Pith. sign in

REVIEW 4 major objections 7 minor 29 references

Vorch-Director: Interactive World Story Model via Noise-Aware Error Rectification

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that matching injected prediction residuals to the flow-matching noise level of each training step substantially mitigates exposure bias in autoregressive audio-visual diffusion generation, keeping identity and audio…

desk verdict Genuinely novel sigma-aware residual injection, but the paper never isolates it—the missing ablation makes the headline claim unsupported as currently written. read the letter →

arxiv 2608.05776 v1 pith:OJ3PKZCI submitted 2026-08-06 cs.CV

Lisai Zhang , Yidi Wu , Qi Liu , Xin Ma , Yang Ding , Gang Yue , Siqian Yang , Jingyuan Chen
show 2 more authors
Lin Ma Yaohui Wang
This is my paper · ORCID
classification cs.CV
keywords autoregressivevideogenerationaudio-visualexposurebiasresidualinjectionflowmatchinglong-horizonidentitydrifttaskembeddings
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Vorch-Director targets the train–inference mismatch that makes autoregressive video generation drift: models are trained on clean ground-truth histories but at inference condition on their own error-laden outputs. The paper's central claim is that the effectiveness of residual-based correction depends on the flow-matching noise level at which the residuals were produced, and that injecting residuals from matched noise regimes during teacher-forcing training produces histories that look like real inference-time histories. On top of an audio-visual diffusion transformer, the method adds task embeddings to distinguish history video, reference images, and target video, plus a short clean conditioning sink, so a single model supports multi-shot, multi-subject, reference-guided audio-visual long-video generation. The paper introduces a long-horizon audio-visual benchmark with anchor-based drift metrics and reports improved cross-shot consistency over the baselines it compares against, while leaving the ablation that isolates noise-level matching to future work.

What carries the argument

The carrier of the argument is the $\sigma$-aware error buffer: a ring buffer of target-token residuals $\delta=\hat{x}_0-x_0$ where each entry stores the flow-matching noise level $\sigma$ at which it was produced, together with the injection rule that samples only residuals satisfying $|\sigma'-\sigma|\le\tau_\sigma$ and applies them as $\tilde{c}_v=c_v+\gamma\delta_s$ with $\gamma\sim U(0.9,1.2)$. This turns teacher-forcing training into exposure to inference-like corruption at a controlled intensity. Secondary machinery is the task-embedding scheme (role tags for history, sink, reference images, reference audio, and target) and the uncorrupted clean sink, which together let the model keep a stable anchor while learning to correct corrupted history.

What would settle it

Run the $\sigma$-agnostic residual-injection variant described in Appendix E against the $\sigma$-aware configuration on the same 22-shot audio-visual benchmark with the same evaluator; if Anchor-relative Consistency and quality-drift scores are statistically indistinguishable, then noise-level matching is not the active ingredient. A second check: collect residuals from a real autoregressive rollout and compare their distribution with the buffer-injected residuals at matched $\sigma$; divergence would show the proxy assumption fails.

Watch

Extended reading notes

Core claim

The paper's discovery is that residual-reuse methods work only when the injected error matches the noise regime of the current denoising step. In flow matching, a target token at noise level $\sigma$ gives the one-step reconstruction $\hat{x}_0 = x_\sigma - \sigma v_\theta(x_\sigma,c)$ and the residual $\delta=\hat{x}_0-x_0$ on each target token; Vorch-Director stores these residuals in a ring buffer tagged with their $\sigma$, and when corrupting the history for a step at level $\sigma$ it samples only residuals with $|\sigma'-\sigma|\le\tau_\sigma$ and injects them as $\tilde{c}_v=c_v+\gamma\delta_s$ with $\gamma\sim U(0.9,1.2)$. Because high-$\sigma$ residuals are coarse and low-$\sigma$ residuals are fine, level-matched injection produces histories that are wrong in the same way inference histories are wrong. Around this, the paper adds learned task embeddings that decouple a token's role from its position, so positional indices can reset each segment, and a short clean sink that is never corrupted, teaching the model to repair a corrupted history rather than propagate it. At inference, no residual injection is applied.

