Pith. sign in

REVIEW 4 major objections 5 minor 33 references

Long video drift tracks collapse of the model's hidden representations, and regularizing those representations stops the compounding error.

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 ↗

Compounding error in autoregressive video diffusion tracks effective-rank collapse of DiT hidden states, and representation regularization (SigReg/Unif) stabilizes long rollouts where data scaling does not.

T0 review reviewed 2026-07-30 challenge →

load-bearing objection Useful Minecraft result and a clean anti-scaling observation, but the erank-collapse story is correlational and the ablations undercut the causal claim. the 4 major comments →

arxiv 2607.27036 v1 pith:4DYKY2IN submitted 2026-07-29 cs.CV cs.LG

Mitigating Compounding Error via Video Representation Regularization

Taiye Chen , Qi Zhang , Yisen Wang This is my paper
classification cs.CV cs.LG
keywords video diffusionautoregressive generationcompounding erroreffective rankrepresentation regularizationworld modelsDiffusion Forcinglong video generation
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 reading

Autoregressive video diffusion models used as world models degrade after tens of frames because small generation mistakes feed back as conditioning and snowball. This paper argues that the drift is not just a sampling nuisance: it coincides with a sharp drop in the effective rank of the model's hidden states, meaning the internal representation loses usable dimensions exactly when the video collapses into noise or overexposure. Scaling training data alone does not fix this and can make rank collapse worse, which undercuts the usual “more data helps” assumption for these models. The authors therefore add a light training regularizer on selected DiT hidden states so the representation stays high-rank and informative. On long Minecraft rollouts, that regularization keeps Aesthetic and Imaging Quality high far past the horizon where plain Diffusion Forcing fails, turning representation geometry into both a diagnostic and a control knob for stable long-horizon generation.

Core claim

Compounding error in sliding-window autoregressive video diffusion is tightly coupled with dimensional collapse of the DiT's hidden representations: effective rank falls sharply at the onset of visual drift, pure data or training-step scaling does not raise resistance to that drift, and adding a representation regularizer (especially SigReg or Uniformity on selected layers) substantially improves long-horizon Aesthetic and Imaging Quality versus Diffusion Forcing.

What carries the argument

Video representation regularization (VRR): the training loss is L_DF plus weighted regularizers on DiT hidden states (mainly SigReg or Uniformity), chosen to keep effective rank high and block shortcut frame-copying that collapses latent expressiveness.

Load-bearing premise

The load-bearing premise is that effective-rank collapse is a primary controllable cause of drift, so forcing healthier hidden-state geometry will systematically stop iterative error buildup rather than only tracking a symptom of distribution shift.

What would settle it

Train the same DiT with and without VRR on the same Minecraft setup, roll out 1100 frames from a 100-frame prompt, and check whether the frame where VBench Imaging/Aesthetic quality collapses still lines up with a sharp erank drop in intermediate layers—and whether VRR both prevents that erank drop and keeps quality high through the full minute.

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

If this is right

  • Effective rank of intermediate DiT states can be used as an online quantitative metric for when autoregressive video generation is about to drift.
  • Simply collecting more trajectories or training longer is not a reliable path to long-horizon robustness in video world models and may worsen representation collapse.
  • A lightweight hidden-state regularizer (SigReg/Unif on early, middle, and late layers) can replace or outperform frame-anchoring and error-recycling tricks for long rollouts.
  • Stable long video world models for robotics, driving, and simulation need explicit representation-capacity constraints, not only better noise schedules or context windows.

