Pith. sign in

REVIEW 3 major objections 6 minor 52 references

Harness Local Rewards for Global Benefits: Effective Text-to-Video Generation Alignment with Patch-level Reward Models

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper claims that post-training text-to-video diffusion models with patch-level rewards, combined with a margin-weighted granular DPO loss, reduces localized artifacts and raises overall generation quality.

desk verdict Plausible patch-level DPO for T2V alignment, but thin evaluation and a weakly validated patch reward model keep the central claim from being secure. read the letter →

arxiv 2502.06812 v2 pith:MLVB2TIW submitted 2025-02-04 cs.LG cs.GR

Shuting Wang , Haihong Tang , Zhicheng Dou , Chenyan Xiong This is my paper · ORCID
classification cs.LGcs.GR
keywords text-to-videogenerationdiffusionmodelsdirectpreferenceoptimizationpatchrewardmodeldistillationvideoalignmentlocalizedartifactspost-training
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

Text-to-video diffusion models can look good globally while still producing localized errors—extra limbs, missing objects, physically wrong details—and this paper argues that those errors are best corrected by training on patch-level rewards rather than only on whole-video scores. The authors construct a patch reward model by distilling GPT-4o patch-quality labels into a VideoScore-based video reward model, which keeps the local and global reward scales consistent. They then introduce Gran-DPO, a diffusion DPO variant whose loss sums a video-level preference term with per-patch preference terms, with pair weights derived from reward margins. In experiments across three base video generators, the combined method raises VBench and VideoScore averages, and the authors report that low-scoring patches improve after training. If true, the work establishes a practical way to add fine-grained, human-aligned local feedback to video generation post-training without collecting large human preference datasets.

What carries the argument

The load-bearing mechanism is Gran-DPO, an adapted Diffusion-DPO objective defined as a weighted sum of a whole-video DPO loss and a DPO loss for each spatial patch. For a patch pair, a ternary function $F(p^w_{ij}, p^l_{ij})$ sets the local preference direction to $+1$, $0$, or $-1$ so that local winner and loser pairs do not need to agree with the global video preference; pair weights $\omega(\cdot,\cdot)$ put more emphasis on pairs with larger reward gaps. The patch reward model that feeds this objective is created by regression fine-tuning VideoScore on GPT-4o-labeled patches from a 3x3 grid, which keeps the patch and video reward distributions on comparable scales. This combination lets the denoiser receive fine-grained spatial feedback at training time without discarding the global video-quality signal.

What would settle it

Recruit human raters to score a larger set of video patches beyond the paper's 360 and compute per-dimension Spearman correlations between the patch reward model and humans; then retrain HALO using only cases where GPT-4o and humans disagree. If the patch reward model's human correlation is close to zero or negative, or if human-rated local defects do not decrease after training, the central claim would be refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that global video rewards miss spatially localized generation defects, and that adding a patch reward model plus a granular DPO objective lets a text-to-video model reduce those defects while keeping overall quality. The patch reward model is made by taking VideoScore and continuing its training on 4,428 GPT-4o-labeled 3x3 spatial patches, with the whole video supplied as context so patches keep their semantics. The Gran-DPO loss then combines the ordinary video Diffusion-DPO loss with patch-level DPO losses: each patch pair uses a sign function to choose its own winner and loser, and both video and patch pairs are weighted by their reward margins. The paper reports that the full model improves average VBench and VideoScore scores over three different base models, that ablating either loss term hurts, and that the distribution of patch scores shifts upward after training. The intended conclusion is that local and global reward feedback are complementary, and jointly optimizing both is what produces the quality gain.

Load-bearing premise

The whole patch-level training signal depends on GPT-4o's patch quality scores being a good proxy for human perception of local defects; if that teacher is biased, the patch DPO loss is biased, and the claimed defect reduction could be an artifact of fitting the teacher rather than a real quality gain.

Editorial extensions

If this is right

  • Because HALO improves models that were already post-trained with video rewards, patch-level rewards can be added on top of existing alignment pipelines as a plug-in stage.
  • Ablating either the patch DPO loss or the video DPO loss lowers average scores, so local and global rewards are complementary rather than redundant.
  • Swapping in reward models with different scales (PickScore or HPSv2) hurts performance, indicating that consistency between patch and video reward distributions is part of what makes the joint training stable.
  • The rightward shift in patch reward distributions after fine-tuning indicates that training moves low-quality patches up rather than merely inflating global scores.