Load-bearing premise

The load-bearing premise is that a one-step reconstruction error measured on a clean target token, once injected into the conditioning history at matched strength, behaves like the accumulated errors the model will actually feed back to itself during long inference. The paper does not yet isolate this premise, since the noise-level-agnostic comparison in its appendix is described but not run.

Editorial extensions

If this is right

  • Standard teacher forcing can be kept, so training avoids the multi-step rollout cost of self-generating histories.
  • Long rollouts should show markedly less identity drift and audio-visual desynchronization than noise-level-agnostic residual reuse.
  • Because task embeddings decouple roles from positions, positional indices can be reset at each segment, removing positional extrapolation from the list of drift sources.
  • One model, trained on a four-way task mixture, can serve pure continuation, subject-reference, image-to-video, and audio-visual continuation without new modules.
  • The anchor-relative and re-appearance metrics should reveal slow identity drift that average or adjacent-shot similarity hides.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the residual proxy holds, the same $\sigma$-matched injection should transfer to other rectified-flow generators and to the audio-history stream, which the paper names as future work.
  • Because the clean sink is drawn from the running context at inference (Appendix D), it may itself carry drift; fixing the sink to a clean reference could strengthen the anchor, and ARC under that variant is a testable extension.
  • The comparison with visual-only baselines is confounded by their lack of audio and lower motion; adding an audio-visual baseline with comparable motion would isolate whether the consistency gain comes from the residual correction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes Vorch-Director, a training-time strategy for autoregressive audio-visual generation that injects prediction residuals into the conditioning history in a noise-level-aware way. The core idea is that residuals collected from one-step reconstructions of noised target tokens should be stored with their originating flow-matching noise level and re-injected into the history at matched noise levels during training, thereby making teacher-forced training histories resemble inference-time self-generated contexts. The method is built on the 22B LTX-2 audio-visual DiT and further includes task embeddings, a clean conditioning sink, a structured prompt template, and a four-way task mixture to support multi-shot, multi-subject, reference-guided long-video generation. Evaluation is conducted on ST-Bench, UnityShots, and a new 16-case long audio-visual benchmark with metrics such as ViCLIP, Self-CIDS, ARC, and Reappear, showing large gains on many cross-shot consistency metrics but mixed results on video quality and audio fidelity.

Significance. If the central claims were fully validated, the paper would make a useful contribution: a training-only method for mitigating exposure bias in autoregressive diffusion models, a unified conditioning interface with task embeddings, and a new long-horizon audio-visual benchmark with anchor-based consistency metrics. The authors also deserve credit for re-scoring all baselines under a single evaluation protocol, which improves comparability across methods. However, the headline mechanism (noise-level-aware residual injection) is not isolated in any experiment, and the paper explicitly defers the decisive ablations to future work. The significance of the empirical gains therefore cannot currently be attributed to the proposed mechanism, and the paper's own results show that audio-visual fidelity is not uniformly improved.