Where Pith is reading between the lines

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

  • If rank collapse is the bottleneck, other sequence generators that feed their own outputs back (not only video DiTs) may show the same erank–drift signature and benefit from analogous regularizers.
  • The reported “more data hurts” pattern suggests curriculum or anti-shortcut objectives may matter more than raw dataset size once the model can copy nearby frames.
  • Layer-selective regularization (first/middle/last) hints that collapse is not uniform across depth; probing which blocks lose rank first could guide cheaper partial regularizers.
  • VBench Imaging/Aesthetic gains on Minecraft leave open whether the same erank control stabilizes semantic consistency in open-world or action-conditioned non-game domains.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. The paper studies compounding error (drift) in sliding-window autoregressive video diffusion world models. It reports that visual collapse during long rollouts coincides with a sharp drop in the effective rank (erank) of DiT hidden states, that training longer on more MineRL Minecraft data does not improve and can worsen erank stability and VBench quality, and that adding a representation regularizer (VRR)—especially SigReg or Uniformity on selected layers—to the Diffusion Forcing objective substantially raises Aesthetic and Imaging Quality on VBench (e.g., 38.65→55.56 and 44.37→72.08 at 16k steps) relative to Diffusion Forcing, Frame Anchor, and SVI. The authors position erank as a quantitative metric for error accumulation and VRR as a lightweight fix motivated by representation collapse.

Significance. If the erank–drift coupling and the failure of pure data scaling generalize beyond the reported setting, the work would give the community a useful internal diagnostic and a concrete reason to treat long-horizon AR video stability as partly a representation-geometry problem rather than only a sampling or data-volume problem. The empirical gains of SigReg/Unif VRR over standard baselines on long Minecraft rollouts are practically relevant for video world models. Strengths include a clear experimental protocol (one-epoch 18k-trajectory scaling sweep, layer ablations, multiple regularizers, time-resolved VBench curves) and an explicit negative result on naïve scaling. The main significance is currently empirical and domain-limited; the stronger claim that dimensional collapse is the controllable mechanism of compounding error is not yet established by the interventions.

major comments (4)
  1. [§3.2, §4.1–4.2, Table 2] §3.2–§4.2 and Table 2: The central interpretive move—from erank collapse coinciding with visual drift (Fig. 2) to the claim that stabilizing representation geometry via L = L_DF + Σ λ_i L_reg(H_i) suppresses compounding error—is not supported by the paper’s own ablations. Direct erank regularization tracks Diffusion Forcing and fails at 16k steps (Aesthetic ~37, Imaging ~37), while SigReg and Uniformity succeed. That dissociation means VRR gains cannot be read as evidence that preventing dimensional collapse is why drift stops. Either provide a causal test (e.g., matched runs where delayed erank collapse predicts delayed drift, or interventions that change erank without changing other geometry) or substantially soften the mechanism language in the abstract, intro, and conclusion to a correlational observation plus an effective regularizer.
  2. [§3.3, Figure 3] §3.3 and Fig. 3: The claim that “more data cannot cure error accumulation” and that this contradicts mainstream scaling rests on one epoch over ~18k MineRL sequences with checkpoints along that single pass. This does not separate data volume from optimization trajectory, capacity, multi-epoch training, data diversity, or domain. A minimal strengthening would include multi-epoch vs. larger unique-data comparisons, at least one non-Minecraft domain, and reporting of diffusion loss alongside erank so under/overfitting can be distinguished from representation collapse. As written, the scaling conclusion is overstated relative to the design.
  3. [§5.1–5.2, Table 1] §5.1–5.2, Table 1, Figure 5–6: Evaluation is confined to Minecraft rollouts and two VBench axes (Aesthetic Quality, Imaging Quality) chosen because other VBench dimensions are “irrelevant.” Those two axes mainly capture noise/collapse, so they align with the failure mode but do not establish general long-video robustness (motion consistency, semantics, action-conditioning fidelity, or other datasets). Baselines (Frame Anchor, SVI) also underperform badly; without stronger AR long-video baselines or broader metrics, the large numerical gains are hard to calibrate. Expand evaluation or clearly scope claims to “collapse resistance on MineRL under this DiT/DF setup.”
  4. [§3.2, Figure 2] §3.2: The argument that erank is a “uniquely informative” indicator because SSIM/PSNR/LPIPS do not “consistently pinpoint the collapse frame” (Fig. 2, bottom) is weak as causal support. Once frames become near-constant or noise-like, low erank of hidden states is expected by construction; coincidence with visual collapse is then partly definitional. The paper should treat erank as a useful correlated diagnostic unless it shows predictive lead time or intervention specificity beyond output-space collapse detectors.
minor comments (5)
  1. [Table 1, Figure 5, §3.3] Typos and naming inconsistency: “Frame Archor” in Table 1 and Fig. 5; “na ïvely” spacing in §3.3; title line breaks (“MITIGATINGCOMPOUNDINGERROR”) look like PDF hyphenation artifacts—clean for camera-ready.
  2. [§4.1, Eq. (4)] Eq. (4): erank definition is standard; briefly state how H is shaped (tokens×dim vs. frames aggregated) and which singular values are kept in practice, so the metric is reproducible.
  3. [§4.2, Table 3] §4.2 / §5.4: Report the actual λ_i values, whether regularization is applied every step, and computational overhead of VRR; Table 3’s layer choices (0, 7, 15) need a one-sentence justification beyond “first/middle/last.”
  4. [§2.3] Related work (§2.3) cites representation regularizers (REPA, DispLoss, LeJEPA/SigReg) appropriately; a short explicit contrast on train-time vs. rollout-time representation degradation would sharpen novelty without overclaiming.
  5. [Figure 1] Figure 1 caption promises “consistent performance gains as training proceeds”; ensure all plotted methods share identical data order and seed reporting where possible.