Reading between the lines

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

  • The paper deliberately avoids splitting patches along time because videos are short; with longer videos, the same Gran-DPO recipe could be run over temporal windows, and the preference-consistency analysis could be repeated per window to test whether temporal local rewards add signal beyond spatial ones.
  • The dependence on GPT-4o as teacher means the method's ceiling is set by the teacher's perceptual agreement with humans; a larger human validation set than the 360 patches, with confidence intervals on the Spearman correlations, would be the direct test of that ceiling.
  • The margin-weighted combination of two preference sources suggests a general principle for multi-granularity DPO: instead of collapsing all rewards into one scalar, let each granularity vote on its own scale and weight disagreements by margin. That recipe could transfer to other generative domains where global metrics are smooth but local structure matters.
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

3 major / 6 minor

Summary. The paper proposes HALO, a post-training framework for text-to-video diffusion models that augments global video-reward alignment with patch-level reward signals. A patch reward model is created by fine-tuning VideoScore-v1.1 on GPT-4o-labeled patch scores, and a Gran-DPO objective combines patch-level and video-level diffusion DPO losses. The method is evaluated on T2V-Turbo-v1/v2 and CogVideoX-2B using VBench (5 of 16 dimensions) and VideoScore, with ablations and reward-distribution analyses.

Significance. If the reported effect is real, HALO provides a useful mechanism for injecting localized quality feedback into video diffusion alignment, and the Gran-DPO formulation is a plausible extension of Diffusion-DPO. The paper also offers an informative ablation study isolating the patch DPO, video DPO, pair weighting, and choice of video reward model. However, the central claim depends on the patch reward model tracking human perception of local defects, and the evidence for that is currently weak (Spearman 0.368, n=360, no confidence intervals). The evaluation also has a partial circularity because VideoScore is both the initialization of the patch reward model and one of the two evaluation metrics. These gaps are fixable and do not invalidate the method's potential, but they currently prevent the paper's 'substantially outperforms' claim from being accepted as stated.

major comments (3)
  1. [§5.3, Table 3] The only evidence that the patch reward model tracks human perception is the Human–Patch Reward Model Spearman correlation of 0.3684 computed on 360 patches, with no confidence interval or significance test. The patch reward model is trained entirely on GPT-4o labels (Section 3.2, Eq. 6), and its correlation with GPT-4o (0.6062) is much higher than its correlation with humans (0.3684), which is also lower than the Human–GPT-4o correlation (0.4023). Since Gran-DPO optimizes the patch reward model's scores (Eqs. 8–9), the observed improvements in patch reward values (Figure 5b) and in evaluation metrics may reflect optimization toward GPT-4o's biases rather than a genuine reduction of human-perceived localized defects. The central claim that local feedback provides useful fine-grained training signal therefore requires stronger human validation, such as a larger annotation sample, confidence intervals, per-dimension correlations, and ideally a human preference study comparing HALO outputs against baselines.
  2. [§3.2, §4, Table 1] The evaluation is partially circular. The patch reward model is obtained by fine-tuning VideoScore-v1.1 (Section 3.2), and VideoScore-v1.1 is one of the two headline evaluation metrics (Section 4). Consequently, gains in VideoScore after HALO training can partly reflect optimization toward the same evaluator family used to create the patch reward model. The second evaluation method, VBench, is reported only on 5 of 16 dimensions, selected because the authors 'focus on and have optimized space for them' (Section 4). This subset choice is disclosed but is not reflected in the abstract's claim of 'substantially outperforms the baselines across two evaluation methods.' Please report full VBench results or a principled selection procedure, and either use an independent video reward model for evaluation or explicitly quantify the circularity risk (e.g., by showing that HALO also improves on metrics not related to VideoScore).
  3. [§5.1, Table 1] No error bars, confidence intervals, or significance tests are reported for any of the main comparisons. The improvement of HALO over T2V-Turbo-v1 is tiny on VideoScore (2.5503 to 2.5526, a difference of 0.0023) and modest on the VBench subset (63.02 to 63.57), and it is not plausible that these differences are meaningful without repeated evaluation runs or a significance test. The statement in Section 5.1 that HALO 'significantly enhances the performance' is therefore unsupported for at least one of the three base models. Please provide confidence intervals or significance tests for the primary comparisons, or at least report per-prompt variance.
minor comments (6)
  1. [§3.3, Eq. (9)] The video-pair weight omega(vw, vl) is normalized by mP (the median patch reward margin) while the text defines mP as the median of pairwise patch reward margins and mV as the median of pairwise video reward margins; this looks like a typo, and the video weight should presumably be normalized by mV.
  2. [§5.3, Appendix C] The human annotation procedure is described only in the appendix and does not state the number of annotators or their agreement; adding this information would help interpret the Spearman correlations in Table 3.
  3. [§3.2] Because the patch reward model does not segment the time axis, 'local feedback' covers only spatial patches; the text acknowledges this, but it should be listed as a limitation in the conclusion since temporal defects (e.g., flicker, motion artifacts) are common in AI-generated videos.
  4. [Appendix C.1] There are several typos in the GPT-4o instruction, including 'patch-leval' and 'Requirments', and the apostrophe in 'video patch´s' is typeset inconsistently; please proofread.
  5. [References] The reference list contains duplicate entries for Wallace et al. 2024a/b (the same CVPR paper cited twice) and for Ouyang et al. 2022a/b; these should be consolidated.
  6. [§4] In the implementation details, 'the epoch is 10' and 'the LoRA rank was set by 64' are ungrammatical; these should be reworded for clarity.

