Pith. sign in

REVIEW 4 major objections 6 minor 29 references

Real video clips fix long-rollout drift in streaming video AI

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 · glm-5.2

2026-07-10 01:30 UTC pith:EV2PXJB5

load-bearing objection Private letter on OPSD-V the 4 major comments →

arxiv 2607.08766 v1 pith:EV2PXJB5 submitted 2026-07-09 cs.CV

OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators

Hongyu Liu , Chun Wang , Feng Gao , Xuanhua He , Yue Ma , Ziyu Wan , Yong Zhang , Xiaoming Wei
show 1 more author
Qifeng Chen
This is my paper
classification cs.CV
keywords few-stepopsd-vvideocacheautoregressivedynamicsmodelson-policy
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.

Few-step autoregressive video generators produce long videos efficiently but degrade over time: errors accumulate in the transformer key-value cache, motion weakens, and artifacts spread. The standard fix—distilling from a short-clip teacher—cannot supervise truly long rollouts because the teacher itself was never trained on long sequences. This paper proposes OPSD-V, which post-trains an existing few-step AR video model using real one-minute videos not as reconstruction targets but as privileged context for a teacher branch. During training, the student model rolls out exactly as it would at inference, writing its own generated chunks into its cache. At each supervised denoising step, a teacher model (an exponential-moving-average copy of the student) is evaluated at the same noisy latent the student is currently denoising, but with a cleaner cache: older generated history is swapped out for the corresponding real-video chunks while the most recent student-generated chunk is retained to preserve autoregressive continuity. The student is then pushed toward the teacher's velocity prediction via a simple mean-squared-error loss applied at every denoising step of every supervised chunk. The result is dense, step-level corrective supervision on the exact cache states the model will encounter at deployment, with zero added inference cost—the teacher branch and real-video cache exist only during training.

Core claim

The central mechanism is the asymmetric cache construction between student and teacher. The student conditions on its own self-generated KV cache (replicating inference), while the teacher conditions on a hybrid cache where older generated history is replaced by real-video chunks but the most recent chunk remains student-generated. This creates a teacher target that is cleaner than what the student sees (because older errors are washed out by real data) yet still autoregressively anchored to the student's current state (because the latest chunk is shared). Supervising the student to match the teacher's velocity at the student's own visited noisy latents then corrects the model on the states它

What carries the argument

The teacher cache (Eq. 12): a hybrid KV cache where all but the most recent chunk come from real video rather than student generation. This is the object that makes the supervision both cleaner than the student's own state and semantically compatible with the student's current rollout position.

Load-bearing premise

The teacher's velocity prediction, computed with a hybrid real-video-plus-student cache, is assumed to be a useful correction for the student's current state. If the student's generated trajectory has diverged too far from the real video's semantic content—different objects, different motion—the teacher's correction direction may not align with what the student actually needs, and the paper does not quantify how much divergence is tolerable before the supervision signal degr

What would settle it

Apply OPSD-V post-training to a model whose initial rollout diverges sharply from the real training video's semantic content (e.g., different subject identity or contradictory motion). If the teacher's velocity at the student's noisy latent points in a direction that increases rather than decreases the student's error on subsequent chunks, the method's core assumption—that real-video context produces cleaner targets regardless of trajectory divergence—would be falsified.

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

If this is right

  • Post-training with real long-video context could become a standard refinement step for any few-step AR video model, analogous to how RLHF post-training refines language models without changing their inference path.
  • The asymmetric-cache self-distillation principle may transfer to other autoregressive modalities with evolving context (e.g., long-form audio or text generation) where error accumulation in cached context is a bottleneck.
  • If the method scales with more training data and longer videos as the authors suggest, it could narrow the quality gap between streaming AR video generators and offline bidirectional video models for long-horizon generation.
  • The finding that test-time cache quality alone affects long-rollout stability (Fig. 2, lower) implies that inference-time cache-replacement or cache-cleaning heuristics might yield immediate gains without any training.

