Pith. sign in

REVIEW 4 major objections 7 minor 22 references

Future frames can train how a video model writes its memory, not just how it reads it.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 10:00 UTC pith:ALRC2VGN

load-bearing objection SGF is a plausible, honestly written fix for a real training-signal gap, but the load-bearing replay-fidelity assumption is only verified in forward space and the closest prior is structurally closer than the main text suggests. the 4 major comments →

arxiv 2607.20368 v2 pith:ALRC2VGN submitted 2026-07-22 cs.CV

Self Gradient Forcing: Native Long Video Extrapolation

Junhao Zhuang , Shiyi Zhang , Yuxuan Bian , Yaowei Li , Yawen Luo , Yijun Liu , Weiyang Jin , Songchun Zhang
show 6 more authors
Xianglong He Xuying Zhang Haoran Li Haoyang Huang Zeyue Xue Nan Duan
This is my paper
classification cs.CV
keywords autoregressive video diffusionlong video extrapolationself forcingKV cachememory writing gradienttwo-pass trainingcontext-gradient reconstructiontemporal consistency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper targets a subtle blind spot in autoregressive video generation: when a model generates video frame by frame, it builds a key–value memory cache of its own past output, and later frames read that cache as history. In standard self-forcing training, losses on future frames teach the model how to read that cache but never how to write it, because the cache is frozen rollout state. The authors call this the historical context-gradient gap and argue that over long rollouts the cache-writing computation drifts, producing identity drift, scene breaks, and layout collapse. Their method, Self Gradient Forcing, runs a gradient-free rollout to record what the model generated, then re-runs just the sampled step in parallel with gradients flowing through the recomputed cache-writing path. If the claim holds, a model trained on five-second clips can extrapolate coherently to several-minute videos.

Core claim

Self Gradient Forcing restores a credit-assignment path that self forcing leaves cut. In autoregressive video diffusion, the model denoises each new latent while attending to a key–value (KV) cache built from its own previously generated clean latents. The cache is written at the clean context timestep, and in self forcing that writing computation is treated as frozen state: future losses can train the model to better read historical memory, but cannot propagate back into how that memory was encoded. SGF performs a no-gradient rollout to collect self-generated context and noisy exit latents, then reconstructs the same exit computation in parallel, this time allowing losses on future video la

What carries the argument

The central object is the historical context-gradient gap and the two-pass reconstruction that closes it. Pass 1 is an ordinary no-gradient autoregressive rollout that records, at a randomly sampled denoising exit step, both the self-generated clean latent context and the noisy latents fed to the model. Pass 2 discards the serial cache and recomputes that exit step in parallel: the recorded context latents are re-encoded at the clean context timestep with stop-gradient on the latents themselves, but with gradients kept through the K/V projections and the causal future-to-context attention under a reconstructed mask that mirrors the sink-plus-FIFO streaming cache. This fixed-window replay tur

Load-bearing premise

The whole method rests on the assumption that the parallel Pass-2 replay faithfully reproduces the serial attention relation used at the sampled rollout step, so that the recovered gradients train the same cache writer that operates at inference.