Circularity Check

2 steps flagged · score 5.0 of 10

The local-defect improvement is measured with the same patch reward model that Gran-DPO is trained to maximize, making the defect-alleviation evidence partially self-referential; VideoScore also serves as both the training reward and one of the evaluation metrics.

  1. fitted input called prediction [Section 5.4, 'Distribution of patch reward variances' / Figure 5, relative to Gran-DPO loss in Eq. (9) and patch RM in Section 3.2]
    "After fine-tuning the VGM by our method, the distributions of all levels shift to the right. These results validate that our proposed method not only could improve the overall quality of generated videos but also could enhance the localized and fine-grained quality of AI-generated videos."

    The patch-reward distributions shown in Figure 5 are outputs of the very patch reward model that the Gran-DPO objective is designed to maximize. Equation (8) defines the patch DPO loss with sign F(p_w,p_l) so that the generator is pushed toward higher patch-RM scores, and Eq. (9) weights pairs by reward margins. Reporting the post-training increase in those scores as evidence of 'alleviating' patch defects therefore measures success on the training objective, not an independent outcome. The only external link between the patch RM's scores and human-perceived defect quality is Table 3's Human-Patch Reward Model Spearman correlation of 0.3684, computed on 360 patches with no confidence interval or significance test.

  2. other [Section 4, 'Evaluation prompts and methods', vs. Section 3.2 and Eqs. (5), (7), (9)]
    "We use VideoScore (He et al., 2024) as our video reward model ... we use VBench and VideoScore (He et al., 2024) as our evaluation methods ... VideoScore-v1.1 serves as our evaluator."

    The same VideoScore-v1.1 model is used both to compute the video rewards that determine the winner/loser pairs and video DPO loss in Eqs. (5) and (7), and as one of the two final evaluation metrics in Table 1. Consequently, the VideoScore column partly measures how well HALO fits its own training-time video reward signal. This would be fully circular only if VideoScore were the sole metric, but the VBench results and the ablation comparisons provide independent content. The overlap nonetheless introduces a partial self-referential bias into the reported global-quality gains, while the patch-level gains in Figure 5 are more directly circular because they use the exact patch RM optimized in Eq. (9).