major comments (4)
  1. [Section 3.3, Appendix E, Section 5] The central claim that noise-level matching is responsible for the reported gains is never tested. Appendix E defines the decisive control, the σ-agnostic residual-injection variant, but no result for that variant is reported in any table, and Section 5 explicitly defers the ablations ('noise strategy, clean sink, and reference conditioning') to future work. Since the final system differs from the baselines in several dimensions simultaneously — task embeddings, clean sink, four-way task mixture, residual injection, and prompt template — the large improvements in Tables 2–4 cannot be attributed to the σ-aware injection. Please run and report the σ-agnostic and Gaussian-context controls on at least one benchmark; without them, the paper's title claim is unsupported by the evidence presented.
  2. [Section 3.4 (Eq. 5) and Section D] The clean conditioning sink has the same train–test mismatch it is meant to remove. Eq. (5) defines the sink at training as the true history's leading window, r = VAE(h[0:Δ]), which is clean ground truth. Section D states that at inference the sink is 'taken from the start of the running context,' i.e., from the model's own generated history, which is corrupted. Thus the inference-time sink is not clean, the τ_sink tokens follow a different distribution from training, and the purported stable anchor is not actually available at test time. Please specify how the sink is obtained at inference; if it is the generated context, either train with a corrupted sink as well or use a fixed clean reference clip that is available at both training and inference.
  3. [Section 4.2 and Table 4] The metric 'Quality Drift' (QD) is promised in the Abstract and Introduction as one of the two long-horizon measures on the new benchmark, but it is never defined and no QD values appear in Table 4 or anywhere else. Appendix F refers to 'the two long-horizon measures QD and ARC of Section 4.5,' yet Section 4.5 does not define QD or report it. Please either provide the definition and the measured values or remove the promise; as written, the paper announces a central metric that does not exist in the empirical section.
  4. [Tables 2–4] The empirical comparisons report no confidence intervals, variance, or significance tests. This is especially concerning for the new long audio-visual benchmark, which contains only 16 cases; the claimed large margins over baselines (e.g., ViCLIP 0.6269 vs. 0.4922 in Table 4) could be within run-to-run variability. Please report at least standard deviations (and ideally significance tests) computed over seeds or bootstrap resamples for the headline metrics, or explicitly justify why the differences are robust despite the small sample size.
minor comments (7)
  1. [Abstract] The Abstract contains missing spaces ('oncleanground-truth', 'onits own'); please proofread the manuscript.
  2. [Section 4.2] The text first names 'Quality Drift and Anchor-relative Consistency' as the two long-horizon metrics, but then defines ARC and Reappear and does not define QD; reconcile the metric naming and definitions.
  3. [Table 2] The header 'T ext Consist.' should read 'Text Consist.' and the column alignment should be checked.
  4. [Section 3.1, Eq. (3)] The token sequence in Eq. (3) contains non-ASCII artifacts ('bracehtipupleft', 'bracehtipdownright', etc.) that should be cleaned up.
  5. [Related Work and Conclusion] The method is referred to as 'Matrix-Game 3.0' in Section 2.2 and 'Matrix-Game' in the Conclusion; use a single consistent citation name.
  6. [Section 4.1] Hyperparameters such as γ∼U(0.9,1.2) and τ_σ=0.05 are given without any sensitivity analysis; a small grid over these values would help the reader assess how robust the method is to their choice.
  7. [Abstract and Table 4] The Abstract claims 'improved stability and audio-visual fidelity over strong baselines,' but in Table 4 Vorch-Director has lower Voice (0.6833 vs. 0.7244) than JoyAI-Echo and equal Acc (0.0680); the claim should be softened to match the reported numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the sigma-aware residual-injection claim is not derived from its own inputs; the noted weakness is an unrun control, not a circular reduction.

full rationale

The paper's derivation chain is self-contained rather than circular. The residual delta = x_hat_0 - x_0 is defined by the one-step reconstruction in Eq. (2), injected into history tokens by Eq. (4), and stabilized by the clean sink in Eq. (5); none of these equations defines the evaluation metrics or the final video-quality scores as inputs. The central empirical results are grounded against external benchmarks, ST-Bench and UnityShots, under a unified evaluator, so the system-level claims are not manufactured from the paper's own fitted values. The author-defined long-horizon metrics ARC and Reappear are applied uniformly to all methods by the same evaluator rather than used to construct the model, so they do not make the comparison circular. The cited prior work, including Matrix-Game and LTX-2, is used as motivation and backbone, not as a self-citation chain that forces the conclusion; the few references overlapping with the authors' own prior work (e.g., LaVie, Latte) are not load-bearing. The manuscript's genuine weakness is the missing ablation: Appendix E defines the sigma-agnostic residual-injection control, but no result for it is reported, and Section 5 states 'Completing these evaluations and the corresponding ablations (noise strategy, clean sink, and reference conditioning) is our immediate next step.' That is a failure to isolate the central mechanism and an attribution risk, but it is not a circularity: the sigma-aware claim is not equivalent by construction to its inputs. The paper's derivation does not reduce to its predictions, so the circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper's central claim depends on hand-set constants (γ, τσ, Δ, λa) and on two unvalidated modeling assumptions: that one-step residuals are faithful proxies for inference-time corruption, and that the sink remains clean at inference. None of these are derived, and the ablations that would justify them are explicitly deferred in the conclusion.