What would settle it

Train SGF exactly as described, then rerun Pass 2 with the context K/V path explicitly stop-gradient, making the reconstruction identical to frozen-cache self forcing. If long-horizon metrics stay at the SGF level rather than reverting to the self-forcing level, the gains are not coming from the memory-writing gradient. Alternatively, perturb the Pass-2 reconstruction mask or sink positions by a few latents; if 60s/240s consistency does not degrade, the method is not actually supervising the cache writer at the sampled exit step.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Training on a 5-second window can extrapolate to 60-second and 240-second videos, so long-horizon consistency does not require a proportionally long training window.
  • Subject identity, background/layout consistency, and temporal stability improve over matched self-forcing baselines, while 5-second short-horizon quality stays on par.
  • The benefit appears in both frame-wise and chunk-wise autoregressive generation, and under teacher-forcing, consistency-distillation, and ODE initializations.
  • SGF is a drop-in gradient-boundary change, not a new architecture, so it can be stacked with existing forcing improvements and context/cache designs.
  • Restoring the memory-writing gradient costs modest extra memory and runtime (about 10% wall-clock in the reported configuration), whereas a directly differentiable cache runs out of memory.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same argument should transfer to any autoregressive sequence model with a KV cache: if future losses cannot reach the writing computation, long-horizon drift should appear there too, and SGF-style replay could be tested on language or world models.
  • If the gap widens as shared parameters shift under few-step distillation, SGF should matter most with short denoising schedules and less with full-step training — a quantitative prediction implied by the paper's mechanism.
  • One could ablate replay fidelity directly by perturbing the Pass-2 reconstruction mask or sink positions; if long-horizon gains collapse while short-horizon metrics hold, the memory-writing supervision is doing the work rather than the extra forward pass.
  • Because Pass 2 is a bounded surrogate, its benefits may saturate as rollout lengthens; combining SGF with longer-context retrieval or recurrent memory could extend write supervision to memory evicted from the FIFO window.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. The paper proposes Self Gradient Forcing (SGF), a two-pass training strategy for autoregressive video diffusion built on Self Forcing. Pass 1 performs a no-gradient serial self-rollout and records sampled noisy latents and predicted clean latents at a sampled denoising exit step; Pass 2 reconstructs the exit-step computation in parallel, re-encoding the recorded context latents at the clean timestep with stop-gradient inputs but permitting gradients through the context KV projections and future-to-context attention, so DMD losses supervise the cache-writing computation without full rollout BPTT. The paper identifies the 'historical context-gradient gap' in frozen-cache Self Forcing and claims that SGF restores this memory-writing supervision, improving long-horizon consistency at 60s and 240s while matching at 5s. Experiments compare matched Self Forcing baselines under frame-wise and chunk-wise settings with TF, causal-CD, and causal-ODE initializations, using VBench-Long metrics, a GSB study, and qualitative strips; an appendix verifies forward recovery fidelity of Pass 2; a feasibility table documents memory/runtime overhead.

Significance. If the central claim holds, the contribution is valuable: it isolates a plausible and previously under-specified credit-assignment problem in frozen-cache self-rollout training and proposes a computationally bounded, implementation-friendly fix that is orthogonal to existing forcing objectives and long-context mechanisms. The paper ships unusually strong empirical infrastructure for this area: controlled matched-pair comparisons (same prompt, seed, initialization, inference geometry), a large GSB study (>1,900 judgments), a Pass-1/Pass-2 recovery diagnostic with error scaled against bf16 precision, and a direct-differentiable-cache feasibility measurement (OOM) that motivates the bounded surrogate. These assets make the empirical claim materially more credible than a purely qualitative report. However, the load-bearing mechanism—that Pass 2's gradients train the same causal memory writer used at inference—is verified only at the level of forward latent recovery, not at the level of gradient identity or the attention relation; the Limitations section (G) concedes this. The main experimental tables also lack error bars and significance tests, so some gains are on the order of small m