Circularity Check

0 steps flagged

No derivation circularity: erank–drift link is correlational observation, VRR is an additive regularizer, and gains are measured on external VBench metrics.

full rationale

The paper’s chain is observational and empirical, not a closed derivation. Section 3.2 reports that effective rank of DiT hidden states drops when autoregressive frames visually collapse; Section 3.3 reports that longer training on more data does not raise erank or long-horizon quality; Section 4 then adds a standard representation regularizer, L = L_DF + Σ λ_i L_reg(H_i), and Section 5 evaluates Aesthetic/Imaging Quality on VBench against Diffusion Forcing, Frame Anchor, and SVI. Erank is an external linear-algebra summary of hidden states (Roy & Vetterli), not defined from the VBench scores or from the claimed drift mechanism. Successful L_reg choices (SigReg, Uniformity) are independent of the diagnostic; the ablation that uses erank itself as L_reg fails to match SigReg/Unif (Table 2), so success is not obtained by optimizing the reported metric by construction. Self-citations (e.g., Diffusion Forcing, VRAG data protocol) supply baselines and dataset setup, not a uniqueness theorem or load-bearing premise that forces the result. Concerns that erank collapse is a symptom rather than a controllable cause of drift are causal/interpretive, not circular reductions of equations or fits to their own inputs. No step reduces a claimed prediction to a fitted input or to a self-definition.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The central story rests on standard diffusion/AR training plus the modeling choice that hidden-state effective rank is the right lens on drift and that off-the-shelf representation regularizers on selected DiT layers are sufficient interventions. Free parameters are the usual training and regularization knobs; invented ‘entities’ are methodological (VRR, erank-as-drift-metric), not physical objects. No formal axioms beyond ordinary deep-learning practice.

