Pith. sign in

REVIEW 6 major objections 5 minor 2 cited by

InfLVG: Reinforce Inference-Time Consistent Long Video Generation with GRPO

T0 review · 6 major / 5 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 a GRPO-trained token-selection policy lets an autoregressive video model extend coherent multi-scene videos up to 9$\times$ its training length, at fixed per-step cost.

desk verdict A genuinely new inference-time context-selection idea for long autoregressive video, but the evaluation is circular: the rewards are the same metrics as Table 1 and the policy is fine-tuned per test scene while baselines are zero-shot. read the letter →

arxiv 2505.17574 v1 pith:HRP6XPKG submitted 2025-05-23 cs.CV

Xueji Fang , Liyuan Ma , Zhiyang Chen , Mingyuan Zhou , Guo-jun Qi This is my paper
classification cs.CV
keywords longvideogenerationautoregressivediffusioncontextselectionKVcacheGRPOreinforcementlearningcross-sceneconsistencypromptalignment
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

Generating a long video scene by scene with an autoregressive model usually degrades: the model keeps attending to the whole history, so the first scene's background and props bleed into later scenes and the new prompt is ignored. This paper claims the fix is to learn which pieces of history matter and to discard the rest. InfLVG trains a small context-selection policy with GRPO to score every past token against the current prompt, sample a fixed top-$K$ subset, and feed only those tokens to the video transformer. The hybrid reward, combining face-identity similarity, text-video similarity, and a vision-language artifact check, steers the policy toward tokens that preserve the subject while following the new scene description. If correct, the paper shows that long, multi-scene video generation can be achieved without long-video training data and with bounded inference cost.

What carries the argument

The load-bearing object is the context selection model $F_\theta$: a small network that fuses the current prompt with the KV cache through cross-attention and outputs a per-token relevance score. Those scores feed a Plackett–Luce sampler, a ranking distribution that draws a ranked list of $K$ tokens in proportion to score without replacement, and the video transformer attends only to the retrieved key-value pairs, so the context window stays at $K$ tokens no matter how long the video becomes. During GRPO training, group-normalized advantages push $F_\theta$ toward selections that maximize the hybrid reward; the paper's ablations show each reward component is necessary, and that $K = 6hw$ balances identity retention against prompt adaptability.

What would settle it

Run InfLVG on a held-out Event Prompt Set whose subject identity and background combinations never appeared during the paper's optimization, with the same 20-iteration per-scene budget, and compare cross-clip face-embedding similarity and text-video alignment against the sliding-window baseline; if the held-out overall consistency score does not exceed the sliding-window score, the generalization claim fails.

Watch

Extended reading notes

Core claim

The central claim is that the failure mode of autoregressive video extension is a context-management problem, not a generation-capacity problem. When the next scene is conditioned on the full key-value memory of past frames (the KV cache), outdated tokens dominate attention and the model reproduces the initial scene instead of the new prompt. InfLVG instead learns a policy $\pi_\theta$ that, for each new prompt, assigns every cached token a relevance score, draws a ranked top-$K$ subset via Plackett–Luce sampling, and restricts attention to those $K$ tokens. Training this policy with GRPO under a hybrid reward $r = r_{\mathrm{content}} + r_{\mathrm{clip}} + r_{\mathrm{artifact}}$ makes the selected context preserve identity, obey the new text, and avoid mosaic artifacts. On the paper's CsVBench benchmark, this yields an overall consistency score of 0.430 versus 0.356 for vanilla extension and up to a 9$\times$ increase in generated length at fixed per-step cost.

Load-bearing premise

The claim stands on the assumption that a context-selection policy trained for just 20 gradient steps per scene generalizes to new scenes, subjects, and backgrounds without retuning, so the reported gains are not an artifact of overfitting to the benchmark prompts.