major comments (4)
  1. [§3.3, Algorithm 1, Appendix D, Limitations G] The central mechanistic claim—that future DMD losses supervise the clean-context K/V writer used at inference—rests on the assumption that Pass 2's parallel reconstruction reproduces the exact causal computation of Pass 1's serial cache, including the backward path through attention and K/V projections. Appendix D verifies forward recovery only (cosine similarity ~0.9999, RelL2 ~1.4%); Limitations G explicitly concedes that a mask, sink, RoPE, timestep, or chunk-alignment deviation would make SGF 'train a different writer.' Forward agreement does not imply backward agreement: small forward differences can be amplified by the backward pass through attention and KV projections, and a mismatched mask could leave forward outputs close while producing gradients for a different relation. The paper should either provide a direct gradient-alignment diagnostic (e.g., gradient cosine similarity be
  2. [Tables 1-2, §4.2] Tables 1 and 2 report no error bars, confidence intervals, or significance tests, and many SGF-vs-SF differences are small (e.g., frame-wise 60s TF-initiated Background 0.970→0.974; CD-initiated Flickering 0.992→0.992). Given the paper's emphasis on 'especially consistent' gains, the reader cannot tell which differences exceed run-to-run noise. Reporting seed-level variance or significance over the 128 prompts would materially strengthen the claim that SGF improves long-horizon metrics, beyond the human study.
  3. [§4.2 and Appendix H] The dynamic-degree results appear to go against the paper's own conclusion: SGF is claimed to improve temporal stability and motion quality, yet dynamics drop substantially in several controlled rows (e.g., frame-wise 60s causal-ODE 0.867→0.566; 240s 0.840→0.556; chunk-wise 240s TF 0.943→0.566). The paper attributes SF's higher dynamics to 'incoherent motion' from scene jumps and deformation (Appendix H), but provides no controlled quantification of this explanation—no correlation between dynamic-degree and human judgments, no perceptual evaluation of the specific videos, and no alternative motion-quality metric. Since dynamics is one of the headline metrics in the tables, this is a load-bearing inconsistency that needs a concrete diagnostic, not just an appeal to appendix examples.
  4. [§4.2, Tables 1-2 and §3.5, Appendix A.1] There is a notable gap between the training context and the long-horizon evaluation context. Frame-wise training uses the full 5-second window with a standard teacher-forcing-style causal mask, while long-horizon inference uses a streaming sink-4/FIFO-16/current-1 window with a different attention geometry (Appendix A.1). The paper asserts the Pass-2 mask 'matches' the sampled rollout, but the sampled rollout is a training-window full-context rollout, not the inference geometry. Since SGF's value proposition is improving the cache writer used at inference, the paper should explicitly state whether SGF's Pass-2 mask is aligned to training rollout or to the inference streaming geometry, and justify why a mask that matches only the training rollout transfers to the streaming policy. This is exactly the alignment risk flagged in Limitations G.
minor comments (7)
  1. [Throughout] Reference dates include 2026 entries and the arXiv ID (2607.20368) is in the future relative to standard CVPR/ICML review cycles; the paper should ensure the bibliography is complete and correctly dated for the venue version.
  2. [§3.3, Eq. (4)] Equation (4) writes the gradient inclusion symbolically; it would be clearer to state explicitly which terms are stop-gradient (context latents, noise samples, scheduler states) and which are not, ideally in the equation or immediately after it.
  3. [Appendix D, Table 7] The recovery metrics are computed on latent predictions before VAE decoding; the claim 'visual pairs are nearly indistinguishable' is fine, but the text should clarify that decoded comparisons are illustrative, not a formal perceptual metric.
  4. [§4.5] The feasibility table reports a 5-step cycle time and peak memory on one hardware setup; specifying the exact GPU type, batch size, and model size would make the numbers reproducible.
  5. [§4.3] The GSB study should state how many raters, how many pairs per comparison, and the inter-annotator agreement; currently it reports only the total number of paired judgments and the aggregate score.
  6. [§4.1] The 240s setting uses 128 MovieGen prompts, but the 60s setting is described only as 'VBench-Long prompts'; the number of prompts per condition should be stated for both.
  7. [Appendix F] The added discussion of Checkpointed Self Forcing in Solaris is appreciated and appropriately positioned as related work; the text should also cite it in the related-work section, not only in an appendix.

Circularity Check

0 steps flagged

No circular derivation: SGF is an operationally defined training objective and its claimed long-horizon benefit is evaluated on held-out benchmarks against matched baselines.

full rationale

The claimed derivation is self-contained rather than circular. SGF is defined operationally in Algorithm 1 and Eqs. (3)-(4): Pass 1 records a no-gradient rollout state, and Pass 2 recomputes the exit-step prediction with gradients flowing through the clean-context K/V path. The paper's central empirical claim that this improves 60s/240s extrapolation is not an input to that definition; it is evaluated on VBench-Long and MovieGen-128 prompts against matched Self Forcing baselines under fixed seeds, prompts, and inference geometry (Sections 4.2-4.3). The two-pass recovery fidelity check (Appendix D) is an internal sanity test of the replay construction, not a fitted prediction. The assumption that the parallel reconstruction faithfully reproduces the serial context relation is explicitly flagged as a limitation (Limitations G), not asserted as a theorem. The only self-references (Zhuang et al. 2026; Luo et al. 2026) appear in related-work context and do not carry the argument. The acknowledgment that Solaris (Savva et al. 2026) contains a closely related checkpointed parallel recompute (Appendix F) is a novelty disclosure, not a circular step. No quantity is defined in terms of the metric it purports to improve, and no prediction reduces to its own input by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The method adds no new fitted physical constants and no invented entities. Its load-bearing choices are the shared-parameter drift assumption, the replay-fidelity assumption, and a set of streaming hyperparameters (sink/FIFO/chunk) fixed through evaluation-side ablations.