free parameters (5)
  • Regularization weights λ_i per layer = Not numerically specified in the text
    Coefficients on L_reg(H_i) in the joint objective; chosen to stabilize training and reported gains, not derived.
  • Which DiT layers receive VRR = layers 0, 7, 15 (best reported)
    Ablation selects layers 0, 7, and 15 as best among tested subsets; this selection affects the headline numbers.
  • Choice and internals of L_reg (SigReg vs Unif vs alternatives) = SigReg and Uniformity preferred
    Family and hyperparameters of the regularizer are experimental choices; SigReg/Unif work, Barlow/VICReg/Erank less so.
  • Training horizon and checkpoint selection (e.g., 4k–16k/40k steps, one epoch) = Primary tables at 4k–16k steps; scaling plot to 40k
    Scaling and quality claims depend on where along training one evaluates; early checkpoints often look better for vanilla DF.
  • Sliding-window length and inference protocol = window=20; prompt=100; gen=1100
    Window size 20, 100 GT prompt frames, 1100 generated frames define the drift measurement setup.
axioms (5)
  • domain assumption Diffusion Forcing / latent video diffusion training and sliding-window autoregressive rollout are a valid model of long video world-model inference.
    §3.1 and §5.1 adopt DF, VAE latents, and windowed AR generation as the problem setting.
  • domain assumption Effective rank of intermediate DiT hidden states is a meaningful scalar for representational expressiveness and collapse during generation.
    §3.2–§4.1 define erank(H) via singular-value entropy and treat drops as representation collapse tied to drift.
  • domain assumption VBench Aesthetic Quality and Imaging Quality adequately quantify compounding-error severity in Minecraft rollouts.
    §5.1 discards other VBench axes as irrelevant and uses these two as primary evidence of mitigation.
  • ad hoc to paper One epoch over ~18k MineRL trajectories without multi-epoch overfitting is a fair test that ‘more data cannot cure’ error accumulation.
    §3.3 equates continued steps within one epoch on this corpus with data scaling and concludes scaling fails.
  • standard math Standard linear algebra and expectation identities used in erank and diffusion loss definitions.
    Eq. (4) and diffusion objectives use conventional SVD/entropy and MSE noise prediction.
invented entities (2)
  • Video Representation Regularization (VRR) no independent evidence
    purpose: Name the training objective that adds hidden-state regularizers to Diffusion Forcing to reduce long-horizon drift.
    Method label for L_DF + Σ λ_i L_reg(H_i); not an external physical entity. Independent evidence is only the paper’s own ablations.
  • Erank-as-quantitative-metric for video error accumulation no independent evidence
    purpose: Promote effective rank of DiT activations as a diagnostic that pins the collapse frame better than SSIM/PSNR/LPIPS.
    Repurposes Roy & Vetterli effective rank as a drift meter; falsifiable outside the paper only if others reproduce the coincidence on new models/domains.

reviewed 2026-07-30 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating Compounding Error via Video Representation Regularization." pith.science (2026). https://pith.science/paper/4DYKY2IN

@misc{pith2026260727036,
  author       = {Pith},
  title        = {Pith review of: Mitigating Compounding Error via Video Representation Regularization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4DYKY2IN}},
  note         = {Machine review of arXiv:2607.27036}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Video diffusion-based world models enable long autoregressive video generation for robotics, autonomous driving and simulation tasks, yet sliding-window autoregressive inference suffers from severe error accumulation that degrades frame quality over time. Although this phenomenon has been widely observed, the underlying mechanism of compounding error and how to achieve stable long-horizon generation remain largely unresolved. In this paper, we investigate the internal representation dynamics of video world models and discover that compounding error is tightly coupled with dimensional collapse of hidden representations. Specifically, the effective rank of model representations sharply decreases at the onset of generation drift, revealing a strong connection between representational degradation and long-term rollout instability. Furthermore, we find that pure training data scaling fails to boost model resistance to error drift, contradicting mainstream scaling paradigms. To address this problem, we propose video representation regularization, a lightweight training constraint that stabilizes latent representations and suppresses iterative error accumulation. Compared with Diffusion Forcing, our method achieves improvements from 38.65 to 55.56 and from 44.37 to 72.08 on the Aesthetic Quality and Imaging Quality metrics of VBench. Our work establishes the first connection between autoregressive video drifting and model internal representations, adopts erank as a quantitative metric for error accumulation, reveals counterintuitive scaling limitations for video world models, and presents a simple yet effective regularization strategy to improve long video generation robustness.

Figures

Figures reproduced from arXiv: 2607.27036 by Qi Zhang, Taiye Chen, Yisen Wang.