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 / 6 minor

Summary. The paper proposes OPSD-V, a post-training method for few-step autoregressive (AR) video diffusion models. The core idea is on-policy self-distillation: the student model follows its exact inference-time rollout (generating chunks from its own KV cache), while a teacher—parameterized as an EMA copy of the student—is evaluated at the same student-visited noisy latents and timesteps but conditioned on a 'cleaner' cache where older generated history is replaced by real-video chunks (Eq. 12). A velocity-matching loss (Eq. 14) provides dense supervision along the denoising trajectory. The method is applied to two base models (Self-Forcing and LongLive) and evaluated on one-minute video generation using VBenchLong and a user study. Results show improvements in Quality Score and Dynamic Degree with no added inference cost.

Significance. The paper addresses a genuine and timely problem: long-horizon degradation in few-step AR video generators whose training supervision is limited by short-clip teachers. The proposed framework is clearly specified, with exact cache construction rules (Eq. 12–13), a well-defined loss (Eq. 14), and a memory-efficient training algorithm (Algorithm 1) that makes long-rollout post-training tractable. The diagnostic in Fig. 2 (test-time cache intervention on the base model) is a clean motivating experiment that isolates cache quality as a bottleneck. The method preserves the original few-step inference path with zero added inference cost, which is a practical strength. Code and project page are promised. The idea of using real long-video data as privileged temporal context rather than direct reconstruction targets is a reasonable and novel instantiation of on-policy self-distillation for the AR video setting.

major comments (4)
  1. §4.3, Eq. (12)–(14): The central supervision signal assumes the teacher velocity v̂^t_{i,k} = f_θ̄(z^s_{i,k}, t_k, c, h^t_i) is a corrective target for the student. However, z^s_{i,k} is produced by the student's trajectory (conditioned on its own generated chunks), while h^t_i replaces older history with real-video chunks that may depict different content, objects, or motion. The paper acknowledges this risk (§4 intro: 'semantic and temporal mismatch') but never measures the divergence between student rollout and real video, nor establishes whether there is a threshold beyond which the teacher target becomes misleading. The consistent Semantic Score decrease on both models (Table 1: LongLive 0.4911→0.4904, Self-Forcing 0.4897→0.4864) could be a symptom of this mismatch. The authors should either (a) measure student-vs-real-video divergence (e.g., feature distance) across the rollout and
  2. §5.1, Table 1: The quantitative evaluation uses only 240 prompts total (120 MovieGenBench + 120 internal 'MeiBench'). For a method claiming improvements in long-horizon generation, this is a relatively small evaluation set, and no confidence intervals or significance tests are reported. The Quality Score improvements (0.8138→0.8242 for LongLive, 0.8259→0.8389 for Self-Forcing) are modest in absolute terms. Without variance estimates, it is difficult to assess whether these gains are statistically meaningful or within noise. The authors should report standard deviations or confidence intervals, or at minimum discuss the statistical reliability of the reported improvements.
  3. §5.4, Fig. 7: The ablation comparing student-trajectory vs. teacher-trajectory supervision is qualitative only (visual comparison of blur). This is a load-bearing design choice—the entire method hinges on evaluating the teacher at student-visited states. A quantitative ablation (VBenchLong scores for both variants) would substantially strengthen the claim that on-policy states are necessary. The current presentation does not allow the reader to verify the magnitude of the effect.
  4. §5.1: The training dataset is small (3,800 videos, ~1 minute each) and the training is only 200 iterations on 24 H800 GPUs. The paper does not discuss whether the method has converged or whether further training would yield additional gains. Given that the improvements are attributed to the supervision signal rather than the data or compute, some analysis of training dynamics (loss curves, quality vs. iterations) would help establish that the gains come from the method rather than from additional exposure to long-video data. A comparison against a simpler baseline—e.g., continued LoRA training on the same long-video data with a standard reconstruction or DMD loss—would isolate the contribution of the on-policy self-distillation mechanism.
minor comments (6)
  1. §4.4, Eq. (15): The warm-up M=7 is justified by the Wan-based AR models' local training window of 7 chunks (21 latent frames). It would help to state the sensitivity of results to this choice, even briefly.
  2. Table 1: The 'Semantic Score' decreases for both models after OPSD-V. The paper notes a 'mild trade-off' (§5.3) but does not discuss whether this is acceptable or whether it indicates the teacher is pulling toward real-video content at the expense of prompt alignment. A brief discussion of this trade-off would be informative.
  3. Fig. 2 (bottom): The diagnostic experiment is described in text but the figure itself is not referenced with specific frame numbers or quantitative metrics. Stating how many frames/chunks were compared and whether the improvement was measured objectively would strengthen the motivation.
  4. §5.3, Fig. 5: The user study has 10 participants and 20 video pairs (10 per backbone). This is a small sample. The paper should acknowledge this limitation more explicitly, and ideally report inter-rater agreement.
  5. References: Several citations use future-dated arXiv IDs (e.g., 2601.xxxxx, 2602.xxxxx, 2603.xxxxx, 2605.xxxxx). These appear to be placeholder or incorrect dates and should be verified.
  6. §3.1: The notation h_i for cache state is introduced but the distinction between h^s_i (student) and h^t_i (teacher) is only made explicit in §4. A forward reference or earlier introduction of the superscript convention would improve readability.