Editorial extensions

If this is right

  • Existing single-scene video models can be turned into long-video generators without being retrained on long-form data; only the lightweight selection policy is updated at inference time.
  • Per-segment computation stays bounded because attention touches exactly $K$ cached tokens, so extending from one scene to $N$ scenes costs about $N$ times a single-scene pass rather than growing quadratically with total length.
  • The learned selector resolves the consistency-versus-alignment tradeoff: at small $K$ it concentrates on identity-critical face tokens, while larger $K$ drags in background tokens that weaken prompt adherence.
  • The same framework transfers to non-human subjects by replacing the face-identity reward with subject-region segmentation features, as demonstrated for dogs and cars.

Reading between the lines

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

  • The paper leaves open whether the 20-iteration per-scene optimization can be amortized; meta-learning the selector across many event-prompt sets would test whether a single prior policy needs almost no adaptation on a new subject.
  • The selected token indices could be read as an interpretable salience map of which earlier regions matter, and could in principle be exposed to a user as an editable attention mask.
  • The hybrid-reward structure suggests a compositional recipe: adding object-permanence, lighting-motion, or camera-motion rewards would likely extend consistency beyond human faces and into arbitrary objects.
  • Because the context stays fixed at $K$ tokens, the selector is effectively a memory-management policy; coupling it with a long-horizon planner could let tokens crucial for later scenes be deliberately pinned across scene cuts.
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

6 major / 5 minor

Summary. The paper proposes InfLVG, an inference-time context-selection framework for autoregressive video diffusion models. It trains a Plackett-Luce policy F_theta via GRPO to select top-K tokens from the history KV cache at each scene transition, using hybrid rewards that combine ArcFace identity similarity, CLIP prompt alignment, and QWen-based artifact detection. The authors introduce a Cross-scene Video Benchmark (CsVBench) with Event Prompt Sets (EPS) and report up to 9x video-length extension with improved consistency and alignment. The core evaluation, however, is complicated by the fact that Section 4.1 states the policy is optimized for 20 iterations per test scene, which makes the comparison against zero-shot baselines difficult to interpret.

Significance. The problem of long, cross-scene video generation is timely, and the idea of learning to select a fixed-size context window at inference time is genuinely interesting. The proposed benchmark CsVBench and the event-prompt construction are useful assets, and the extension to non-human subjects via SAM/DINOv2 in Appendix C indicates generality beyond the ArcFace-based main experiments. The paper also states that code will be released, which would aid reproducibility. However, the central quantitative claim is currently undermined by the combination of per-scene test-time optimization and the overlap between the reward functions and the evaluation metrics. If the authors can demonstrate the learned policy's effectiveness in a zero-shot setting or against equally expensive baselines, the contribution would be much stronger.