Figure 1
Figure 1. Figure 1: Overview of our proposed method. Existing autoregressive diffusion models suffer from [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Top row: Video frames sampled from a 40k-step checkpoint trained with the vanilla dif￾fusion forcing method. Hidden states from DiT’s 7th layer are extracted to analyze the effective rank (Erank), which reveals a direct correlation between visual collapse and abrupt Erank reduction. Middle row: Samples generated by a 4k-step checkpoint under the same training scheme; stable Er￾ank values are observed when … view at source ↗
Figure 3
Figure 3. Figure 3: We trained for one epoch on the 18k training set using the vanilla diffusion forcing method [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Architecture of DiT model. Model Architecture As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Frames of videos generated from the checkpoint of all methods at the 16,000th training [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The video quality of vanilla Diffusion Forcing continuously degrades as the video pro [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] 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

33 extracted references · 26 linked inside Pith

  1. [2]

    Lejepa: Provable and scalable self-supervised learning with- out the heuristics.arXiv preprint arXiv:2511.08544,

    Randall Balestriero and Yann LeCun. Lejepa: Provable and scalable self-supervised learning with- out the heuristics.arXiv preprint arXiv:2511.08544,

    Pith/arXiv arXiv

  2. [3]

    Amir Bar, Gaoyue Zhou, Danny Tran, Trevor Darrell, and Yann LeCun

    URLhttps: //deepmind.google/blog/genie-3-a-new-frontier-for-world-models/. Amir Bar, Gaoyue Zhou, Danny Tran, Trevor Darrell, and Yann LeCun. Navigation world models. arXiv preprint arXiv:2412.03572,

    Pith/arXiv arXiv

  3. [6]

    URLhttps://oasis-model.github. io/. Zihan Ding, Chi Jin, Difan Liu, Haitian Zheng, Krishna Kumar Singh, Qiang Zhang, Yan Kang, Zhe Lin, and Yuchen Liu. Dollar: Few-step video generation via distillation and latent reward optimization.arXiv preprint arXiv:2412.15689,

    Pith/arXiv arXiv

  4. [7]

    Pre-trained language model representations for language generation

    Sergey Edunov, Alexei Baevski, and Michael Auli. Pre-trained language model representations for language generation. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, V olume 1 (Long and Short Papers), pp. 4052–4059,

    2019

  5. [9]

    Minerl: A large-scale dataset of minecraft demonstrations

    William H Guss, Brandon Houghton, Nicholay Topin, Phillip Wang, Cayden Codel, Manuela Veloso, and Ruslan Salakhutdinov. Minerl: A large-scale dataset of minecraft demonstrations. arXiv preprint arXiv:1907.13440,

    Pith/arXiv arXiv 1907

  6. [13]

    Cogvideo: Large-scale pretraining for text-to-video generation with transformers.arXiv preprint arXiv:2205.15868,

    Yu Hong, Jing Wei, Xing Liu, Xiaodi Wang, Yutong Bai, Haitao Li, Ming Zhang, and Hao Xu. Cogvideo: Large-scale pretraining for text-to-video generation with transformers.arXiv preprint arXiv:2205.15868,

    Pith/arXiv arXiv

  7. [14]

    Gaia-1: A generative world model for autonomous driving.arXiv preprint arXiv:2309.17080,

    Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shot- ton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving.arXiv preprint arXiv:2309.17080,

    Pith/arXiv arXiv

  8. [15]

    Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin

    doi: 10.1109/CVPR52733.2024.02060. 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,

    arXiv 2024

  9. [16]

    URLhttps://proceedings.neurips.cc/paper_files/paper/ 2024/file/e304d374c85e385eb217ed4a025b6b63-Paper-Conference.pdf

    doi: 10.52202/ 079017-3985. URLhttps://proceedings.neurips.cc/paper_files/paper/ 2024/file/e304d374c85e385eb217ed4a025b6b63-Paper-Conference.pdf. Wuyang Li, Wentao Pan, Po-Chien Luan, Yang Gao, and Alexandre Alahi. Stable video infinity: Infinite-length video generation with error recycling. InInternational Conference on Learning Representations 2025 (ICLR 2025),