Circularity Check

0 steps flagged

No significant circularity: the teacher's corrective signal derives from external real-video data, not from the student's own outputs or fitted parameters.

full rationale

The central derivation chain of OPSD-V is not circular. The student rollout (Eq. 5-9) produces on-policy states (z^s_{i,k}, h^s_i) using the model's own generated chunks. The teacher target (Eq. 10, 12) is evaluated at the same student-visited noisy latents z^s_{i,k} but conditioned on a cache h^t_i where older history is replaced by real-video chunks x^data. The velocity matching loss (Eq. 14) then minimizes the distance between the student prediction and this teacher target. The teacher's corrective signal originates from external ground-truth data (real long videos), not from the student's own predictions or a parameter fitted to the target variable. The EMA teacher (bar_theta) is a copy of the student, but its predictions are conditioned on privileged real-video context, providing an independent supervision source rather than a self-referential definition. The ablation in Fig. 7 (teacher-trajectory supervision fails) further confirms that the result is not trivially forced by construction: if the method were circular, switching to teacher-trajectory states would not produce the observed catastrophic failure. The only minor self-referential element is that the teacher parameters bar_theta are an EMA copy of the student theta, but this is a standard self-distillation technique and does not make the target equivalent to the input by definition, since the conditioning context differs. The Semantic Score degradation noted in Table 1 is a correctness concern (potential semantic mismatch between student trajectory and real-video context), not a circularity issue. The derivation is self-contained against external benchmarks (VBenchLong, user study) and the central claim has independent empirical content. Score: 2 (one minor self-referential element via EMA teacher that is not load-bearing for the logical validity of the supervision signal).

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The axiom ledger is lean. The method introduces one invented entity (the AR-consistent teacher cache) whose effect is tested through ablation. Free parameters are few and mostly justified by architecture constraints rather than fitted to evaluation metrics. Two training hyperparameters (LoRA rank, learning rate) are missing from the paper, which slightly reduces reproducibility confidence.