major comments (6)
  1. [Section 4.1; Table 1] The evaluation is not apples-to-apples. The paper states, 'For each scene generation, we optimized the policy model for 20 iterations during inference.' With group size 10, this amounts to roughly 200 generated candidate scenes per test scene plus policy gradient updates before the reported output is produced. The baselines (Vanilla, Random Per-token/Frame, Sliding Window, Global-Local, Random Global-Local) are zero-shot and receive no test-time search. Therefore the improvements in Table 1 may be due to per-scene reward maximization rather than to the learned context-selection mechanism. The 'up to 9x with better consistency and alignment' claim is not supported unless the policy is evaluated zero-shot after training on separate scenes, or the baselines are given an equivalent test-time budget.
  2. [Section 3.3, Eq. (9); Table 1] The optimized quantity is also the measured quantity. The reward in Eq. (9) combines rcontent (ArcFace pairwise cosine), rclip (CLIP similarity), and rartifact (QWen2.5-VL artifact detection), while Table 1 reports ArcFace-42M/ArcFace-360K, CLIP-Flan/ViCLIP, and QWen scores. Since the policy is tuned per scene on exactly these reward families (the exact checkpoints are not specified), the reported gains on those metrics are circular with respect to the method's objective. The paper should report at least one set of metrics not used as rewards, and should state whether the reward and evaluation checkpoints are identical.
  3. [Section 3.1, Eq. (4)] The removal of the KL penalty from GRPO is not justified. The text says 'We remove the KL penalty term from vanilla GRPO objective to eliminate reference model overhead.' Without KL regularization, 20 iterations of per-scene policy optimization can exploit idiosyncrasies of the reward models, especially when the reward and evaluation metrics overlap. The authors should either keep a KL term or ablate this choice, for instance by comparing against KL-regularized GRPO on held-out scenes.
  4. [Section 3.2; Section 4.1] The 'fixed computational budget' claim is misleading at the system level. The fixed-KV-cache argument bounds only per-step attention cost because K is fixed. The end-to-end inference budget includes the 20 GRPO iterations per scene, group size 10, generation of roughly 200 candidate scenes per test scene, and policy gradient updates. The paper should report total wall-clock time or FLOPs for Ours versus the baselines if it claims a fixed budget.
  5. [Section 3.2; Section 4.1] The context selection model and training procedure are underspecified. The paper gives only 'N1 cross-attention blocks' and 'N2 lightweight linear projection layers' and states N1=1, N2=2, but does not specify how token features are extracted from the KV cache, the parameter count, initialization, whether F_theta is shared across scenes or reinitialized per scene, the policy loss masking, or the exact GRPO update. These details are necessary to reproduce the method and to assess whether the policy generalizes beyond the scenes on which it was optimized.
  6. [Table 1] No error bars or statistical significance tests are reported. Several key differences are within 0.01 (e.g., ViCLIP 0.2003 for Ours vs. 0.1999 for Random Global-local; QWen text alignment 0.5484 for Ours vs. 0.5500 for Random Per-token; HPSv2 0.2720 for Ours vs. 0.2730 for Global-Local). Without variance over EPS subsets or seeds, the claimed superiority of InfLVG over the strongest baselines is not established.
minor comments (5)
  1. [Table 2] Table 2 reports CLIP-Flan and ViCLIP values as 60.51 and 20.08, while Table 1 reports the same measures as decimals (0.5523 and 0.2003); the units should be stated consistently.
  2. [Appendix B.2; Listing 1] The paper contains several typos and artifacts, including 'degration' in Appendix B.2 and 'te l ep ho ne' in Listing 1; the main text also contains placeholder-like strings such as 'figure2，abstract pipeline' before Figure 3.
  3. [Section 4.2; Table 2] The symbol K=6hw is used without defining h and w; they should be defined as the latent spatial dimensions of the video token grid.
  4. [References] References [23] and [44] appear to be the same paper; this should be consolidated.
  5. [Appendix C] Appendix C claims generalization to non-human subjects but provides only qualitative examples (Figure 11); a quantitative table for the DINOv2-based reward variant would strengthen the claim.

Circularity Check

3 steps flagged · score 6.0 of 10

Reward functions double as evaluation metrics and per-scene inference-time GRPO tunes the policy on the evaluated scenes, making the reported consistency and alignment gains partly an optimization artifact.

  1. fitted input called prediction [Section 3.3, Eq. (7)-(9); Section 4.2 Table 1; Appendix A.2]
    "rcontent(Vnl:n(l+1)−1, V<nl) = 1/E2 EXi=1 EXj=1 sim(φ(V cur j ), φ(V prev i )), (7) ... r = rcontent + rclip + rartifact. (9) ... we employ two versions of ArcFace: ArcFace-360k [10] ... and ArcFace-42M [26] ... to assess the consistency of human identity across different scenes."

    The content-consistency reward in Eq. (7) is the same ArcFace pairwise-cosine construction reported as the ArcFace-42M/ArcFace-360K cross-scene consistency metrics in Table 1 and Appendix A.2. GRPO maximizes this reward during inference, so the reported consistency scores are the training objective itself rather than independent evidence for the learned context-selection policy. The per-scene optimization also means the policy is fitted to the evaluated scenes, while the zero-shot baselines receive no such reward maximization.

  2. fitted input called prediction [Section 3.3, Eq. (9); Section B.1; Appendix A.2; Table 1]
    "we leverage a Vision-Language Model (VLM) to detect artifacts through binary classification ... rartifact is set to 1 if no artifact is detected, and 0 otherwise. ... QWen2.5-VL [3] is utilized to evaluate the aforementioned aspects ... provides a score to judge the presence of color mosaic artifacts in the generated videos."