free parameters (3)
  • Sink latents (frame-wise) = 4
    Chosen via VAE boundary diagnostic and a 60s sink ablation (Appendix E, Table 8); sink 4 enters the stable range, but the selection uses the same 60s evaluation family as the main results.
  • FIFO window / total context (frame-wise) = FIFO 16, total window 21
    Streaming context policy shared by both baselines; fixed to isolate SGF, but it determines which history is available and is not independently justified.
  • Chunk-wise context policy = sink 3, FIFO 6, chunk size 3
    Coarser memory-update granularity selected for chunk-wise experiments; no ablation is reported for chunk size or sliding-window width.
axioms (4)
  • domain assumption A causal DiT shares parameters between noisy denoising steps and the clean t_ctx=0 cache-writing computation, so updates at noisy steps can drift the cache writer (Eq. 2).
    Section 3.1 uses this parameter-sharing drift as the motivation for restoring cache-writing gradients; it is a mechanistic claim about training dynamics, not proven.
  • domain assumption Pass-2 parallel reconstruction faithfully reproduces the serial cache relation at the sampled exit step (mask M_rec, positional indices, deterministic layers).
    Section 3.3 and Limitations G; only checked empirically in Appendix D at the latent level. If mask/RoPE/chunk alignment deviates, SGF may train a different writer.
  • domain assumption Distribution-matching distillation (DMD) from a bidirectional video model is a valid supervision signal for self-rollout training.
    Inherited from Self Forcing (Huang et al., 2025); SGF uses DMD losses in Pass 2 without re-validating the distillation objective.
  • domain assumption The Wan video VAE requires a 4-latent sink to cover its start-boundary group, and this boundary diagnostic transfers to training/evaluation.
    Appendix E measures local re-encoding mismatch on 10 videos and motivates sink=4; this is an empirical domain assumption for streaming generation.

pith-pipeline@v1.3.0-alltime-deepseek · 21593 in / 11354 out tokens · 101937 ms · 2026-08-01T10:00:23.219295+00:00 · methodology

0 comments
read the original abstract

Recent autoregressive video diffusion methods are increasingly built upon Self Forcing, where the student is trained on histories produced by its own rollout rather than ground-truth video contexts. This reduces exposure bias, but the historical key-value cache is still used by future frames only as frozen rollout state. As a result, future losses cannot supervise how earlier generated latents should be written into more useful keys and values for later video-latent generation. We call this the historical context-gradient gap. We propose Self Gradient Forcing (SGF), a two-pass training strategy that restores this missing supervision signal without backpropagating through the full serial rollout. Pass 1 performs a no-gradient autoregressive rollout matching inference and, at a sampled denoising exit step, records both the self-generated context and the noisy latents fed to the model. Pass 2 performs parallel context-gradient reconstruction for the recorded exit step. The generated context is used as stop-gradient clean-latent input, while the model recomputes the context KV representations and future-to-context causal attention. Thus, SGF provides the missing memory-writing supervision within the native autoregressive training objective, using losses on future video latents to train the model to encode context into more effective causal memory. Across extensive long-horizon frame-wise and chunk-wise experiments under different initializations, SGF achieves stronger native long-video extrapolation than Self Forcing, especially in subject identity, background/layout consistency, and temporal stability. Remarkably, using only a 5-second training window, SGF can extrapolate to videos lasting several minutes. Code and models will be released to advance research on autoregressive video generation.