free parameters (5)
  • M (warm-up chunks) = 7
    Chosen to match the original local training horizon of Wan-based AR models (3×7=21 latent frames). Not fitted to the target metric; justified by architecture.
  • EMA decay rate = 0.9999
    Standard EMA hyperparameter for teacher copy; not tuned to the target result.
  • Training iterations = 200
    Number of post-training iterations; small number chosen empirically but not fitted to evaluation metrics.
  • LoRA rank = not specified
    The paper states LoRA-based training but does not specify the rank. This is a missing parameter.
  • Learning rate = not specified
    Not stated in the paper. A missing training hyperparameter.
axioms (4)
  • domain assumption Real long videos provide a cleaner target distribution for supervising long AR rollouts than short-clip teacher distributions.
    §1: 'Compared with relying only on a short-clip teacher distribution, real long videos offer a richer and cleaner target distribution.' This is the foundational premise; it is supported qualitatively by the Fig. 2 diagnostic but not quantitatively proven.
  • domain assumption Replacing older generated KV-cache history with real-video chunks yields a teacher velocity prediction that is a useful correction for the student's current state.
    §4.3: The teacher cache construction (Eq. 12) assumes semantic compatibility between real-video context and student trajectory. The paper acknowledges potential mismatch but does not bound the divergence.
  • domain assumption Velocity matching in the native prediction space is superior to clean-latent (x0) matching for long-rollout supervision.
    §5.4 ablation (Fig. 6) provides qualitative support; the claim about timestep-dependent reweighting under flow parameterization is stated but not formally derived.
  • domain assumption Stop-gradient through the solver transition (Eq. 6) does not materially harm the training signal.
    §4.2: sg(·) prevents backprop through AR history. This is a standard approximation in self-forcing methods but truncates credit assignment across chunks.
invented entities (1)
  • AR-consistent real-video teacher cache independent evidence
    purpose: Provides cleaner temporal context for teacher velocity predictions by replacing older generated history with real-video KV entries while retaining the most recent student-generated chunk.
    The construction is defined by Eq. 12 and its effect is demonstrated through the ablation in Fig. 7 (teacher-trajectory supervision fails) and the diagnostic in Fig. 2 (data-context intervention improves stability). The entity is a training-time construct, not a new physical or mathematical object.

pith-pipeline@v1.1.0-glm · 21239 in / 3321 out tokens · 237060 ms · 2026-07-10T01:30:45.473521+00:00 · methodology

0 comments
read the original abstract

We propose OPSD-V, an on-policy self-distillation paradigm for post-training few-step autoregressive (AR) video diffusion models. Existing few-step AR video generators can produce long videos with low latency, but still suffer from error accumulation and weakened motion dynamics during long autoregressive rollout. OPSD-V reduces long-horizon degradation while preserving the original few-step inference path. The key idea is to introduce real long-video data as temporal context during training and use it to provide dense trajectory-level supervision. Specifically, the student follows the exact inference-time rollout, generating each chunk conditioned on its own previously generated KV cache. In parallel, the teacher is evaluated at the same student-visited denoising states, but uses a cleaner AR-consistent temporal cache in which older history can be replaced by real-video context. This provides dense denoising-level corrective targets under on-policy AR cache dynamics, without changing the sampler, number of denoising steps, or inference-time cache mechanism. We apply OPSD-V to representative few-step AR video models, including Self-Forcing and LongLive. Experiments show consistent improvements in visual quality, motion dynamics, and VBenchLong scores. A user study with 10 participants comparing 20 video pairs shows that OPSD-V is preferred over the base models in 66.0% of overall-preference judgments (82.5% excluding ties).

Figures

Figures reproduced from arXiv: 2607.08766 by Chun Wang, Feng Gao, Hongyu Liu, Qifeng Chen, Xiaoming Wei, Xuanhua He, Yong Zhang, Yue Ma, Ziyu Wan.

