pith. machine review for the scientific record. sign in

arxiv: 2605.07503 · v1 · submitted 2026-05-08 · 💻 cs.CV

Recognition: 2 theorem links

· Lean Theorem

Diffusion-APO: Trajectory-Aware Direct Preference Alignment for Video Diffusion Transformers

Jingyuan Zhu , Biaolong Chen , Le Zhang , Aixi Zhang , Hao Jiang , Pipei Huang
Authors on Pith no claims yet

Pith reviewed 2026-05-11 01:55 UTC · model grok-4.3

classification 💻 cs.CV
keywords video diffusionpreference optimizationdirect preference alignmenttrajectory synchronizationRLHFdiffusion transformersmodel alignment
0
0 comments X

The pith

Diffusion-APO aligns video diffusion models by synchronizing training noise distributions with inference denoising trajectories to strengthen preference optimization.

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

The paper proposes Diffusion-APO to fix the mismatch between the noise used in training video diffusion transformers and the actual paths taken during inference. Existing methods like DPO and GRPO fall short because of biased rewards or poor timestep choices, so this approach matches the distributions directly. The result is stronger gradient signals that guide the model toward human-preferred outputs. A modular framework then wraps this core idea with online ranking, offline refinement, and drift correction to support different data and compute setups. Experiments show gains in visual quality and instruction following while keeping the model's original generative ability intact during speedups.

Core claim

Diffusion-APO is a trajectory-aware algorithm that resolves misalignment in direct preference alignment for video diffusion transformers by synchronizing training noise with inference-time denoising paths to maximize gradient signal efficacy. This is realized through a unified RLHF framework that combines online ranking, half-online anchoring, offline refinement, and distillation-aware drift correction, enabling scalable alignment without scalar-reward policy gradients.

What carries the argument

Trajectory-aware synchronization of training noise distributions with inference denoising paths, which aligns the noise schedules used in optimization to those encountered at generation time so that preference gradients become more effective.

If this is right

  • Diffusion-APO outperforms standard DPO and GRPO baselines in visual quality and instruction following on video tasks.
  • The method preserves generative fidelity when the model is accelerated after alignment.
  • The modular framework supports flexible multi-stage alignment under varying data and compute budgets without scalar rewards.
  • No reliance on complex reward models is needed for the preference optimization step.

Where Pith is reading between the lines

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

  • The same noise-trajectory matching idea could be tested on image or 3D diffusion models that share the training-inference gap.
  • Stronger gradients might shorten the number of preference-tuning steps required to reach a target quality level.
  • Combining Diffusion-APO with existing distillation or quantization techniques could further cut inference latency while retaining alignment gains.

Load-bearing premise

Matching the noise distributions seen during training to those encountered at inference will reliably improve gradient signals and produce better results than current DPO or GRPO baselines without adding new instabilities.

What would settle it

A side-by-side evaluation on the same video generation benchmarks in which models trained with Diffusion-APO receive equal or lower human preference scores for visual quality and instruction following than models trained with standard DPO or GRPO.

Figures

Figures reproduced from arXiv: 2605.07503 by Aixi Zhang, Biaolong Chen, Hao Jiang, Jingyuan Zhu, Le Zhang, Pipei Huang.