Figures

Figures reproduced from arXiv: 2607.20368 by Haoran Li, Haoyang Huang, Junhao Zhuang, Nan Duan, Shiyi Zhang, Songchun Zhang, Weiyang Jin, Xianglong He, Xuying Zhang, Yaowei Li, Yawen Luo, Yijun Liu, Yuxuan Bian, Zeyue Xue.

Figure 1
Figure 1. Figure 1: Long-horizon consistency with Self Gradient Forcing. Under the same prompt, seed, and TF initialization, Self Forcing trains with self-generated history but does not supervise the clean￾timestep KV writing computation with future losses; it eventually exhibits view jumps, scene breaks, and identity drift. SGF restores a bounded context-gradient path for self-generated memory writing and better preserves su… view at source ↗
Figure 2
Figure 2. Figure 2: From frozen-cache Self Forcing to Self Gradient Forcing. Self Forcing trains on self-generated histories but treats historical K/V entries as detached cache state. SGF keeps the no￾gradient rollout unchanged and adds a parallel reconstruction pass, where detached context latents are re-encoded at the clean context timestep so future DMD losses supervise K/V writing without serial rollout backpropagation. A… view at source ↗
Figure 3
Figure 3. Figure 3: Frame-wise 240-second comparison under TF initialization. Both rows use matched prompts, random seeds, initialization, horizon, and inference geometry. 21, FIFO 16, and current chunk 1; for chunk-wise generation, we use sink 3, total window 12, FIFO 6, current chunk 3, and chunk size 3. The 5s evaluation follows the standard VBench protocol and serves as a short-horizon sanity check. Full 5s VBench results… view at source ↗
Figure 4
Figure 4. Figure 4: Decoded Pass-1 and Pass-2 recovery comparison. Rows show paired decoded outputs from Pass 1 and Pass 2 at exit steps 1000, 750, 500, and 250. The decoded pairs are visually nearly indistinguishable, matching the latent-space recovery results in Tab. 7. 0 1 2 3 4 5 6 7 0 n n+1 n+2 n+3 n+4 n+5 n+6 0 1 2 3 n n+1 n+2 n+3 Train Infer sink = 1 Infer sink = 4 … … … Local encoding mismatch Relative L2 to full-pref… view at source ↗
Figure 5
Figure 5. Figure 5: VAE boundary and sink-latent choice. Left: sink 1 preserves only the stream-start an￾chor, while sink 4 preserves the short boundary-transition prefix. Right: local re-encoding mismatch to the full-prefix 21st latent drops sharply after adding previous latent groups and plateaus around the first four positions. This motivates using four sink latents for frame-wise streaming inference. size preserves more p… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative sink ablation at 60 seconds. Very small sinks preserve too little early stream￾boundary information, while a multi-latent sink better maintains subject and scene consistency dur￾ing long extrapolation. The ablation supports the selected frame-wise streaming policy; the core SGF claim remains the recovery of gradients through self-generated context K/V formation. F ADDITIONAL RELATED-WORK DISCUS… view at source ↗
Figure 7
Figure 7. Figure 7: Frame-wise 60-second comparison under TF initialization. In the car-and-motorcycle prompt, both methods remain plausible, but SGF keeps the car position, road geometry, and motor￾cycle relation more fixed. In the squatting-person prompt, Self Forcing changes the subject pose, identity, and framing over time, whereas SGF maintains a consistent squatting action and park back￾ground. This example illustrates … view at source ↗
Figure 8