full rationale

The paper's central algorithmic claim—that patch-level rewards provide useful signal beyond global video rewards—does have independent content. The ablation results in Table 2 show that removing the patch DPO loss degrades both VBench and VideoScore, which is a non-circular comparison because the patch RM was trained on GPT-4o labels rather than on VBench or VideoScore labels. There is also an external, though weak, human-alignment check for the patch RM (Spearman 0.3684 on 360 patches). No load-bearing self-citation or imported uniqueness theorem appears: Diffusion-DPO is cited as a standard external method, and the other self-citations are not doing justificatory work. However, the paper's headline evidence that HALO 'effectively alleviates' localized patch defects rests on Figure 5, which plots scores from the same patch reward model that the Gran-DPO objective is trained to maximize; that increase is close to a training-loss artifact unless the patch RM is independently shown to track human perception, and the provided human correlation is too weak and underpowered for that role. In addition, VideoScore serves both as the video reward used to construct DPO pairs and as one of the two evaluation metrics, adding a partial self-referential bias to the global-score evidence. These issues make the local-defect and part of the global-quality evidence partially circular, but they do not reduce the entire contribution to a tautology, so a moderate score of 5 is appropriate. Score 5.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several assumptions that are not independently verified: the reliability of GPT-4o as a patch-quality teacher, the meaningfulness of spatial patch decomposition of latents, and the validity of VideoScore as a global reward. The training procedure also relies on several hand-chosen hyperparameters and a data-selection rule based on reward margins. No code, data, or checkpoints are released, so the empirical results cannot be independently reproduced from the paper.

free parameters (6)
  • Patch grid size hn, wn = 3 x 3
    Chosen by hand. It defines the granularity of local feedback and the number of patch reward scores per video.
  • DPO temperature beta
    Regularization hyperparameter in the DPO loss (Eqs. 4, 8, 9). The value is not reported, so the exact training objective is not reproducible.
  • Prompt filtering threshold tau = 0.85
    Ad hoc threshold in Eq. 10 to remove training prompts that are too similar to evaluation or earlier generated prompts.
  • Pair selection medians mV and mP = Statistical medians of pairwise reward margins
    Data builder in Section 3.3 keeps only pairs whose video or patch reward margin exceeds the median. This selection may inflate apparent DPO gains.
  • Distillation training set size = 4,428 labeled patches from 492 videos
    Chosen to cap GPT-4o API cost; the patch reward model is trained on this small set.
  • LoRA rank, learning rate, steps = 64, 1e-4, 8k
    Fine-tuning hyperparameters for the base VGMs; no sensitivity analysis is provided.
assumptions (5)
  • domain assumption VideoScore provides a valid global quality measure for generated videos.
    Used both as the video reward in Gran-DPO and as one of the two evaluation metrics (VideoScore-v1.1). If its scores are biased, training and evaluation are biased together.
  • domain assumption Spatial patches of the latent representation are a meaningful unit for local defect detection and optimization.
    Gran-DPO splits latents into a 3x3 grid and applies patch-level DPO (Eq. 8). This assumes defects manifest in local spatial regions and that patch-level noise prediction error is a valid proxy.
  • domain assumption GPT-4o patch quality labels are reliable enough to supervise a reward model.
    Section 3.2 uses GPT-4o as the sole teacher. Section 5.3 supports this with a weak Spearman correlation of 0.368 between the resulting patch reward model and human annotations on 360 patches.
  • standard math Diffusion-DPO is a valid post-training objective for VGMs.
    Adopted from Wallace et al. 2024; Gran-DPO is built on this loss, so the correctness of the base objective is assumed.
  • domain assumption The five chosen VBench dimensions are sufficient to measure overall video generation quality.
    Only image quality, multiple objects, human action, spatial relationship, and scene are reported, selected because the authors say they focus on and have optimized space for them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Harness Local Rewards for Global Benefits: Effective Text-to-Video Generation Alignment with Patch-level Reward Models." pith.science (2026). https://pith.science/paper/MLVB2TIW

@misc{pith2026250206812,
  author       = {Pith},
  title        = {Pith review of: Harness Local Rewards for Global Benefits: Effective Text-to-Video Generation Alignment with Patch-level Reward Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MLVB2TIW}},
  note         = {Machine review of arXiv:2502.06812}
}
read the original abstract

The emergence of diffusion models (DMs) has significantly improved the quality of text-to-video generation models (VGMs). However, current VGM optimization primarily emphasizes the global quality of videos, overlooking localized errors, which leads to suboptimal generation capabilities. To address this issue, we propose a post-training strategy for VGMs, HALO, which explicitly incorporates local feedback from a patch reward model, providing detailed and comprehensive training signals with the video reward model for advanced VGM optimization. To develop an effective patch reward model, we distill GPT-4o to continuously train our video reward model, which enhances training efficiency and ensures consistency between video and patch reward distributions. Furthermore, to harmoniously integrate patch rewards into VGM optimization, we introduce a granular DPO (Gran-DPO) algorithm for DMs, allowing collaborative use of both patch and video rewards during the optimization process. Experimental results indicate that our patch reward model aligns well with human annotations and HALO substantially outperforms the baselines across two evaluation methods. Further experiments quantitatively prove the existence of patch defects, and our proposed method could effectively alleviate this issue.