free parameters (5)
  • Injection strength γ (per-step range) = U(0.9, 1.2)
    Hand-chosen in Section 3.3/4.1 with no sensitivity analysis; directly sets the size of corruption added to the conditioning history.
  • Sigma-match tolerance τσ = 0.05
    Hand-picked in Section 3.3 for accepting residuals into the pool; no study of the tolerance or the nearest-σ fallback.
  • Clean sink length Δ = 1 second
    Hand-picked in Section 3.4 and used at inference in Section D; no ablation of sink length.
  • Audio loss weight λa = unreported
    Appears in Eq. 7 but its value is never given, leaving the video-audio loss balance a hidden tuning parameter.
  • Error buffer size and sampling constants = 1,048,576 tokens; warmup 100 steps; injection probability 1.0; top-norm fraction 25%
    Implementation constants in Section 4.1/Appendix B, treated as defaults without ablation.
assumptions (5)
  • standard math Rectified-flow interpolation xσ = (1-σ)x0 + σϵ with velocity target u = ϵ - x0 (Eq. 1).
    Standard flow-matching formulation adopted from prior work; not disputed.
  • domain assumption One-step reconstruction residual δ = x̂0 - x0 is a faithful proxy for inference-time history corruption.
    Section 3.3 collects residuals from the one-step estimate and treats them as exactly the error a fed-back token would carry; this is the core premise of the method and is never validated empirically.
  • domain assumption Residuals with noise levels within τσ=0.05 are interchangeable error regimes.
    The σ-matched sampling pools residuals within a fixed tolerance (Eq. 4); there is no evidence that this tolerance is the right granularity for matching error structure.
  • domain assumption A clean conditioning sink is available at inference.
    Training uses a clean ground-truth sink (Eq. 5), but Section D re-encodes the sink from the running context of the model's own output, so it is not clean; the paper treats it as clean without discussion.
  • ad hoc to paper Four-way task mixture with 25% weights covers the target distribution.
    The mixture in Section 3.2 is a design choice with no analysis or ablation of the weight balance.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Vorch-Director: Interactive World Story Model via Noise-Aware Error Rectification." pith.science (2026). https://pith.science/paper/OJ3PKZCI

@misc{pith2026260805776,
  author       = {Pith},
  title        = {Pith review of: Vorch-Director: Interactive World Story Model via Noise-Aware Error Rectification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OJ3PKZCI}},
  note         = {Machine review of arXiv:2608.05776}
}
read the original abstract

Autoregressive continuation provides a natural path toward minute-scale audio-visual generation by repeatedly extending a short-window generator conditioned on previously generated video and audio. However, models are trained on clean ground-truth histories, while inference relies on their own generated histories, where accumulated errors cause identity drift, over-smoothing, and audio-visual desynchronization. Recent methods reduce this mismatch by reusing prediction residuals as synthetic corruption, but we observe that the effectiveness of residual correction critically depends on the flow-matching noise level at which residuals are produced. We propose Vorch-Director, a noise-level-aware residual correction strategy that associates each residual with its originating noise level and injects residuals from matched noise regimes during training. By aligning injected errors with the denoising process, Vorch-Director produces more realistic autoregressive histories while retaining efficient teacher-forcing training. Built on the audio-visual LTX-2 diffusion transformer, Vorch-Director further introduces task embeddings to distinguish historical video, reference images, and target video, enabling unified conditioning for long-horizon generation. Together with a clean conditioning sink and mixed-task training, Vorch-Director supports multi-shot, multi-subject, reference-guided audio-visual long-video generation. We evaluate Vorch-Director on ST-Bench and introduce a new long-horizon audio-visual benchmark with metrics for quality drift and long-range consistency. Extensive experiments demonstrate improved stability and audio-visual fidelity over strong baselines.