Figure 8. Figure 8: Frame-wise 60-second comparison under causal CD initialization. Under causal CD initialization, Self Forcing does not always collapse, but it exhibits systematic crop and scale drift. The sea-turtle and elephant examples progressively zoom into the subject, reducing scene context and sometimes clipping the object. SGF keeps the animal scale, background, and camera distance more stable, showing that SGF imp… view at source ↗
Figure 9
Figure 9. Figure 9: Frame-wise 240-second comparison under causal CD initialization. The otter prompt shows that Self Forcing can preserve broad semantics but accumulates pose and boundary distor￾tions, while SGF keeps the otter centered on the surfboard with a more stable tropical-water layout. In the anime reading prompt, Self Forcing drifts into partial crops and high-saturation color streaks by the late timestamps; SGF pr… view at source ↗
Figure 10
Figure 10. Figure 10: Frame-wise 60-second comparison under causal ODE initialization. In the library￾reading prompt, Self Forcing introduces chromatic background artifacts by 24 seconds, drops the seated reader by 36 seconds, and devolves into bookshelf/window fragments by 48–60 seconds; SGF keeps the reader, armchair, open book, and bookcase layout stable. In the clock-and-vase prompt, Self Forcing drifts from the intended f… view at source ↗
Figure 11
Figure 11. Figure 11: Frame-wise 240-second comparison under causal ODE initialization. Under matched initialization and frame-wise streaming settings, Self Forcing undergoes severe long-horizon drift: the candle/creature example degenerates into color blocks and partial subject replacements, and the disco kangaroo example repeatedly loses the stage layout and subject identity. SGF preserves the prompt-specific subject, pose f… view at source ↗
Figure 12
Figure 12. Figure 12: Chunk-wise 60-second comparison under TF initialization. In the apple-peeling prompt, Self Forcing with TF initialization changes the person identity and camera viewpoint, at times replacing the tabletop action with close-ups of a smiling face. SGF keeps the hands, apples, and countertop action more consistent. In the eyebrow-makeup prompt, Self Forcing changes the number and identity of people across tim… view at source ↗
Figure 13
Figure 13. Figure 13: Chunk-wise 240-second comparison under TF initialization. The coarser chunk￾wise rollout tests whether a generated block remains useful as context several chunks later. In the bacterium prompt, SGF maintains a population of translucent spherical cells more consistently, whereas Self Forcing often turns the scene into unrelated circular textures or sparse particles. In the lizard prompt, Self Forcing frequ… view at source ↗
Figure 14
Figure 14. Figure 14: Chunk-wise 60-second comparison under causal CD initialization. The strip includes bidirectional-ODE and causal-ODE Self Forcing references in addition to the matched causal-CD pair. In the umbrella-couple prompt, SGF under causal CD initialization keeps two people, umbrella colors, and rainy street layout more coherent, while Self Forcing variants often change viewpoint or subject arrangement. In the cor… view at source ↗
Figure 15
Figure 15. Figure 15: Chunk-wise 240-second comparison under causal CD initialization. This strip in￾cludes ODE-initialized Self Forcing references and a matched causal-CD pair. In the running-man prompt, the ODE baselines drift into night streets or back-view shots, while matched Self Forcing under causal CD initialization preserves the subject intermittently but still changes viewpoint and fa￾cial identity. SGF under causal … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 16 linked inside Pith