Figure 1
Figure 1. Figure 1: OPSD-V improves long-horizon autoregressive video generation. We compare the original base model and its OPSD-V post-trained version under the same prompt and seed. The first example is based on LongLive, and the second example is based on Self-Forcing. In each example, the first row shows the base model result, while the second row shows the result after applying OPSD-V. Our method produces more dynamic a… view at source ↗
Figure 2
Figure 2. Figure 2: Overview and motivating cache intervention. Top: Unlike Self-Forcing training, which applies DMD su￾pervision after self-rollout, OPSD-V continues post-training an existing few-step AR generator by aligning an on-policy student with a teacher conditioned on cleaner long-video context. The student writes its own generated chunks into the KV cache, while the teacher is evaluated at the same denoising states … view at source ↗
Figure 3
Figure 3. Figure 3: OPSD-V post-trains a few-step autoregressive video generator with cache-aware on-policy self-distillation. The student follows the exact inference-time rollout and writes its own generated chunks into the KV cache. The teacher is evaluated at the same student-visited denoising states, but uses an AR-consistent real-video cache, where older history is replaced by real-video chunks while the most recent chun… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on long autoregressive video generation. The first two examples are based on [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: User study preference results. Ten participants compare 20 paired videos each, consisting of 10 LongLive pairs and 10 Self-Forcing pairs. Each pair contains a base-model video and its OPSD-V post-trained counterpart under the same prompt. For each pair, participants choose Model A, Model B, or Same for overall preference, motion quality, and visual quality. Each panel shows percentages over 100 judgments p… view at source ↗
Figure 6
Figure 6. Figure 6: Velocity versus clean-latent matching. We compare distillation in the predicted clean-latent (x0) space and in the native velocity space over a one-minute rollout. Both variants remain plausible at the beginning, but x0 matching progressively smooths fine structures and introduces geometric degradation in the bridge truss and foliage. Velocity matching better preserves these details at later timestamps. Re… view at source ↗
Figure 7
Figure 7. Figure 7: Failure of teacher-trajectory supervision. When distillation states are sampled from the teacher’s own denois￾ing trajectory, the teacher rollout remains sharp and coherent (bottom), but the student becomes severely blurred under its own autoregressive rollout (top). This discrepancy reveals the off-policy state mismatch introduced by teacher-trajectory supervision. Student versus teacher trajectory. A cen… 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

29 extracted references · 29 canonical work pages · 23 internal anchors

  1. [1]

    MAGI-1: Autoregressive Video Generation at Scale

    URLhttps: //arxiv.org/abs/2505.13211. Philip J. Ball, Jakob Bauer, Frank Belletti, Bethanie Brownfield, Ariel Ephrat, Shlomi Fruchter, Agrim Gupta, Kristian Holsheimer, Aleksander Holynski, Jiri Hron, Christos Kaplanis, Marjorie Limont, Matt McGill, Yanko Oliveira, Jack Parker-Holder, Frank Perbet, Guy Scully, Jeremy Shar, Stephen Spencer, Omer Tov, Ruben...

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion

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

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    SkyReels-V2: Infinite-length Film Generative Model

    URLhttps://arxiv.org/abs/2504.13074. Justin Cui, Jie Wu, Ming Li, Tao Yang, Xiaojie Li, Rui Wang, Andrew Bai, Yuanhao Ban, and Cho-Jui Hsieh. Self- forcing++: Towards minute-scale high-quality video generation.arXiv preprint arXiv:2510.02283,

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    Flow-OPD: On-Policy Distillation for Flow Matching Models

    Zhen Fang, Wenxuan Huang, Yu Zeng, Yiming Zhao, Shuang Chen, Kaituo Feng, Yunlong Lin, Lin Chen, Zehui Chen, Shaosheng Cao, and Feng Zhao. Flow-opd: On-policy distillation for flow matching models.arXiv preprint arXiv:2605.08063,

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [5]

    AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation

    Yuchao Gu, Guian Fang, Yuxin Jiang, Weijia Mao, Song Han, Han Cai, and Mike Zheng Shou. Anyflow: Any-step video diffusion model with on-policy flow map distillation.arXiv preprint arXiv:2605.13724,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    LTX-2: Efficient Joint Audio-Visual Foundation Model