Figure 1
Figure 1. Figure 1: Overview of the Unified Preference Optimization Framework. The five-stage pipeline enables progressive, scalable preference alignment governed by the Diffusion-APO objective. Hybrid High–low Frequency Sliding Window. Directly optimizing isolated diffusion timesteps often destabilizes training due to strong temporal dependencies and severe gradient variance across noise scales, particularly the volatile gra… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative Comparison: Diffusion-DPO vs. Diffusion-APO. Under identical training configurations and datasets, Diffusion-APO produces videos with substantially fewer structural artifacts and superior visual fidelity. This performance gap underscores the efficacy of our trajectory￾aware, inference-aligned training strategy in optimizing the generative process. 4.2 Diffusion-APO Evaluation We curate our alig… view at source ↗
Figure 3
Figure 3. Figure 3: Human Preference Evaluation. Evaluators performed side-by-side, double-blind com￾parisons of videos generated under identical conditions, selecting the counterpart that demonstrates superior photorealism and reduced visual artifacts across our video quality benchmark (400 videos) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Training loss comparison. The blue line represents the baseline training without the sliding-window mechanism, exhibiting significant instability and periodic loss spikes due to gradient variance across noise scales. The red line, utilizing our hybrid high-low frequency sliding-window, demonstrates a significantly more stable convergence path, which directly translates to improved visual coherence and redu… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of generated results. The SFT baseline model (upper row) suffers from common generative failures in complex scenarios, including severe background distortion, motion-induced deformation, spurious artifacts (circled in red). In contrast, the model fine-tuned with the proposed fully aligned policy (bottom) effectively resolves these artifacts, consistently producing temporally coherent… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of generated results. The SFT baseline model (upper row) suffers from common generative failures in complex scenarios, including severe missing limbs and object deformation (circled in red). In contrast, the model fine-tuned with the proposed fully aligned policy (bottom row) effectively resolves these artifacts, consistently producing temporally coherent, anatomically accurate, and … view at source ↗
Figure 7
Figure 7. Figure 7: Visual demonstration of enhanced instruction following via Half-Online DPO. In I2V generation settings, the Half-Online DPO stage explicitly improves the model’s ability to accurately interpret and execute complex, action-driven text prompts, ensuring the generated visual dynamics perfectly align with user instructions. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Illustrative Example 1 of Post-Distillation Recovery and Enhancement through DPO. The top row shows the original teacher model’s high-fidelity output. The middle row presents the naively distilled student model, which exhibits visual degradation despite a 3× speedup. The bottom row displays our DPO-optimized model, successfully restoring and often surpassing the teacher’s generative fidelity, thus proving … view at source ↗
Figure 9
Figure 9. Figure 9: Illustrative Example 2 of Post-Distillation Recovery and Enhancement through DPO. The top row shows the original teacher model’s high-fidelity output. The middle row presents the naively distilled student model, which exhibits visual degradation despite a 3× speedup. The bottom row displays our DPO-optimized model, successfully restoring and often surpassing the teacher’s generative fidelity, thus proving … view at source ↗
Figure 10
Figure 10. Figure 10: Illustrative Example 3 of Post-Distillation Recovery and Enhancement through DPO. The top row shows the original teacher model’s high-fidelity output. The middle row presents the naively distilled student model, which exhibits visual degradation despite a 3× speedup. The bottom row displays our DPO-optimized model, successfully restoring and often surpassing the teacher’s generative fidelity, thus proving… view at source ↗
read the original abstract

Efficiently aligning large-scale video diffusion models with human intent requires a scalable and trajectory-aware pathway that bridges the inherent discrepancy between training noise distributions and practical inference trajectories. While existing paradigms such as Direct Preference Optimization (DPO) and Group Relative Policy Optimization (GRPO) attempt to address this, they are often hindered by either reliance on bias-prone, complex reward models or suboptimal timestep sampling. In this paper, we propose Diffusion-APO (Aligned Preference Optimization), a trajectory-aware algorithm that resolves this misalignment by synchronizing training noise with inference-time denoising paths to maximize gradient signal efficacy. To translate this algorithmic innovation into a practical solution, we introduce a unified and modular RLHF framework that integrates online ranking, half-online anchoring, offline refinement, and distillation-aware drift correction. This framework enables flexible, multi-stage preference alignment across diverse data and computational constraints without relying on scalar-reward-based policy gradients. Through extensive experiments, we demonstrate that Diffusion-APO consistently outperforms standard baselines in visual quality and instruction following, while effectively preserving generative fidelity during model acceleration, providing a robust, end-to-end pathway for scalable video diffusion alignment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces Diffusion-APO, a trajectory-aware direct preference optimization method for video diffusion transformers. It aims to bridge the gap between training noise distributions and inference trajectories by synchronizing them to enhance gradient signal efficacy. A modular RLHF framework is proposed, combining online ranking, half-online anchoring, offline refinement, and distillation-aware drift correction. The authors claim through extensive experiments that this approach outperforms baselines in visual quality and instruction following while preserving generative fidelity.