  1. [1]

    Lumiere: A space-time diffusion model for video generation.arXiv preprint arXiv:2401.12945,

    Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, Yuanzhen Li, Michael Rubinstein, Tomer Michaeli, Oliver Wang, Deqing Sun, Tali Dekel, and Inbar Mosseri. Lumiere: A space-time diffusion model for video generation.arXiv preprint arXiv:2401.12945,

    Pith/arXiv arXiv

  2. [3]

    Metrics are rows and model variants are columns, and boldface marks the better value within each matched Self Forcing–SGF pair

    B FULLMETRICTABLES This appendix reports the complete 5-second VBench results. Metrics are rows and model variants are columns, and boldface marks the better value within each matched Self Forcing–SGF pair. These 13 Table 5: Frame-wise 5-second VBench metrics. Evaluation uses sink 4, FIFO 16, and current chunk 1, for a total context window of 21 latent fr...

    2060

  3. [6]

    Rolling forcing: Autoregressive long video diffusion in real time.arXiv preprint arXiv:2509.25161,

    Kunhao Liu, Wenbo Hu, Jiale Xu, Ying Shan, and Shijian Lu. Rolling forcing: Autoregressive long video diffusion in real time.arXiv preprint arXiv:2509.25161,

    Pith/arXiv arXiv

  4. [7]

    Shotstream: Streaming multi-shot video generation for interactive storytelling

    Yawen Luo, Xiaoyu Shi, Junhao Zhuang, Yutian Chen, Quande Liu, Xintao Wang, Pengfei Wan, and Tianfan Xue. Shotstream: Streaming multi-shot video generation for interactive storytelling. arXiv preprint arXiv:2603.25746,

    arXiv

  5. [8]

    Tethercache: Stabilizing autoregressive long-form video generation with gated recall and trusted alignment.arXiv preprint arXiv:2606.13035,

    Yu Meng, Xiangyang Luo, Letian Li, Wenyuan Jiang, Chen Gao, Xinlei Chen, Yong Li, and Xiao- Ping Zhang. Tethercache: Stabilizing autoregressive long-form video generation with gated recall and trusted alignment.arXiv preprint arXiv:2606.13035,

    Pith/arXiv arXiv

  6. [12]

    Head forcing: Long autoregressive video generation via head heterogeneity.arXiv preprint arXiv:2605.14487,

    Jiahao Tian, Yiwei Wang, Gang Yu, and Chi Zhang. Head forcing: Long autoregressive video generation via head heterogeneity.arXiv preprint arXiv:2605.14487,

    Pith/arXiv arXiv

  7. [13]

    Gen-l-video: Multi-text to long video generation via temporal co-denoising.arXiv preprint arXiv:2305.18264,

    Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-l-video: Multi-text to long video generation via temporal co-denoising.arXiv preprint arXiv:2305.18264,

    Pith/arXiv arXiv

  8. [14]

    Matrix-game 3.0: Real-time and streaming interactive world model with long- horizon memory.arXiv preprint arXiv:2604.08995,

    Zile Wang, Zexiang Liu, Jiaxing Li, Kaichen Huang, Baixin Xu, Fei Kang, Mengyin An, Peiyu Wang, Biao Jiang, Yichen Wei, Yidan Xietian, Jiangbo Pei, Liang Hu, Boyi Jiang, Hua Xue, Zidong Wang, Haofeng Sun, Wei Li, Wanli Ouyang, Xianglong He, Yang Liu, Yangguang Li, and Yahui Zhou. Matrix-game 3.0: Real-time and streaming interactive world model with long- ...

    Pith/arXiv arXiv

  9. [15]

    Sparse forcing: Native trainable sparse attention for real-time autoregressive diffusion video generation.arXiv preprint arXiv:2604.21221, 2026a

    Boxun Xu, Yuming Du, Zichang Liu, Siyu Yang, Ziyang Jiang, Siqi Yan, Rajasi Saha, Albert Pumarola, Wenchen Wang, and Peng Li. Sparse forcing: Native trainable sparse attention for real-time autoregressive diffusion video generation.arXiv preprint arXiv:2604.21221, 2026a. Gangwei Xu, Qihang Zhang, Jiaming Zhou, Xing Zhu, Yujun Shen, Xin Yang, and Yinghao X...

    Pith/arXiv arXiv

  10. [16]

    Infinity- RoPE: Action-controllable infinite video generation emerges from autoregressive self-rollout

    Hidir Yesiltepe, Tuna Han Salih Meral, Adil Kaan Akan, Kaan Oktay, and Pinar Yanardag. Infinity- RoPE: Action-controllable infinite video generation emerges from autoregressive self-rollout. arXiv preprint arXiv:2511.20649,

    arXiv

  11. [17]

    Video-mirai: Autore- gressive video diffusion models need foresight.arXiv preprint arXiv:2606.03971,

    Yonghao Yu, Lang Huang, Runyi Li, Zerun Wang, and Toshihiko Yamasaki. Video-mirai: Autore- gressive video diffusion models need foresight.arXiv preprint arXiv:2606.03971,

    Pith/arXiv arXiv

  12. [18]

    Causal forcing++: Scalable few-step autoregressive diffusion distil- lation for real-time interactive video generation.arXiv preprint arXiv:2605.15141,

    Min Zhao, Hongzhou Zhu, Kaiwen Zheng, Zihan Zhou, Bokai Yan, Xinyuan Li, Xiao Yang, Chongxuan Li, and Jun Zhu. Causal forcing++: Scalable few-step autoregressive diffusion distil- lation for real-time interactive video generation.arXiv preprint arXiv:2605.15141,