    The artifact reward is computed by QWen2.5-VL answering whether colorful mosaic-like artifacts are present, and the same QWen2.5-VL model is then used as a headline evaluator in Table 1, including judging color mosaic artifacts and video quality. Optimizing the policy to satisfy QWen2.5-VL's binary judgment and then reporting QWen2.5-VL scores as evidence of artifact suppression is a closed evaluation loop.

1 more flagged steps
  1. other [Section 4.1, Inference-time GRPO Training Details]
    "During GRPO training, we set the group size to 10 to ensure thorough exploration. ... For each scene generation, we optimized the policy model for 20 iterations during inference."

    The policy is optimized at inference time for every scene that is subsequently scored in Table 1, with 20 GRPO iterations and group size 10, i.e., roughly 200 candidate scene generations per evaluated scene. The baselines are zero-shot and receive no such test-time search. Thus the headline '9× longer with better consistency and alignment' compares a per-scene fitted policy against unfitted baselines, so the improvement does not demonstrate generalization of the learned context-selection mechanism.

full rationale

The core context-selection mechanism is not circular by construction: a GRPO-trained policy scores and samples top-K KV-cache tokens, and the attention computation in Eq. (6) genuinely depends on that new module. No load-bearing self-citation chain or imported uniqueness theorem is present. The circularity is concentrated in the reward/evaluation identity and test-time fitting. Eq. (7)'s content reward is the ArcFace pairwise-cosine score that Appendix A.2 and Table 1 report as cross-scene consistency; Eq. (9)'s artifact term is a binary QWen2.5-VL judgment, and the same QWen2.5-VL model is a headline scorer in Table 1. Moreover, Section 4.1 states the policy is optimized for 20 iterations per scene at inference, i.e., on the very CsVBench scenes that are scored, while all baselines are zero-shot. Reporting 'up to 9× longer with strong consistency and semantic fidelity' after per-scene reward maximization on those same metrics is therefore partly a statement about the optimization objective, not an independent validation of the learned selection policy. The claim that the method maintains a fixed KV-cache context is also only a per-step attention bound; the end-to-end inference budget includes the roughly 200 candidate generations per scene. Overall, the paper has a real mechanism and independent backbone, but its central empirical evidence is substantially contaminated by the reward-metric overlap and per-scene test-time fitting, giving a partial circularity score of 6.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim depends on several fitted parameters (K, reward weights, iterations) and domain assumptions about the base model and reward functions. No new physical or conceptual entities are introduced.

free parameters (3)
  • Context length K = 6hw
    K is the number of context tokens selected. The paper tests K=3hw, 6hw, 12hw and chooses 6hw as the best balance, making it a hyperparameter fitted to the benchmark.
  • Reward weights = 1 (equal weights)
    The hybrid reward is a sum of three components with equal weight. This weighting is not justified or explored, so it is an arbitrary choice affecting the optimization.
  • Number of GRPO iterations per scene = 20
    The policy is optimized for 20 iterations per scene during inference; this number is reported but not varied or justified.
assumptions (3)
  • domain assumption The pretrained base model (CausVid) is a suitable backbone for long video generation when given selective context.
    The method builds on CausVid, a distilled autoregressive video diffusion model. The paper assumes its generation quality remains acceptable when only a subset of context tokens is provided.
  • domain assumption ArcFace similarity is a valid proxy for content consistency, and CLIP similarity is a valid proxy for prompt alignment.
    The reward and evaluation rely on these pretrained models, assuming they capture the desired semantic properties for generated videos.
  • ad hoc to paper The Plackett-Luce sampling probability can be optimized by GRPO without a KL penalty.
    The GRPO objective removes the KL term to avoid a reference model, which is a non-standard modification that may affect training stability but is adopted without analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of InfLVG: Reinforce Inference-Time Consistent Long Video Generation with GRPO." pith.science (2026). https://pith.science/paper/HRP6XPKG

@misc{pith2026250517574,
  author       = {Pith},
  title        = {Pith review of: InfLVG: Reinforce Inference-Time Consistent Long Video Generation with GRPO},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HRP6XPKG}},
  note         = {Machine review of arXiv:2505.17574}
}
abstract