    2024

  10. [17]

    Genie envisioner: A unified world foundation platform for robotic manipulation.arXiv preprint arXiv:2508.05635,

    Yue Liao, Pengfei Zhou, Siyuan Huang, Donglin Yang, Shengcong Chen, Yuxin Jiang, Yue Hu, Jingbin Cai, Si Liu, Jianlan Luo, Liliang Chen, Shuicheng Yan, Maoqing Yao, and Guanghui Ren. Genie envisioner: A unified world foundation platform for robotic manipulation.arXiv preprint arXiv:2508.05635,

    Pith/arXiv arXiv

  11. [18]

    Leworld- model: Stable end-to-end joint-embedding predictive architecture from pixels.arXiv preprint arXiv:2603.19312,

    Lucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun, and Randall Balestriero. Leworld- model: Stable end-to-end joint-embedding predictive architecture from pixels.arXiv preprint arXiv:2603.19312,

    Pith/arXiv arXiv

  12. [19]

    URL https://arxiv.org/abs/2501.03575. Jack Parker-Holder, Philip Ball, Jake Bruce, Vibhavari Dasagi, Kristian Holsheimer, Chris- tos Kaplanis, Alexandre Moufarek, Guy Scully, Jeremy Shar, Jimmy Shi, Stephen Spencer, Jessica Yung, Michael Dennis, Sultan Kenjeyev, Shangbang Long, Vlad Mnih, Harris Chan, Maxime Gazeau, Bonnie Li, Fabio Pardo, Luyu Wang, Lei ...

    Pith/arXiv arXiv

  13. [20]

    Charlotte Peale, Vinod Raman, and Omer Reingold

    URLhttps://deepmind.google/discover/blog/ genie-2-a-large-scale-foundation-world-model/. Charlotte Peale, Vinod Raman, and Omer Reingold. Representative language generation.arXiv preprint arXiv:2505.21819,

    Pith/arXiv arXiv

  14. [21]

    Freenoise: Tuning-free longer video diffusion via noise rescheduling

    Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. InInternational Conference on Learning Representations, volume 2024, pp. 5260–5274,

    2024

  15. [22]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer

    URLhttps://arxiv.org/ abs/2506.09042. Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition, pp. 10684–10695,

    Pith/arXiv arXiv

  16. [23]

    Make-a-video: Text-to-video generation without text-video data.arXiv preprint arXiv:2209.14792,

    Uriel Singer, Adam Polyak, Eliya Nachmani, Guy Dahan, Eli Shechtman, and Haggai Ha- cohen. Make-a-video: Text-to-video generation without text-video data.arXiv preprint arXiv:2209.14792,

    Pith/arXiv arXiv

  17. [24]

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu

    URLhttps://arxiv.org/abs/2502.06764. Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: En- hanced transformer with rotary position embedding.Neurocomputing, 568:127063,

    Pith/arXiv arXiv

  18. [25]

    Diffusion models are real-time game engines.arXiv preprint arXiv:2408.14837,

    Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines.arXiv preprint arXiv:2408.14837,

    Pith/arXiv arXiv

  19. [26]

    Mod- elscope text-to-video technical report.arXiv preprint arXiv:2308.06571,

    Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Mod- elscope text-to-video technical report.arXiv preprint arXiv:2308.06571,

    Pith/arXiv arXiv

  20. [27]

    Wenming Weng, Ruoyu Feng, Yanhui Wang, Qi Dai, Chunyu Wang, Dacheng Yin, Zhiyuan Zhao, Kai Qiu, Jianmin Bao, Yuhui Yuan, et al

    URLhttps://arxiv.org/abs/2506.09027. Wenming Weng, Ruoyu Feng, Yanhui Wang, Qi Dai, Chunyu Wang, Dacheng Yin, Zhiyuan Zhao, Kai Qiu, Jianmin Bao, Yuhui Yuan, et al. Art-v: Auto-regressive text-to-video generation with diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7395–7405,