Figures

Figures reproduced from arXiv: 2502.06812 by the authors.

Figure 1
Figure 1. A generated video with local flaws (red boxed). plications, such as video editing and artistic creation (Pic, 2025; run, 2025; syn, 2025). Typically, the pre-training of VGMs relies on extensive text￾to-video datasets (Bain et al., 2021), leading to strong foun￾dational generation abilities. However, the varied quality of these datasets still limits the potential of VGMs. Inspired by the alignment of large language … view at source ↗
Figure 2
Figure 2. The framework of our proposed method, HALO. Distill GPT-4o to build the patch RM. With the labeled data from GPT-4o, we fine-tune the video reward model by regression loss following VideoScore, as expressed below, Lreg(v) = 1 hnwn|DV S| Xhn i=1 Xwn j=1 X d∈DV S ∥p d ij − g d ij∥ 2 2. (6) p d ij denotes the reward score predicted by the patch reward model for the video patch in the index of (i, j), and g d ij is the … view at source ↗
Figure 3
Figure 3. The visualized comparison between our proposed model HALO and baselines. 0.00 0.25 0.50 0.75 1.00 (a) Consistency between video and patch rewards VQ TC DD TVA FC 54.0% 51.2% 43.2% 54.0% 51.6% 9.8% 5.4% 9.4% 9.3% 5.4% 36.2% 43.4% 47.5% 36.6% 43.1% 0k 0.2k 0.4k 0.8k 1.2k (c) Performance trends with the training pair amount 2.47 2.48 2.49 2.50 2.51 VideoScore VideoScore 64 66 68 VBench Vbench 0 2500 5000 7500 (b) Rewar… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Further analyses about our reward models and training process. lowing VideoScore (He et al., 2024), we define five evalua￾tion dimensions for video patches with scores ranging from 1 to 4. The annotated videos are sampled from VideoScore￾collected AI-generated videos a…
Figure 5
Figure 5. Figure 5: Two types of patch reward distributions. if one of the rewards has no preferences between the two videos (reward values are identical), we label the pair as “None”, if the preferences differ, they are labeled “Inverse”. We conduct this analysis on all our training prom…
Figure 6
Figure 6. Figure 6: Cases to show the evaluated patch rewards. A. Implementation Details All our models are trained on 4 NVIDIA A100 GPUs. For the training instance generation, the filtering threshold, tau, is set as 0.85. 878 text prompts were generated by prompting GPT-3.5-Turbo. We spl…
Figure 7
Figure 7. Figure 7: Instructions for human annotators to label video patch scores. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Instruction for GPT-4o to Generate Evaluation Scores for the Input Video Patch. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Instruction for GPT-4o to generate evaluation scores for the input video patch. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 25 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    URL https://pictory.ai/blog

    The pictory blog, 2025. URL https://pictory.ai/blog. Accessed: 2025-01-27

    work page 2025

  3. [3]

    URL https://runwayml.com/

    Advancing creativity with artificial intelligence., 2025. URL https://runwayml.com/. Accessed: 2025-01-27

    work page 2025

  4. [4]

    URL https://www.synthesia.io/tools/ai-video-editor

    Ai video editor., 2025. URL https://www.synthesia.io/tools/ai-video-editor. Accessed: 2025-01-27

    work page 2025

  5. [5]

    Frozen in time: A joint video and image encoder for end-to-end retrieval

    Bain, M., Nagrani, A., Varol, G., and Zisserman, A. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV , pp.\ 1708--1718. IEEE , 2021

    work page 2021

  6. [6]

    Training diffusion models with reinforcement learning

    Black, K., Janner, M., Du, Y., Kostrikov, I., and Levine, S. Training diffusion models with reinforcement learning. In ICLR . OpenReview.net, 2024

    work page 2024

  7. [7]

    W., Fidler, S., and Kreis, K

    Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S. W., Fidler, S., and Kreis, K. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR , pp.\ 22563--22575. IEEE , 2023

    work page 2023

  8. [8]