Recent advances in text-to-video generation, particularly with autoregressive models, have enabled the synthesis of high-quality videos depicting individual scenes. However, extending these models to generate long, cross-scene videos remains a significant challenge. As the context length grows during autoregressive decoding, computational costs rise sharply, and the model's ability to maintain consistency and adhere to evolving textual prompts deteriorates. We introduce InfLVG, an inference-time framework that enables coherent long video generation without requiring additional long-form video data. InfLVG leverages a learnable context selection policy, optimized via Group Relative Policy Optimization (GRPO), to dynamically identify and retain the most semantically relevant context throughout the generation process. Instead of accumulating the entire generation history, the policy ranks and selects the top-$K$ most contextually relevant tokens, allowing the model to maintain a fixed computational budget while preserving content consistency and prompt alignment. To optimize the policy, we design a hybrid reward function that jointly captures semantic alignment, cross-scene consistency, and artifact reduction. To benchmark performance, we introduce the Cross-scene Video Benchmark (CsVBench) along with an Event Prompt Set (EPS) that simulates complex multi-scene transitions involving shared subjects and varied actions/backgrounds. Experimental results show that InfLVG can extend video length by up to 9$\times$, achieving strong consistency and semantic fidelity across scenes. Our code is available at https://github.com/MAPLE-AIGC/InfLVG.

Figures

Figures reproduced from arXiv: 2505.17574 by the authors.

Figure 1
Figure 1. Challenges in autoregressive long video generation across scenes. (Left) Baseline models [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Different scene extension paradigms with InfLVG. (a) Single-scene extension, (b) Multi [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. GRPO training pipeline. The DiT-based autoregressive video model generates a group [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Illustration of Context Selection Model [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Comparison of different context selection designs under cross-scene video generation . [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Ablation study on hybrid rewards. We compare our proposed InfLVG against various context selection strategies for data-free cross￾scene autoregressive video generation. Initially, we employ the original autoregressive model for long video generation beyond its training…
Figure 7
Figure 7. Figure 7: Analysis of context selection strategies under two settings: (a) Varying selected context [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Illustration of clip-wise cross-scene cosine similarity calculation. [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Visualization of mosaic color artifact detection via VLM [ [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Vanilla autoregressive long video extension with different implementations . (a) Official [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: InfLVG generates consistent long videos of generalized subjects (e.g., dogs, cars) across [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. LoViC: Efficient Long Video Generation with Context Compression

    cs.CV 2025-07 conditional novelty 6.0 of 10

    LoViC uses FlexFormer, a single-query-token Q-Former with interpolated rotary positional encoding, to compress long video-text context for efficient long-video generation.

  2. Rethinking Reward Signals in Video GRPO: When Scores Become Targets

    cs.CV 2025-11 reject novelty 5.0 of 10

    A self-paced GRPO variant that adaptively reweights visual, temporal, and text-alignment reward components as the generator improves reports small VBench gains over static-reward baselines.

Reference graph

Works this paper leans on

50 extracted references · 25 canonical work pages · cited by 2 Pith papers

  1. [1]

    X. An, X. Zhu, Y . Gao, Y . Xiao, Y . Zhao, Z. Feng, L. Wu, B. Qin, M. Zhang, D. Zhang, et al. Partial fc: Training 10 million identities on a single machine. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1445–1449, 2021

    work page 2021

  2. [2]

    S. A. Aytes, J. Baek, and S. J. Hwang. Sketch-of-thought: Efficient llm reasoning with adaptive cognitive-inspired sketching, 2025

    work page 2025

  3. [3]

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025

    arXiv 2025

  4. [4]

    Black, M

    K. Black, M. Janner, Y . Du, I. Kostrikov, and S. Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023

    arXiv 2023

  5. [5]

    Brooks, B

    T. Brooks, B. Peebles, C. Holmes, W. DePue, Y . Guo, L. Jing, D. Schnurr, J. Taylor, T. Luhman, E. Luhman, C. Ng, R. Wang, and A. Ramesh. Video generation models as world simulators. Preprint, 2024

    work page 2024

  6. [6]