Significance. Should the noise synchronization and multi-stage alignment prove effective and stable, this could represent a meaningful step forward in aligning video diffusion models with human preferences in a scalable manner, reducing dependence on potentially biased reward models. It may influence future work on preference optimization for generative models in computer vision.

major comments (2)
  1. Abstract: The statement that Diffusion-APO 'consistently outperforms standard baselines in visual quality and instruction following' is presented without any quantitative evidence, error bars, specific metrics, or named baselines, which is critical for substantiating the central claim of improved gradient efficacy via trajectory synchronization.
  2. Abstract: No derivation or equation is provided for how the training noise is synchronized with inference-time denoising paths, leaving open whether this mechanism reliably maximizes gradient signal or introduces new biases, as assumed in the method.
minor comments (2)
  1. Consider adding a figure or diagram in the method section to illustrate the noise synchronization process and the multi-stage framework.
  2. The abstract is quite dense; breaking it into clearer sentences could improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment point by point below, indicating revisions where appropriate to improve clarity and substantiation.

read point-by-point responses

  1. Referee: Abstract: The statement that Diffusion-APO 'consistently outperforms standard baselines in visual quality and instruction following' is presented without any quantitative evidence, error bars, specific metrics, or named baselines, which is critical for substantiating the central claim of improved gradient efficacy via trajectory synchronization.

    Authors: We agree that the abstract would benefit from greater specificity to support its claims. The full manuscript presents quantitative results with named baselines (including DPO and GRPO), specific metrics, and error bars in the experimental evaluation. To directly address this point, we will revise the abstract to briefly reference key quantitative improvements from those experiments. revision: yes

  2. Referee: Abstract: No derivation or equation is provided for how the training noise is synchronized with inference-time denoising paths, leaving open whether this mechanism reliably maximizes gradient signal or introduces new biases, as assumed in the method.

    Authors: The trajectory-aware noise synchronization is formally derived with supporting equations in Section 3 of the manuscript, where we detail the alignment between training noise distributions and inference denoising paths to improve gradient signals. Abstracts conventionally omit derivations for brevity. We will revise the abstract to include a concise reference to this synchronization mechanism. Ablation studies in the paper confirm improved performance and stability without evidence of introduced biases. revision: partial

Circularity Check

0 steps flagged

No circularity in visible derivation chain

full rationale

The provided abstract and description introduce Diffusion-APO as an algorithmic proposal that synchronizes training noise with inference trajectories to address misalignment in video diffusion models, extending DPO/GRPO without any equations, derivations, or parameter fits shown. No self-definitional reductions, fitted inputs renamed as predictions, or load-bearing self-citations appear in the text. The multi-stage framework (online ranking, drift correction, etc.) is presented as a practical implementation rather than a closed loop reducing to its inputs by construction. The derivation chain is therefore self-contained as a novel framing and engineering contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Based solely on abstract: the central claim rests on the unproven premise that noise-trajectory synchronization maximizes gradient efficacy and that the modular framework avoids scalar-reward issues. No explicit free parameters, axioms, or invented entities are detailed.

axioms (2)
  • domain assumption Existing DPO and GRPO are hindered by bias-prone reward models or suboptimal timestep sampling
    Stated in abstract as motivation for the new method.
  • ad hoc to paper Synchronizing training noise with inference denoising paths improves preference alignment
    Core algorithmic assumption not derived in abstract.

pith-pipeline@v0.9.0 · 5506 in / 1349 out tokens · 27119 ms · 2026-05-11T01:55:53.665802+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

66 extracted references · 66 canonical work pages · 12 internal anchors

  1. [1]

    Y . Bai, S. Kadavath, S. Kundu, A. Askell, J. Kernion, A. Jones, A. Chen, A. Goldie, A. Mirhoseini, C. McKinnon, et al. Constitutional ai: Harmlessness from ai feedback.arXiv preprint arXiv:2212.08073, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  2. [2]

    F. Bao, C. Xiang, G. Yue, G. He, H. Zhu, K. Zheng, M. Zhao, S. Liu, Y . Wang, and J. Zhu. Vidu: a highly consistent, dynamic and skilled text-to-video generator with diffusion models.arXiv preprint arXiv:2405.04233, 2024

    work page arXiv 2024

  3. [3]

    Bar-Tal, H

    O. Bar-Tal, H. Chefer, O. Tov, C. Herrmann, R. Paiss, S. Zada, A. Ephrat, J. Hur, G. Liu, A. Raj, et al. Lumiere: A space-time diffusion model for video generation. InSIGGRAPH Asia 2024 Conference Papers, pages 1–11, 2024

    work page 2024

  4. [4]

    Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets

    A. Blattmann, T. Dockhorn, S. Kulal, D. Mendelevitch, M. Kilian, D. Lorenz, Y . Levi, Z. English, V . V oleti, A. Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  5. [5]

    Blattmann, R

    A. Blattmann, R. Rombach, H. Ling, T. Dockhorn, S. W. Kim, S. Fidler, and K. Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22563–22575, 2023

    work page 2023

  6. [6]

    Brooks, B

    T. Brooks, B. Peebles, C. Holmes, W. DePue, Y . Guo, L. Jing, D. Schnurr, J. Taylor, T. Luhman, E. Luhman, et al. Video generation models as world simulators.OpenAI Blog, 1(8):1, 2024

    work page 2024

  7. [7]

    H. Chen, M. Xia, Y . He, Y . Zhang, X. Cun, S. Yang, J. Xing, Y . Liu, Q. Chen, X. Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation.arXiv preprint arXiv:2310.19512, 2023

    work page internal anchor Pith review arXiv 2023

  8. [8]

    H. Chen, Y . Zhang, X. Cun, M. Xia, X. Wang, C. Weng, and Y . Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7310–7320, 2024

    work page 2024

  9. [9]

    Z. Chen, Y . Deng, H. Yuan, K. Ji, and Q. Gu. Self-play fine-tuning converts weak language models to strong language models. InInternational Conference on Machine Learning, pages 6621–6642. PMLR, 2024

    work page 2024

  10. [10]

    Y . Guo, C. Yang, A. Rao, Z. Liang, Y . Wang, Y . Qiao, M. Agrawala, D. Lin, and B. Dai. Animated- iff: Animate your personalized text-to-image diffusion models without specific tuning. InThe Twelfth International Conference on Learning Representations, 2024

    work page 2024

  11. [11]

    J. Ho, W. Chan, C. Saharia, J. Whang, R. Gao, A. Gritsenko, D. P. Kingma, B. Poole, M. Norouzi, D. J. Fleet, et al. Imagen video: High definition video generation with diffusion models.arXiv preprint arXiv:2210.02303, 2022

    work page internal anchor Pith review arXiv 2022

  12. [12]

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

    work page 2020

  13. [13]

    Ho and T

    J. Ho and T. Salimans. Classifier-free diffusion guidance. InNeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications

    work page 2021

  14. [14]

    J. Ho, T. Salimans, A. Gritsenko, W. Chan, M. Norouzi, and D. J. Fleet. Video diffusion models.Advances in neural information processing systems, 35:8633–8646, 2022

    work page 2022

  15. [15]

    W. Hong, M. Ding, W. Zheng, X. Liu, and J. Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. InThe Eleventh International Conference on Learning Representations, 2022

    work page 2022

  16. [16]

    Huang, Y

    Z. Huang, Y . He, J. Yu, F. Zhang, C. Si, Y . Jiang, Y . Zhang, T. Wu, Q. Jin, N. Chanpaisit, Y . Wang, X. Chen, L. Wang, D. Lin, Y . Qiao, and Z. Liu. VBench: Comprehensive benchmark suite for video generative models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    work page 2024

  17. [17]

    Huang, F

    Z. Huang, F. Zhang, X. Xu, Y . He, J. Yu, Z. Dong, Q. Ma, N. Chanpaisit, C. Si, Y . Jiang, Y . Wang, X. Chen, Y .-C. Chen, L. Wang, D. Lin, Y . Qiao, and Z. Liu. VBench++: Comprehensive and versatile benchmark suite for video generative models.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    work page 2025

  18. [18]

    Hyung, K

    J. Hyung, K. Kim, S. Hong, M.-J. Kim, and J. Choo. Spatiotemporal skip guidance for enhanced video diffusion sampling. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11006–11015, 2025. 10

    work page 2025

  19. [19]

    Khachatryan, A

    L. Khachatryan, A. Movsisyan, V . Tadevosyan, R. Henschel, Z. Wang, S. Navasardyan, and H. Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954–15964, 2023

    work page 2023

  20. [20]

    J. Li, Y . Cui, T. Huang, Y . Ma, C. Fan, Y . Cheng, M. Yang, Z. Zhong, and L. Bo. Mixgrpo: Unlocking flow-based grpo efficiency with mixed ode-sde.arXiv preprint arXiv:2507.21802, 2025

    work page internal anchor Pith review arXiv 2025

  21. [21]