    Videocrafter1: Open diffusion models for high-quality video generation

    Chen, H., Xia, M., He, Y., Zhang, Y., Cun, X., Yang, S., Xing, J., Liu, Y., Chen, Q., Wang, X., Weng, C., and Shan, Y. Videocrafter1: Open diffusion models for high-quality video generation. CoRR, abs/2310.19512, 2023

    arXiv 2023

Show all 52 references
  1. [9]

    Videocrafter2: Overcoming data limitations for high-quality video diffusion models

    Chen, H., Zhang, Y., Cun, X., Xia, M., Wang, X., Weng, C., and Shan, Y. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In CVPR , pp.\ 7310--7320. IEEE , 2024

    2024

  2. [10]

    S., Brox, T., and Ronneberger, O

    C i c ek, \" O ., Abdulkadir, A., Lienkamp, S. S., Brox, T., and Ronneberger, O. 3d u-net: Learning dense volumetric segmentation from sparse annotation. In MICCAI (2) , volume 9901 of Lecture Notes in Computer Science, pp.\ 424--432, 2016

    2016

  3. [11]

    Clark, K., Vicol, P., Swersky, K., and Fleet, D. J. Directly fine-tuning diffusion models on differentiable rewards. In ICLR . OpenReview.net, 2024

    2024

  4. [12]

    DPOK: reinforcement learning for fine-tuning text-to-image diffusion models

    Fan, Y., Watkins, O., Du, Y., Liu, H., Ryu, M., Boutilier, C., Abbeel, P., Ghavamzadeh, M., Lee, K., and Lee, K. DPOK: reinforcement learning for fine-tuning text-to-image diffusion models. CoRR, abs/2305.16381, 2023

    2023 arXiv

  5. [13]

    D., Ni, Y., Lyu, B., Narsupalli, Y., Fan, R., Lyu, Z., Lin, B

    He, X., Jiang, D., Zhang, G., Ku, M., Soni, A., Siu, S., Chen, H., Chandra, A., Jiang, Z., Arulraj, A., Wang, K., Do, Q. D., Ni, Y., Lyu, B., Narsupalli, Y., Fan, R., Lyu, Z., Lin, B. Y., and Chen, W. Videoscore: Building automatic metrics to simulate fine-grained human feedba...

    2024

  6. [14]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. In NeurIPS, 2020

    2020

  7. [15]

    Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., and Fleet, D. J. Video diffusion models. arXiv:2204.03458, 2022

    2022 arXiv

  8. [16]

    Cogvideo: Large-scale pretraining for text-to-video generation via transformers

    Hong, W., Ding, M., Zheng, W., Liu, X., and Tang, J. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In ICLR . OpenReview.net, 2023

    2023

  9. [17]

    J., Shen, Y., Wallis, P., Allen - Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W

    Hu, E. J., Shen, Y., Wallis, P., Allen - Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W. Lora: Low-rank adaptation of large language models. In ICLR . OpenReview.net, 2022

    2022

  10. [18]

    Vbench: Comprehensive benchmark suite for video generative models

    Huang, Z., He, Y., Yu, J., Zhang, F., Si, C., Jiang, Y., Zhang, Y., Wu, T., Jin, Q., Chanpaisit, N., Wang, Y., Chen, X., Wang, L., Lin, D., Qiao, Y., and Liu, Z. Vbench: Comprehensive benchmark suite for video generative models. In CVPR , pp.\ 21807--21818. IEEE , 2024

    2024

  11. [19]

    MANTIS: interleaved multi-image instruction tuning

    Jiang, D., He, X., Zeng, H., Wei, C., Ku, M., Liu, Q., and Chen, W. MANTIS: interleaved multi-image instruction tuning. CoRR, abs/2405.01483, 2024

    2024 arXiv

  12. [20]

    Text2video-zero: Text-to-image diffusion models are zero-shot video generators

    Khachatryan, L., Movsisyan, A., Tadevosyan, V., Henschel, R., Wang, Z., Navasardyan, S., and Shi, H. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In ICCV , pp.\ 15908--15918. IEEE , 2023

    2023

  13. [21]

    Pick-a-pic: An open dataset of user preferences for text-to-image generation

    Kirstain, Y., Polyak, A., Singer, U., Matiana, S., Penna, J., and Levy, O. Pick-a-pic: An open dataset of user preferences for text-to-image generation. In NeurIPS, 2023

    2023

  14. [22]