Figures

Figures reproduced from arXiv: 2608.05776 by the authors.

Figure 1
Figure 1. Vorch-Director generates minute-scale, multi-shot, audio-visual video by autoregressive con￾tinuation. Given a few reference images that fix subject identity, our model extends an audio-visual video over dozens of shots with camera cuts, keeping character appearance, scene, and speaker voice consistent while emitting synchronized audio. audio, whereas at inference it consists entirely of the model’s own predictions,… view at source ↗
Figure 2
Figure 2. Overview of Vorch-Director. During training the model predicts velocities on noised target tokens; the resulting reconstruction residuals are collected into a buffer tagged by their noise level σ. When forming a later step, residuals of a matching σ are sampled and injected into the conditioning-history tokens with strength γ, so the history carries inference-like error at the right intensity for the current denoisi… view at source ↗
Figure 3
Figure 3. Data-construction pipeline. From same-source long videos we filter shots and build a per￾character subject index; a multimodal captioner (Doubao) produces separate video, audio, and joint audio￾visual captions; each source video is split into a history (Video-1) and target (Video-2) pair. The result is a multi-task mixture (continuation, continuation+image, subject-IP, and audio-visual continuation) rendered into a … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison over a three-minute rollout on the long audio-visual benchmark. One frame is sampled a quarter of the way into every third shot, so the eight columns span shots 1, 4, 7, . . . , 22, i.e. 0s to 177s of a 22-shot continuation. Memento and StoryMem …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 8 canonical work pages

  1. [1]

    Towards Error-Free Long Video Generation

    Shuning Chang, Weihua Chen, Jiasheng Tang, Hao Xu, Zeyu Zhang, Hangjie Yuan, Yu Lu, and Ruigang Niu. Towards error-free long video generation.arXiv preprint arXiv:2606.22370,

    work page Pith review arXiv

  2. [4]

    Avatarforcing: One- step streaming talking avatars via local-future sliding-window denoising.arXiv preprint arXiv:2603.14331,

    Liyuan Cui, Wentao Hu, Wenyuan Zhang, Zesong Yang, Fan Shi, and Xiaoqiang Liu. Avatarforcing: One- step streaming talking avatars via local-future sliding-window denoising.arXiv preprint arXiv:2603.14331,

    arXiv

  3. [5]

    Dreamid-omni: Unified framework for controllable human-centric audio-video generation.arXiv preprint arXiv:2602.12160,

    Xu Guo, Fulong Ye, Qichao Sun, Liyang Chen, Bingchuan Li, Pengze Zhang, Jiawei Liu, Songtao Zhao, Qian He, and Xiangwang Hou. Dreamid-omni: Unified framework for controllable human-centric audio-video generation.arXiv preprint arXiv:2602.12160,

    arXiv

  4. [6]

    Unityshots: Memory-driven multi-shot audio-video generation with boundary-aware gating.arXiv preprint arXiv:2606.21661,

    Jiehui Huang, Yuechen Zhang, Bin Xia, Jiahao Wang, Xu He, Zhenchao Tang, Meng Chu, and Xin Tao. Unityshots: Memory-driven multi-shot audio-video generation with boundary-aware gating.arXiv preprint arXiv:2606.21661,

    arXiv

  5. [7]

    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,

    arXiv

  6. [9]

    GitHub: jd- opensource/JoyAI-Echo

    Technical report. GitHub: jd- opensource/JoyAI-Echo. Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, and Jianwei Zhang. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603,

    arXiv

  7. [10]

    Stable video infinity: Infinite- length video generation with error recycling.arXiv preprint arXiv:2510.09212,