    J. Li, W. Feng, T.-J. Fu, X. Wang, S. Basu, W. Chen, and W. Y . Wang. T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback.Advances in neural information processing systems, 37:75692–75726, 2024

    work page 2024

  22. [22]

    S. Li, K. Kallidromitis, A. Gokul, Y . Kato, and K. Kozuka. Aligning diffusion models by optimizing human utility.Advances in Neural Information Processing Systems, 37:24897–24925, 2024

    work page 2024

  23. [23]

    J. Liu, G. Liu, J. Liang, Y . Li, J. Liu, X. Wang, P. Wan, D. ZHANG, and W. Ouyang. Flow-grpo: Training flow matching models via online rl. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    work page 2025

  24. [24]

    C. Meng, R. Rombach, R. Gao, D. Kingma, S. Ermon, J. Ho, and T. Salimans. On distillation of guided diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14297–14306, 2023

    work page 2023

  25. [25]

    Sora: Creating video from text.https://openai.com/sora, 2024

    OpenAI. Sora: Creating video from text.https://openai.com/sora, 2024

    work page 2024

  26. [26]

    Ouyang, J

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, et al. Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    work page 2022

  27. [27]

    Peebles and S

    W. Peebles and S. Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    work page 2023

  28. [28]

    Video diffusion align- ment via reward gradients.arXiv preprint arXiv:2407.08737,

    M. Prabhudesai, R. Mendonca, Z. Qin, K. Fragkiadaki, and D. Pathak. Video diffusion alignment via reward gradients.arXiv preprint arXiv:2407.08737, 2024

    work page arXiv 2024

  29. [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

    work page 2023

  30. [30]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    work page 2022

  31. [31]

    Salimans and J

    T. Salimans and J. Ho. Progressive distillation for fast sampling of diffusion models. InInternational Conference on Learning Representations, 2022

    work page 2022

  32. [32]

    Seedance 1.5 pro: A native audio-visual joint generation foundation model.arXiv preprint arXiv:2512.13507, 2025

    T. Seedance, H. Chen, S. Chen, X. Chen, Y . Chen, Y . Chen, Z. Chen, F. Cheng, T. Cheng, X. Cheng, et al. Seedance 1.5 pro: A native audio-visual joint generation foundation model.arXiv preprint arXiv:2512.13507, 2025

    work page arXiv 2025

  33. [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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  34. [34]

    Singer, A

    U. Singer, A. Polyak, T. Hayes, X. Yin, J. An, S. Zhang, Q. Hu, H. Yang, O. Ashual, O. Gafni, et al. Make-a-video: Text-to-video generation without text-video data. InThe Eleventh International Conference on Learning Representations, 2022

    work page 2022

  35. [35]

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative mod- eling through stochastic differential equations. InInternational Conference on Learning Representations, 2021

    work page 2021

  36. [36]

    K. A. Team. Kling: A large-scale text-to-video generation model. https://klingai.kuaishou.com,

    work page

  37. [37]

    Kuaishou AI Technical Report

    work page

  38. [38]

    T., Castro, S., Kunze, J., and Erhan, D

    R. Villegas, M. Babaeizadeh, P.-J. Kindermans, H. Moraldo, H. Zhang, M. T. Saffar, S. Castro, J. Kunze, and D. Erhan. Phenaki: Variable length video generation from open domain textual description.arXiv preprint arXiv:2210.02399, 2022. 11

    work page arXiv 2022

  39. [39]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228–8238, 2024

    work page 2024

  40. [40]

    T. Wan, A. Wang, B. Ai, B. Wen, C. Mao, C.-W. Xie, D. Chen, F. Yu, H. Zhao, J. Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  41. [41]

    Coefficients-preserving sampling for reinforcement learning with flow matching

    F. Wang and Z. Yu. Coefficients-preserving sampling for reinforcement learning with flow matching.arXiv preprint arXiv:2509.05952, 2025

    work page arXiv 2025

  42. [42]

    J. Wang, H. Yuan, D. Chen, Y . Zhang, X. Wang, and S. Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023

    work page internal anchor Pith review arXiv 2023

  43. [43]