    Step-dpo: Step-wise preference optimization for long-chain reasoning of llms

    Lai, X., Tian, Z., Chen, Y., Yang, S., Peng, X., and Jia, J. Step-dpo: Step-wise preference optimization for long-chain reasoning of llms. CoRR, abs/2406.18629, 2024

    2024 arXiv

  15. [23]

    Lee, K., Liu, H., Ryu, M., Watkins, O., Du, Y., Boutilier, C., Abbeel, P., Ghavamzadeh, M., and Gu, S. S. Aligning text-to-image models using human feedback. CoRR, abs/2302.12192, 2023

    2023 arXiv

  16. [24]

    Li, J., Feng, W., Fu, T., Wang, X., Basu, S., Chen, W., and Wang, W. Y. T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback. CoRR, abs/2405.18750, 2024 a

    2024 arXiv

  17. [25]

    Li, J., Long, Q., Zheng, J., Gao, X., Piramuthu, R., Chen, W., and Wang, W. Y. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. CoRR, abs/2410.05677, 2024 b

    2024

  18. [26]

    Reinforcement learning with human feedback: Learning dynamic choices via pessimism

    Li, Z., Yang, Z., and Wang, M. Reinforcement learning with human feedback: Learning dynamic choices via pessimism. CoRR, abs/2305.18438, 2023

    2023 arXiv

  19. [27]

    Training diffusion models towards diverse image generation with reinforcement learning

    Miao, Z., Wang, J., Wang, Z., Yang, Z., Wang, L., Qiu, Q., and Liu, Z. Training diffusion models towards diverse image generation with reinforcement learning. In CVPR , pp.\ 10844--10853. IEEE , 2024

    2024

  20. [28]

    Openvid-1m: A large-scale high-quality dataset for text-to-video generation

    Nan, K., Xie, R., Zhou, P., Fan, T., Yang, Z., Chen, Z., Li, X., Yang, J., and Tai, Y. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. CoRR, abs/2407.02371, 2024

    2024 arXiv

  21. [29]

    GPT-4 technical report

    OpenAI. GPT-4 technical report. CoRR, abs/2303.08774, 2023

    2023 arXiv

  22. [30]

    L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P. F., Leike, J., and Lowe, R. Training language models t...

    2022

  23. [31]

    L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P. F., Leike, J., and Lowe, R. Training language models t...

    2022

  24. [32]

    and Xie, S

    Peebles, W. and Xie, S. Scalable diffusion models with transformers. In ICCV , pp.\ 4172--4182. IEEE , 2023

    2023

  25. [33]

    SDXL: improving latent diffusion models for high-resolution image synthesis

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., M \" u ller, J., Penna, J., and Rombach, R. SDXL: improving latent diffusion models for high-resolution image synthesis. In ICLR . OpenReview.net, 2024

    2024

  26. [34]

    High-resolution image synthesis with latent diffusion models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In CVPR , pp.\ 10674--10685. IEEE , 2022

    2022

  27. [35]

    U-net: Convolutional networks for biomedical image segmentation

    Ronneberger, O., Fischer, P., and Brox, T. U-net: Convolutional networks for biomedical image segmentation. In MICCAI (3) , volume 9351 of Lecture Notes in Computer Science, pp.\ 234--241. Springer, 2015

    2015

  28. [36]

    LAION-5B: an open large-scale dataset for training next generation image-text models

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S., Crowson, K., Schmidt, L., Kaczmarczyk, R., and Jitsev, J. LAION-5B: an open large-scale dataset for training next gen...

    2022

  29. [37]

    A., Maheswaranathan, N., and Ganguli, S

    Sohl - Dickstein, J., Weiss, E. A., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML , volume 37 of JMLR Workshop and Conference Proceedings , pp.\ 2256--2265. JMLR.org, 2015

    2015

  30. [38]

    Denoising diffusion implicit models

    Song, J., Meng, C., and Ermon, S. Denoising diffusion implicit models. In ICLR . OpenReview.net, 2021

    2021

  31. [39]

    Consistency models

    Song, Y., Dhariwal, P., Chen, M., and Sutskever, I. Consistency models. In ICML , volume 202 of Proceedings of Machine Learning Research, pp.\ 32211--32252. PMLR , 2023

    2023

  32. [40]

    R., and Shah, M

    Soomro, K., Zamir, A. R., and Shah, M. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012

    2012 arXiv

  33. [41]