    12 Vorch-Director: Interactive World Story Model via Noise-Aware Error Rectification Wuyang Li, Wentao Pan, Po-Chien Luan, Yang Gao, and Alexandre Alahi. Stable video infinity: Infinite- length video generation with error recycling.arXiv preprint arXiv:2510.09212,

    arXiv

  8. [11]

    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,

    arXiv

Show all 29 references
  1. [12]

    Filmweaver: Weaving consistent multi-shot videos with cache-guided autoregressive diffusion.arXiv preprint arXiv:2512.11274,

    Xiangyang Luo, Qingyu Li, Xiaokun Liu, Wenyu Qin, Miao Yang, Meng Wang, Pengfei Wan, Di Zhang, Kun Gai, and Shao-Lun Huang. Filmweaver: Weaving consistent multi-shot videos with cache-guided autoregressive diffusion.arXiv preprint arXiv:2512.11274,

  2. [13]

    Corruption-aware training of latent video diffusion models for robust text-to-video generation.arXiv preprint arXiv:2505.21545,

    Chika Maduabuchi, Hao Chen, Yujin Han, and Jindong Wang. Corruption-aware training of latent video diffusion models for robust text-to-video generation.arXiv preprint arXiv:2505.21545,

  3. [14]

    Holocine: Holistic generation of cinematic multi-shot long video narratives.arXiv preprint arXiv:2510.20822,

    Yihao Meng, Hao Ouyang, Yue Yu, Qiuyu Wang, Wen Wang, Ka Leong Cheng, Hanlin Wang, Yixuan Li, Cheng Chen, and Yanhong Zeng. Holocine: Holistic generation of cinematic multi-shot long video narratives.arXiv preprint arXiv:2510.20822,

  4. [15]

    Tethercache: Stabilizingautoregressivelong-formvideogenerationwithgatedrecallandtrustedalignment

    Yu Meng, Xiangyang Luo, Letian Li, Wenyuan Li, Chen Gao, Xinlei Chen, Yong Li, and Xiao-Ping Zhang. Tethercache: Stabilizingautoregressivelong-formvideogenerationwithgatedrecallandtrustedalignment. arXiv preprint arXiv:2606.13035,

    arXiv

  5. [16]

    Flowlong: Inference-time long video generation via manifold-constrained tweedie matching.arXiv preprint arXiv:2605.20910,

    Jangho Park, Geon Yeong Park, Gihyun Kwon, and Jong Chul Ye. Flowlong: Inference-time long video generation via manifold-constrained tweedie matching.arXiv preprint arXiv:2605.20910,

    arXiv

  6. [18]

    Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314,

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, and Jianxiao Yang. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314,

    arXiv

  7. [19]

    Error analyses of auto-regressive video diffusion models: A unified framework.arXiv preprint arXiv:2503.10704,

    Jing Wang, Fengzhuo Zhang, Xiaoli Li, Vincent YF Tan, Tianyu Pang, Chao Du, Aixin Sun, and Zhuoran Yang. Error analyses of auto-regressive video diffusion models: A unified framework.arXiv preprint arXiv:2503.10704,

  8. [20]

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

    arXiv

  9. [21]

    Memento: Reconstructtorememberforconsistentlongvideogeneration.arXiv preprint arXiv:2606.14667,

    XuanWei, LongbinJi, GuanWang, XiangruiLiu, ZhenyuZhang, ShuohuanWang, YuSun, andQingqiHong. Memento: Reconstructtorememberforconsistentlongvideogeneration.arXiv preprint arXiv:2606.14667,

  10. [22]

    Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072,

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, and Guanyu Feng. Cogvideox: Text-to-video diffusion models with an expert transformer.arXiv preprint arXiv:2408.06072,

    arXiv

  11. [23]

    Talker-t2av: Joint talking audio-video generation with autoregressive diffusion modeling.arXiv preprint arXiv:2604.23586,

    13 Vorch-Director: Interactive World Story Model via Noise-Aware Error Rectification Zhen Ye, Xu Tan, Aoxiong Yin, Hongzhan Lin, Guangyan Zhang, Peiwen Sun, Yiming Li, Chi-Min Chan, Wei Ye, Shikun Zhang, and Wei Xue. Talker-t2av: Joint talking audio-video generation with autor...