    J. Chen, F. Long, J. An, Z. Qiu, T. Yao, J. Luo, and T. Mei. Ouroboros-diffusion: Exploring con- sistent content generation in tuning-free long video diffusion. arXiv preprint arXiv:2501.09019, 2025

    arXiv 2025

  7. [7]

    N. Chen, Z. Hu, Q. Zou, J. Wu, Q. Wang, B. Hooi, and B. He. Judgelrm: Large reasoning models as a judge, 2025

    work page 2025

  8. [8]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025

    DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025

    2025

Show all 50 references
  1. [9]

    H. Deng, T. Pan, H. Diao, Z. Luo, Y . Cui, H. Lu, S. Shan, Y . Qi, and X. Wang. Autoregressive video generation without vector quantization. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  2. [10]

    J. Deng, J. Guo, X. Niannan, and S. Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019

    2019

  3. [11]

    Esser, S

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024

    2024

  4. [12]

    Y . Fan, O. Watkins, Y . Du, H. Liu, M. Ryu, C. Boutilier, P. Abbeel, M. Ghavamzadeh, K. Lee, and K. Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36:79858–79885, 2023

    2023

  5. [13]

    Z. Fei, M. Fan, C. Yu, D. Li, and J. Huang. Scaling diffusion transformers to 16 billion parameters. arXiv preprint arXiv:2407.11633, 2024

    2024 arXiv

  6. [14]

    X. He, D. Jiang, G. Zhang, M. Ku, A. Soni, S. Siu, H. Chen, A. Chandra, Z. Jiang, A. Arulraj, et al. Videoscore: Building automatic metrics to simulate fine-grained human feedback for video generation. arXiv preprint arXiv:2406.15252, 2024

    2024 arXiv

  7. [15]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  8. [16]

    Huang, B

    W. Huang, B. Jia, Z. Zhai, S. Cao, Z. Ye, F. Zhao, Z. Xu, Y . Hu, and S. Lin. Vision-r1: Incentiviz- ing reasoning capability in multimodal large language models.arXiv preprint arXiv:2503.06749, 2025

    2025 arXiv

  9. [17]

    J. Kim, J. Kang, J. Choi, and B. Han. Fifo-diffusion: Generating infinite videos from text without training. arXiv preprint arXiv:2405.11473, 2024

    2024 arXiv

  10. [18]

    Kirstain, A

    Y . Kirstain, A. Polyak, U. Singer, S. Matiana, J. Penna, and O. Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:36652–36663, 2023. 10

    2023

  11. [19]

    Lab and T

    P.-Y . Lab and T. A. etc. Open-sora-plan, Apr. 2024

    2024

  12. [20]

    J. Lei, X. Hu, Y . Wang, and D. Liu. Pyramidflow: High-resolution defect contrastive localization using pyramid normalizing flow. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14143–14152, 2023

    2023

  13. [21]

    Z. Lin, D. Pathak, B. Li, J. Li, X. Xia, G. Neubig, P. Zhang, and D. Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2024

    2024

  14. [22]

    Lipman, R

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022

    2022 arXiv

  15. [23]

    J. Liu, G. Liu, J. Liang, Z. Yuan, X. Liu, M. Zheng, X. Wu, Q. Wang, W. Qin, M. Xia, et al. Improving video generation with human feedback. arXiv preprint arXiv:2501.13918, 2025

    2025 arXiv

  16. [24]

    Z. Mou, B. Xia, Z. Huang, W. Yang, and J. Jia. Gradeo: Towards human-like evaluation for text-to-video generation via multi-step reasoning. arXiv preprint arXiv:2503.02341, 2025

    2025

  17. [25]

    Muennighoff, Z

    N. Muennighoff, Z. Yang, W. Shi, X. L. Li, L. Fei-Fei, H. Hajishirzi, L. Zettlemoyer, P. Liang, E. Candès, and T. Hashimoto. s1: Simple test-time scaling, 2025

    2025

  18. [26]

    Paraperas Papantoniou, A

    F. Paraperas Papantoniou, A. Lattas, S. Moschoglou, J. Deng, B. Kainz, and S. Zafeiriou. Arc2face: A foundation model for id-consistent human faces. In Proceedings of the European Conference on Computer Vision (ECCV), 2024