    Yoav HaCohen, Benny Brazowski, Nisan Chiprut, Yaki Bitterman, Andrew Kvochko, Avishai Berkowitz, Daniel Shalem, Daphna Lifschitz, Dudu Moshe, Eitan Porat, et al. Ltx-2: Efficient joint audio-visual foundation model.arXiv preprint arXiv:2601.03233,

    work page internal anchor Pith review Pith/arXiv arXiv

  7. [7]

    Self-Distillation Zero: Self-Revision Turns Binary Rewards into Dense Supervision

    15 Yinghui He, Simran Kaur, Adithya Bhaskar, Yongjin Yang, Jiarui Liu, Narutatsu Ri, Liam Fowl, Abhishek Panigrahi, Danqi Chen, and Sanjeev Arora. Self-distillation zero: Self-revision turns binary rewards into dense supervision.arXiv preprint arXiv:2604.12002,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion

    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, 2025a. Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, et al. Vbench++: Comprehensive and versatile ...

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models

    Dengyang Jiang, Xin Jin, Dongyang Liu, Zanyi Wang, Mingzhe Zheng, Ruoyi Du, Xiangpeng Yang, Qilong Wu, Zhen Li, Peng Gao, Harry Yang, and Steven Hoi. D-opsd: On-policy self-distillation for continuously tuning step-distilled diffusion models.arXiv preprint arXiv:2605.05204,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [10]

    Pyramidal flow matching for efficient video generative modeling

    Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. InInternational Conference on Learning Representations, volume 2025, pp. 23378–23402,

    work page 2025

  11. [11]

    HunyuanVideo: A Systematic Framework For Large Video Generative Models

    Weijie Kong, Qi Tian, Zijian Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603,

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Rolling Sink: Bridging Limited-Horizon Training and Open-Ended Testing in Autoregressive Video Diffusion

    Haodong Li, Shaoteng Liu, Zhe Lin, and Manmohan Chandraker. Rolling sink: Bridging limited-horizon training and open-ended testing in autoregressive video diffusion.arXiv preprint arXiv:2602.07775, 2026a. Lin Li, Qihang Zhang, Yiming Luo, Shuai Yang, Ruilin Wang, Fei Han, Mingrui Yu, Zelin Gao, Nan Xue, Xing Zhu, Yujun Shen, and Yinghao Xu. Causal world m...

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Decoupled dmd: Cfg augmentation as the spear, distribution matching as the shield.arXiv preprint arXiv:2511.22677, 2025

    Dongyang Liu, Peng Gao, David Liu, Ruoyi Du, Zhen Li, Qilong Wu, Xin Jin, Sihan Cao, Shifeng Zhang, Hongsheng Li, and Steven Hoi. Decoupled dmd: Cfg augmentation as the spear, distribution matching as the shield.arXiv preprint arXiv:2511.22677, 2025a. Kunhao Liu, Wenbo Hu, Jiale Xu, Ying Shan, and Shijian Lu. Rolling forcing: Autoregressive long video dif...

    work page arXiv

  14. [14]

    Longcat-video technical report.arXiv preprint arXiv:2510.22200,

    Meituan LongCat Team, Xunliang Cai, Qilong Huang, Zhuoliang Kang, Hongyu Li, Shijun Liang, Liya Ma, Siyu Ren, Xiaoming Wei, Rixu Xie, and Tong Zhang. Longcat-video technical report.arXiv preprint arXiv:2510.22200,

    work page arXiv

  15. [15]

    Scalable Diffusion Models with Transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers.arXiv preprint arXiv:2212.09748,

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    Movie Gen: A Cast of Media Foundation Models

    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 foundation models.arXiv preprint arXiv:2410.13720,

    work page internal anchor Pith review Pith/arXiv arXiv

  17. [17]

    Self-Distillation Enables Continual Learning

    Idan Shenfeld, Mehul Damani, Jonas Hübotter, and Pulkit Agrawal. Self-distillation enables continual learning.arXiv preprint arXiv:2601.19897,

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Motionstream: Real-time video generation with interactive motion controls.arXiv preprint arXiv:2511.01266,