    arXiv

  12. [24]

    Storymem: Multi-shot long video storytelling with memory.arXiv preprint arXiv:2512.19539,

    Kaiwen Zhang, Liming Jiang, Angtian Wang, Jacob Zhiyuan Fang, Tiancheng Zhi, Qing Yan, Hao Kang, Xin Lu, and Xingang Pan. Storymem: Multi-shot long video storytelling with memory.arXiv preprint arXiv:2512.19539,

  13. [25]

    Kaiwen Zheng, Guande He, Min Zhao, Jintao Zhang, Huayu Chen, Jianfei Chen, Chen-Hsuan Lin, and Ming- Yu Liu. Causal-rcm: A unified teacher-forcing and self-forcing open recipe for autoregressive diffusion distillation in streaming video generation and interactive world models....

    arXiv

  14. [26]

    Videomemory: Toward consistent video generation via memory integration.arXiv preprint arXiv:2601.03655,

    Jinsong Zhou, Yihua Du, Xinli Xu, Luozhou Wang, Zijie Zhuang, Yehang Zhang, Shuaibo Li, Xiaojun Hu, Bolan Su, and Ying-cong Chen. Videomemory: Toward consistent video generation via memory integration.arXiv preprint arXiv:2601.03655,

  15. [27]

    Storydiffusion: Consistent self- attention for long-range image and video generation.Advances in Neural Information Processing Systems, arXiv preprint arXiv:2405.01434,

    Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self- attention for long-range image and video generation.Advances in Neural Information Processing Systems, arXiv preprint arXiv:2405.01434,

    arXiv

  16. [28]

    Hiar: Efficient autoregressive long video generation via hierarchical denoising.arXiv preprint arXiv:2603.08703,

    Kai Zou, Dian Zheng, Hongbo Liu, Tiankai Hang, Bin Liu, and Nenghai Yu. Hiar: Efficient autoregressive long video generation via hierarchical denoising.arXiv preprint arXiv:2603.08703,

  17. [29]

    For segmentt≥2we condition on the references, the clean sinkr(length∆) taken from the start of the running context, the previous segment’s generated video and audio as the history(c(t) v ,c (t) a ), and promptpt, each tagged by its task embedding as in training. Starting from ...

    2026

  18. [2023]

    Seedance 2.0: Advancing video generation for world complexity

    Team Seedance, De Chen, Liyang Chen, Xin Chen, Ying Chen, Zhuo Chen, Zhuowei Chen, Feng Cheng, Tianheng Cheng, and Yufeng Cheng. Seedance 2.0: Advancing video generation for world complexity. arXiv preprint arXiv:2604.14148,

    arXiv

  19. [2024]

    Cinedance: Towards next-generation multi-shot long-form cinematic audio-video generation.arXiv preprint arXiv:2606.09639,

    Yuheng Chen, Teng Hu, Yuji Wang, Qingdong He, Zhucun Xue, Qianyu Zhou, Jason Li, Lizhuang Ma, Jiangning Zhang, and Dacheng Tao. Cinedance: Towards next-generation multi-shot long-form cinematic audio-video generation.arXiv preprint arXiv:2606.09639,

    arXiv

  20. [2025]

    Videoar: Autoregressive video generation via next-frame and scale prediction.arXiv preprint arXiv:2601.05966,

    Longbin Ji, Xiaoxiong Liu, Junyuan Shang, Shuohuan Wang, Yu Sun, Hua Wu, and Haifeng Wang. Videoar: Autoregressive video generation via next-frame and scale prediction.arXiv preprint arXiv:2601.05966,

  21. [2026]

    Diffusion forcing: Next-token prediction meets full-sequence diffusion.arXiv preprint arXiv:2407.01392,

    Boyuan Chen, Diego Marti Monso, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion.arXiv preprint arXiv:2407.01392,

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.