    Pith/arXiv arXiv

  21. [28]

    Progressive autoregressive video diffusion models.arXiv preprint arXiv:2410.08151,

    Desai Xie, Zhan Xu, Yicong Hong, Hao Tan, Difan Liu, Feng Liu, Arie Kaufman, and Yang Zhou. Progressive autoregressive video diffusion models.arXiv preprint arXiv:2410.08151,

    Pith/arXiv arXiv

  22. [29]

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

    Pith/arXiv arXiv

  23. [30]

    Gamefactory: Creating new games with generative interactive videos.arXiv preprint arXiv:2501.08325, 2025a

    Jiwen Yu, Yiran Qin, Xintao Wang, Pengfei Wan, Di Zhang, and Xihui Liu. Gamefactory: Creating new games with generative interactive videos.arXiv preprint arXiv:2501.08325, 2025a. Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is ...

    arXiv

  24. [31]

    13 A Preprint Boyang Zheng, Nanye Ma, Shengbang Tong, and Saining Xie

    URLhttps://arxiv.org/abs/2504.12626. 13 A Preprint Boyang Zheng, Nanye Ma, Shengbang Tong, and Saining Xie. Diffusion transformers with repre- sentation autoencoders,

    arXiv

  25. [32]

    Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You

    URLhttps://arxiv.org/abs/2510.11690. 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,

    Pith/arXiv arXiv

  26. [33]

    URLhttps://arxiv.org/abs/ 2411.04983. 14

    Pith/arXiv arXiv

  27. [2018]

    Flexible diffusion modeling of long videos.arXiv preprint arXiv:2205.11495,

    William Harvey, Søren Nørskov, Niklas K¨olch, and George V ogiatzis. Flexible diffusion modeling of long videos.arXiv preprint arXiv:2205.11495,

    Pith/arXiv arXiv

  28. [2019]

    World models.arXiv preprint arXiv:1803.10122,

    David Ha and J¨urgen Schmidhuber. World models.arXiv preprint arXiv:1803.10122,

    Pith/arXiv arXiv

  29. [2022]

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models.arXiv preprint arXiv:2204.03458,

    Pith/arXiv arXiv

  30. [2023]

    Junliang Guo, Yang Ye, Tianyu He, Haoyu Wu, Yushu Jiang, Tim Pearce, and Jiang Bian

    doi: 10.1109/ ICCV51070.2023.00675. Junliang Guo, Yang Ye, Tianyu He, Haoyu Wu, Yushu Jiang, Tim Pearce, and Jiang Bian. Mineworld: a real-time and open-source interactive world model on minecraft.arXiv preprint arXiv:2504.08388,

    arXiv 2023

  31. [2024]

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al

    URLhttps://arxiv.org/abs/2404.08471. Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023a. Andreas Blattmann, Robin Rombach, Huan Lin...

    Pith/arXiv arXiv

  32. [2025]

    Alisson Azzolini, Hannah Brandon, Prithvijit Chattopadhyay, Huayu Chen, Jinju Chu, Yin Cui, Jenna Diamond, Yifan Ding, Francesco Ferroni, Rama Govindaraju, et al

    URL https://arxiv.org/abs/2506.09985. Alisson Azzolini, Hannah Brandon, Prithvijit Chattopadhyay, Huayu Chen, Jinju Chu, Yin Cui, Jenna Diamond, Yifan Ding, Francesco Ferroni, Rama Govindaraju, et al. Cosmos-reason1: From physical common sense to embodied reasoning.arXiv preprint arXiv:2503.15558,

    Pith/arXiv arXiv

  33. [2026]

    Self-forcing++: Towards minute-scale high-quality video generation.arXiv preprint arXiv:2510.02283,

    10 A Preprint 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,

    Pith/arXiv arXiv

This paper was first reviewed by grok-4.5 on July 30, 2026.