    Joonghyuk Shin, Zhengqi Li, Richard Zhang, Jun-Yan Zhu, Jaesik Park, Eli Shechtman, and Xun Huang. Motionstream: Real-time video generation with interactive motion controls.arXiv preprint arXiv:2511.01266,

    work page arXiv

  19. [19]

    Denoising Diffusion Implicit Models

    16 Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502,

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  20. [20]

    LongCat-Video-Avatar 1.5 Technical Report

    Meituan LongCat Team, Xunliang Cai, Meng Cheng, Feng Gao, Zhe Kong, Jiamu Li, Le Li, Weiheng Li, Hongyu Liu, Shuai Tan, et al. Longcat-video-avatar 1.5 technical report.arXiv preprint arXiv:2605.26486, 2026a. Robbyant Team, Zelin Gao, Qiuyu Wang, Yanhong Zeng, Jiapeng Zhu, Ka Leong Cheng, Yixuan Li, Hanlin Wang, Yinghao Xu, Shuailei Ma, Yihang Chen, Jie L...

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    Flowact-r1: Towards interactive humanoid video generation.arXiv preprint arXiv:2601.10103,

    Lizhen Wang, Yongming Zhu, Zhipeng Ge, Youwei Zheng, Longhao Zhang, Tianshu Hu, Shiyang Qin, Mingshuang Luo, Jiaxu Zhang, Xin Chen, et al. Flowact-r1: Towards interactive humanoid video generation.arXiv preprint arXiv:2601.10103,

    work page arXiv

  22. [22]

    InfiniteTalk: Audio-driven Video Generation for Sparse-Frame Video Dubbing

    Shaoshu Yang, Zhe Kong, Feng Gao, Meng Cheng, Xiangyu Liu, Yong Zhang, Zhuoliang Kang, Wenhan Luo, Xunliang Cai, Ran He, and Xiaoming Wei. Infinitetalk: Audio-driven video generation for sparse-frame video dubbing, 2025a. URLhttps://arxiv.org/abs/2508.14033. Shuai Yang, Wei Huang, Ruihang Chu, Yicheng Xiao, Yuyang Zhao, Xianbang Wang, Muyang Li, Enze Xie,...

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [23]

    World Action Models are Zero-shot Policies

    URLhttps: //arxiv.org/abs/2602.15922. Hidir Yesiltepe, Tuna Meral, Adil Kaan Akan, Kaan Oktay, and Pinar Yanardag. Infinity-rope: Action-controllable infinite video generation emerges from autoregressive self-rollout. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 40256–40265,

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    Deep forcing: Training-free long video generation with deep sink and participative compression.arXiv preprint arXiv:2512.05081,

    Jung Yi, Wooseok Jang, Paul Hyunbin Cho, Jisu Nam, Heeji Yoon, and Seungryong Kim. Deep forcing: Training-free long video generation with deep sink and participative compression.arXiv preprint arXiv:2512.05081,

    work page arXiv

  25. [25]

    One-step Diffusion with Distribution Matching Distillation

    Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation.arXiv preprint arXiv:2311.18828,

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [26]

    Reward-Forcing: Autoregressive Video Generation with Reward Feedback

    Jingran Zhang, Ning Li, Yuanhao Ban, Andrew Bai, and Justin Cui. Reward-forcing: Autoregressive video generation with reward feedback.arXiv preprint arXiv:2601.16933,

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models

    17 Siyan Zhao, Zhihui Xie, Mengchen Liu, Jing Huang, Guan Pang, Feiyu Chen, and Aditya Grover. Self-distilled reasoner: On-policy self-distillation for large language models.arXiv preprint arXiv:2601.18734,

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    Open-Sora: Democratizing Efficient Video Production for All

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all.arXiv preprint arXiv:2412.20404,

    work page internal anchor Pith review Pith/arXiv arXiv

  29. [29]

    Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation

    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 generation.arXiv preprint arXiv:2602.02214, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026