    Diffusion model alignment using direct preference optimization

    Wallace, B., Dang, M., Rafailov, R., Zhou, L., Lou, A., Purushwalkam, S., Ermon, S., Xiong, C., Joty, S., and Naik, N. Diffusion model alignment using direct preference optimization. In CVPR , pp.\ 8228--8238. IEEE , 2024 a

    2024

  34. [42]

    Diffusion model alignment using direct preference optimization

    Wallace, B., Dang, M., Rafailov, R., Zhou, L., Lou, A., Purushwalkam, S., Ermon, S., Xiong, C., Joty, S., and Naik, N. Diffusion model alignment using direct preference optimization. In CVPR , pp.\ 8228--8238. IEEE , 2024 b

    2024

  35. [43]

    Modelscope text-to-video technical report

    Wang, J., Yuan, H., Chen, D., Zhang, Y., Wang, X., and Zhang, S. Modelscope text-to-video technical report. CoRR, abs/2308.06571, 2023 a

    2023 arXiv

  36. [44]

    Richrag: Crafting rich responses for multi-faceted queries in retrieval-augmented generation

    Wang, S., Yu, X., Wang, M., Chen, W., Zhu, Y., and Dou, Z. Richrag: Crafting rich responses for multi-faceted queries in retrieval-augmented generation. CoRR, abs/2406.12566, 2024

    2024 arXiv

  37. [45]

    C., Dai, B., Lin, D., Qiao, Y., and Liu, Z

    Wang, Y., Chen, X., Ma, X., Zhou, S., Huang, Z., Wang, Y., Yang, C., He, Y., Yu, J., Yang, P., Guo, Y., Wu, T., Si, C., Jiang, Y., Chen, C., Loy, C. C., Dai, B., Lin, D., Qiao, Y., and Liu, Z. LAVIE: high-quality video generation with cascaded latent diffusion models. CoRR, ab...

    2023 arXiv

  38. [46]

    A., Khashabi, D., and Hajishirzi, H

    Wang, Y., Kordi, Y., Mishra, S., Liu, A., Smith, N. A., Khashabi, D., and Hajishirzi, H. Self-instruct: Aligning language models with self-generated instructions. In ACL (1) , pp.\ 13484--13508. Association for Computational Linguistics, 2023 c

    2023

  39. [47]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis

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

    2023 arXiv

  40. [48]

    Easyanimate: A high-performance long video generation method based on transformer architecture

    Xu, J., Zou, X., Huang, K., Chen, Y., Liu, B., Cheng, M., Shi, X., and Huang, J. Easyanimate: A high-performance long video generation method based on transformer architecture. CoRR, abs/2405.18991, 2024

    2024

  41. [49]

    Cogvideox: Text-to-video diffusion models with an expert transformer

    Yang, Z., Teng, J., Zheng, W., Ding, M., Huang, S., Xu, J., Yang, Y., Hong, W., Zhang, X., Feng, G., Yin, D., Gu, X., Zhang, Y., Wang, W., Cheng, Y., Liu, T., Xu, B., Dong, Y., and Tang, J. Cogvideox: Text-to-video diffusion models with an expert transformer. CoRR, abs/2408.06...

    2024 arXiv

  42. [50]

    S., Eom, S., Han, G., Nam, D

    Yoon, E., Yoon, H. S., Eom, S., Han, G., Nam, D. W., Jo, D., On, K., Hasegawa - Johnson, M., Kim, S., and Yoo, C. D. TLCR: token-level continuous reward for fine-grained reinforcement learning from human feedback. In ACL (Findings) , pp.\ 14969--14981. Association for Computat...

    2024

  43. [51]

    Instructvideo: Instructing video diffusion models with human feedback

    Yuan, H., Zhang, S., Wang, X., Wei, Y., Feng, T., Pan, Y., Zhang, Y., Liu, Z., Albanie, S., and Ni, D. Instructvideo: Instructing video diffusion models with human feedback. In CVPR , pp.\ 6463--6474. IEEE , 2024

    2024

  44. [52]

    J., Wu, J

    Zhang, D. J., Wu, J. Z., Liu, J., Zhao, R., Ran, L., Gu, Y., Gao, D., and Shou, M. Z. Show-1: Marrying pixel and latent diffusion models for text-to-video generation. CoRR, abs/2309.15818, 2023

    2023 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.