_files/paper/2025/hash/2bde8fef08f7ebe42b584266cbcfc909-Abstract-Conference.html; fixed context does not make it real-time, and anti-drift/inverted ordering is not always strictly causal
• SkyReels-V2: https://arxiv.org/abs/2504.13074
• MAGI-1: https://arxiv.org/abs/2505.13211; reported streaming uses a 24-GPU pipeline and separates/overlaps the VAE, so it is not one-GPU latency evidence
• LongVie 2: https://arxiv.org/abs/2512.13604

Claim-handling rules

1. Performance numbers are introduced as author-reported, not independently replicated.
2. FPS is not compared across different GPU counts, resolutions, precisions, model sizes or denoising steps.
3. A long-duration demo is not evidence of real-time execution.
4. Dynamic prompting is not evidence of action-conditioned world modeling.
5. Causal attention is an information constraint, not proof of causal physical reasoning.
6. New August 2026 preprints are marked provisional and are not used to rewrite older peer-reviewed conclusions.
7. Causal codec, causal generator, streaming commit and real-time SLO are separate contracts; success at one layer is not inherited by the next.
8. Sampling-step labels are not treated as NFE until clean-context encoding, CFG branches and recomputation are counted.
9. Bounded resident KV/GPU memory is not described as lossless long-term memory or constant total system cost; CPU, external index and retrieval latency remain separate.
10. Author priority claims such as "first" are omitted unless an exhaustive search could support them; the chapter instead states the concrete capability placed at the center of the paper.

Proposed falsification package

The chapter now defines StreamFork-1, but the package is explicitly marked proposal, not run. Its four required probes are:

1. equal prefix/seed with perturbed hidden suffix or future condition, checked by per-commit hashes outside the declared revision window;
2. matched GT/noised-GT/self-rollout histories at equal backbone, teacher, cache and measured NFE, evaluated at 1×/2×/6×/12× training horizon;
3. 5 s/1 min/4 min resident-memory and long-recall traces with object-return, scene-cut, small-object and cache-poisoning probes;
4. end-to-end cold/warm and single/concurrent timing including text encoder, DiT, codec/VAE, queue, transport and display, with deadline recovery and at least a 60 s soak.

The completion surface is a frozen manifest, commands/environment lock, raw trace.jsonl, commits.csv and hashes, NFE hook log, all videos/failures and survival/memory/latency plots. A selected demo or copied author table does not complete the experiment.

Figure generation and verification

The chapter overview figure was generated with the built-in image-generation tool after the scientific-schematics script could not run because OPENROUTER_API_KEY was not configured.

• Workspace asset: assets/diagrams/causal-streaming-video-generation.png
• Dimensions: 1672 × 941
• SHA-256: 41a0f7564b40803164b92af3938c8a933f16d89e16fe5676d724d046db9bc079
• Visual checks: all 11 requested labels rendered correctly; offline full-clip/bidirectional arrows are distinct from the causal chunk loop; the rolling memory is visibly bounded; no logo, watermark, extra text or malformed typography was observed.

Exact mechanism relationships were added as editable Mermaid rather than another generated raster:

1. five-layer causal/streaming stack;
2. next-unit versus full-sequence versus rolling per-token noise and commit frontier;
3. bounded generator-memory write/age/compress/retrieve/drop/read lifecycle;
4. speculative generation → decode/gate → immutable commit/hash → display, with backpressure and reset branches.

All four contain accTitle and accDescr. Mermaid CLI 11.16.0 with system Chrome rendered them to non-empty SVG and PNG; the inspected PNG sizes were 936×726, 671×1932, 1584×828 and 1324×1189. Individual color renders and a grayscale contact sheet preserved every branch label, terminal state and arrow meaning; the diagrams do not rely on color alone. The rolling-noise chart deliberately uses a tall reading order so the A/B/C schedules remain legible instead of shrinking three mechanism panels into one wide strip.

Documentation validation

The final causal/streaming integration batch produced the following checks:

• Markdown: 24 changed/new non-README files, 0 issues with markdownlint-cli2 0.23.2 / markdownlint 0.41.1; README retains exactly two pre-existing warnings at its historical heading/blockquote locations.
• Reference closure: causal chapter 28 references / 71 citation occurrences; no missing, orphan, duplicate or gapped reference anchors. The same check passed for every changed ref-based chapter.
• Local links/images: 417 relative targets across 25 changed/new Markdown files; 0 missing.
• Mermaid: 39 blocks across the changed/new Markdown set; all 39 contain accessibility metadata and render to non-empty SVG.
• Causal visuals: all four Mermaid charts inspected individually and in color/grayscale contact sheets; no clipped labels, broken terminals or color-only distinction.
• Timeline preservation: 75 HTML images retained and all 75 have non-empty alt text.
• External URLs in the causal chapter: 38 unique links. The first pass returned HTTP 200 for 37 and a policy 403 only for an OpenReview PDF endpoint; that link was replaced by its official forum page, which returned 200. A later rapid GitHub-only retry produced transport code 000 after those same repository pages had already returned 200 and had been read through the browser layer; no deterministic 404 remained.
• Patch hygiene: git diff --check returned no error.

These checks validate documentation structure and source accessibility, not model behavior. StreamFork-1 remains a proposed experiment; no checkpoint, runtime trace or independent speed/long-horizon reproduction was created in this batch.

Review limitations

• This is a focused scoping review, not a formal meta-analysis; the field is too new and the evaluation setups are not homogeneous.
• Some 2026 works had only arXiv/technical-report versions at the review date; formal status was upgraded only for works with an official venue page or accepted-paper list.
• OpenAlex and Crossref counts were too broad to serve as screening statistics.
• Semantic Scholar rate limiting prevented its use as the third metadata cross-check; DBLP, Crossref and official proceedings were used instead.
• Independent reproduction of speed, commit correctness and long-horizon quality claims was outside this repository review; the chapter therefore supplies StreamFork-1 and a fair-reporting protocol rather than a ranking.