    Pith/arXiv arXiv

  13. [19]

    Causal forcing: Autoregressive diffusion distillation done right for high-quality real-time interactive video gener- ation.arXiv preprint arXiv:2602.02214,

    Hongzhou Zhu, Min Zhao, Guande He, Hang Su, Chongxuan Li, and Jun Zhu. Causal forcing: Autoregressive diffusion distillation done right for high-quality real-time interactive video gener- ation.arXiv preprint arXiv:2602.02214,

    Pith/arXiv arXiv

  14. [20]

    The 5-second setting follows the standard VBench protocol and reports all 16 quality and semantic dimensions

    A ADDITIONALEXPERIMENTALDETAILS Benchmark protocol.We evaluate matched Self Forcing and SGF pairs at 5s, 60s, and 240s. The 5-second setting follows the standard VBench protocol and reports all 16 quality and semantic dimensions. The 60-second setting uses the VBench-Long protocol, while the 240-second setting uses 128 randomly sampled MovieGen prompts wi...

    2026

  15. [22]

    The controlled SGF comparisons are the matched causal-CD and TF pairs; ODE-initialized Self Forcing rows are reference baselines. Metric Bidirectional ODE init Causal ODE init Causal CD init TF init SF SF SF SGF SF SGF Aesthetics 0.659 0.660 0.6610.663 0.6630.678 Background 0.961 0.959 0.9450.958 0.9480.964 Dynamics 0.647 0.836 0.7530.875 0.9080.692 Imagi...

    2000

  16. [2020]

    Kingma, Ben Poole, Mohammad Norouzi, David J

    Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models.arXiv preprint arXiv:2210.02303, 2022a. Jonathan Ho, William Chan, Chitwan Saharia, Jay Whang, Ruiqi Gao, Alexey Gritsenk...

    Pith/arXiv arXiv

  17. [2021]

    Roformer: En- hanced transformer with rotary position embedding.arXiv preprint arXiv:2104.09864,

    Jianlin Su, Yu Lu, Shengfeng Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. Roformer: En- hanced transformer with rotary position embedding.arXiv preprint arXiv:2104.09864,

    Pith/arXiv arXiv

  18. [2022]

    Solaris: Building a multiplayer video world model in minecraft.arXiv preprint arXiv:2602.22208,

    Georgy Savva, Oscar Michel, Daohan Lu, Suppakit Waiwitlikhit, Timothy Meehan, Dhairya Mishra, Srivats Poddar, Jack Lu, and Saining Xie. Solaris: Building a multiplayer video world model in minecraft.arXiv preprint arXiv:2602.22208,

    arXiv

  19. [2023]

    Movie gen: A cast of media founda- tion models.arXiv preprint arXiv:2410.13720,

    Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media founda- tion models.arXiv preprint arXiv:2410.13720,

    Pith/arXiv arXiv

  20. [2024]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023a

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, Varun Jampani, and Robin Rom- bach. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023a. Andreas Blattmann, Robin Rombach, Huan Ling, Tim D...

    Pith/arXiv arXiv

  21. [2025]

    Forcing-kv: Hybrid kv cache compression for efficient autoregressive video diffusion models.arXiv preprint arXiv:2605.09681,

    Yicheng Ji, Zhizhou Zhong, Jun Zhang, Qin Yang, Xitai Jin, Ying Qin, Wenhan Luo, Shuiyang Mao, Wei Liu, and Huan Li. Forcing-kv: Hybrid kv cache compression for efficient autoregressive video diffusion models.arXiv preprint arXiv:2605.09681,

    Pith/arXiv arXiv

  22. [2026]

    Self forcing: Bridging the train-test gap in autoregressive video diffusion.arXiv preprint arXiv:2506.08009,

    Xun Huang, Zhengqi Li, Guande He, Mingyuan Zhou, and Eli Shechtman. Self forcing: Bridging the train-test gap in autoregressive video diffusion.arXiv preprint arXiv:2506.08009,

    Pith/arXiv arXiv