    2024

  19. [27]

    Peebles and S

    W. Peebles and S. Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195–4205, 2023

    2023

  20. [28]

    H. Qiu, M. Xia, Y . Zhang, Y . He, X. Wang, Y . Shan, and Z. Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. arXiv preprint arXiv:2310.15169, 2023

    2023 arXiv

  21. [29]

    Rafailov, A

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728–53741, 2023

    2023

  22. [30]

    Ragain and J

    S. Ragain and J. Ugander. Choosing to rank. arXiv preprint arXiv:1809.05139, 2018

    2018 arXiv

  23. [31]

    Schuhmann, R

    C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278– 25294, 2022

    2022

  24. [32]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  25. [33]

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . Li, Y . Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  26. [34]

    J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  27. [35]

    G. Team. Mochi 1. https://github.com/genmoai/models, 2024

    2024

  28. [36]

    G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023

    2023 arXiv

  29. [37]

    H. L. Turner, J. van Etten, D. Firth, and I. Kosmidis. Modelling rankings in R: The PlackettLuce package. Computational Statistics, 35:1027–1057, 2020

    2020

  30. [38]

    Wallace, M

    B. Wallace, M. Dang, R. Rafailov, L. Zhou, A. Lou, S. Purushwalkam, S. Ermon, C. Xiong, S. Joty, and N. Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228–8238, 2024

    2024

  31. [39]

    T. Wan. Wan: Open and advanced large-scale video generative models, 2025

    2025

  32. [40]

    Y . Wang, Y . He, Y . Li, K. Li, J. Yu, X. Ma, X. Chen, Y . Wang, P. Luo, Z. Liu, Y . Wang, L. Wang, and Y . Qiao. Internvid: A large-scale video-text dataset for multimodal understanding and generation. arXiv preprint arXiv:2307.06942, 2023. 11

    2023 arXiv

  33. [41]

    Y . Wang, S. Wu, Y . Zhang, S. Yan, Z. Liu, J. Luo, and H. Fei. Multimodal chain-of-thought reasoning: A comprehensive survey, 2025

    2025

  34. [42]

    X. Wu, Y . Hao, K. Sun, Y . Chen, F. Zhu, R. Zhao, and H. Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023

    2023 arXiv

  35. [44]

    X. Wu, S. Huang, G. Wang, J. Xiong, and F. Wei. Boosting text-to-video generative model with mllms feedback. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  36. [45]

    X. Wu, K. Sun, F. Zhu, R. Zhao, and H. Li. Human preference score: Better aligning text- to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096–2105, 2023

    2023

  37. [46]

    J. Xu, X. Liu, Y . Wu, Y . Tong, Q. Li, M. Ding, J. Tang, and Y . Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation.Advances in Neural Information Processing Systems, 36, 2024

    2024

  38. [47]

    Z. Yang, J. Teng, W. Zheng, M. Ding, S. Huang, J. Xu, Y . Yang, W. Hong, X. Zhang, G. Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024

    2024 arXiv

  39. [48]

    T. Yin, Q. Zhang, R. Zhang, W. T. Freeman, F. Durand, E. Shechtman, and X. Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024

    2024

  40. [49]

    H. Yuan, S. Zhang, X. Wang, Y . Wei, T. Feng, Y . Pan, Y . Zhang, Z. Liu, S. Albanie, and D. Ni. Instructvideo: Instructing video diffusion models with human feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6463–6474, 2024

    2024

  41. [50]

    Zaheer, G

    M. Zaheer, G. Guruganesh, K. A. Dubey, J. Ainslie, C. Alberti, S. Ontanon, P. Pham, A. Ravula, Q. Wang, L. Yang, et al. Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33:17283–17297, 2020

    2020

  42. [51]

    0 " : " A man play with a dog at the train platform

    Z. Zhu, G. Huang, J. Deng, Y . Ye, J. Huang, X. Chen, J. Zhu, T. Yang, J. Lu, D. Du, and J. Zhou. Webface260m: A benchmark unveiling the power of million-scale deep face recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p...

    2021

Pith tools

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