    Y . Wang, X. Chen, X. Ma, S. Zhou, Z. Huang, Y . Wang, C. Yang, Y . He, J. Yu, P. Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models.International Journal of Computer Vision, 133(5):3059–3078, 2025

    work page 2025

  44. [44]

    J. Z. Wu, Y . Ge, X. Wang, S. W. Lei, Y . Gu, Y . Shi, W. Hsu, Y . Shan, X. Qie, and M. Z. Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. InProceedings of the IEEE/CVF international conference on computer vision, pages 7623–7633, 2023

    work page 2023

  45. [45]

    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, 36:15903–15935, 2023

    work page 2023

  46. [46]

    Z. Xue, J. Wu, Y . Gao, F. Kong, L. Zhu, M. Chen, Z. Liu, W. Liu, Q. Guo, W. Huang, et al. Dancegrpo: Unleashing grpo on visual generation.arXiv preprint arXiv:2505.07818, 2025

    work page internal anchor Pith review arXiv 2025

  47. [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. InThe Thirteenth International Conference on Learning Representations, 2024

    work page 2024

  48. [48]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6463–6474, 2024

    work page 2024

  49. [49]

    Zhang, J

    J. Zhang, J. Wu, W. Chen, Y . Ji, X. Xiao, W. Huang, and K. Han. Align video diffusion model with online video-centric preference optimization. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6142–6152, 2026

    work page 2026

  50. [50]

    VBench-2.0: Advancing Video Generation Benchmark Suite for Intrinsic Faithfulness

    D. Zheng, Z. Huang, H. Liu, K. Zou, Y . He, F. Zhang, Y . Zhang, J. He, W.-S. Zheng, Y . Qiao, and Z. Liu. VBench-2.0: Advancing video generation benchmark suite for intrinsic faithfulness.arXiv preprint arXiv:2503.21755, 2025

    work page internal anchor Pith review arXiv 2025

  51. [51]

    DiffusionNFT: Online Diffusion Reinforcement with Forward Process

    K. Zheng, H. Chen, H. Ye, H. Wang, Q. Zhang, K. Jiang, H. Su, S. Ermon, J. Zhu, and M.-Y . Liu. Diffusionnft: Online diffusion reinforcement with forward process.arXiv preprint arXiv:2509.16117, 2025

    work page internal anchor Pith review arXiv 2025

  52. [52]

    MagicVideo: Efficient Video Generation With Latent Diffusion Models

    D. Zhou, W. Wang, H. Yan, W. Lv, Y . Zhu, and J. Feng. Magicvideo: Efficient video generation with latent diffusion models.arXiv preprint arXiv:2211.11018, 2022. 12 A Limitations Although Diffusion-APO effectively mitigates train-inference timestep discrepancies, its optimization efficacy remains fundamentally bottlenecked by the fidelity of the preferenc...

    work page internal anchor Pith review arXiv 2022

  53. [53]

    Primary Criterion - Degree of Visual Distortion: • a) Severe Defects:Anomalies violating physical laws (e.g., anti-gravity cars, unnatural movement of objects without applied force), unnatural transitions/black screens/objects appearing or disappearing out of nowhere, severe anatomical distortions (severed limbs, misplaced facial features), or visible mos...

    work page

  54. [54]

    • Avoid selecting completely static or nearly motionless videos

    Secondary Criterion - Dynamic Performance: • When both videos have no obvious distortions, prioritize the video with larger motion amplitude and a more vibrant visual dynamic. • Avoid selecting completely static or nearly motionless videos. Please compare Video 1 and Video 2, and outputONLYone of the following four options:

    work page

  55. [55]

    Video 1 has a lower degree of visual distortion

    work page

  56. [56]

    Video 2 has a lower degree of visual distortion

    work page

  57. [57]

    Both videos have similar degrees of distortion, but Video 1 features larger camera movement or richer human actions

    work page

  58. [58]

    Prompt for Non-Human-Centric Videos <video><video> You are sequentially given two videos, Video 1 and Video 2

    Both videos have similar degrees of distortion, but Video 2 features larger camera movement or richer human actions. Prompt for Non-Human-Centric Videos <video><video> You are sequentially given two videos, Video 1 and Video 2. Please strictly evaluate them based on the following criteria and select the better video: Evaluate whether the following issues ...

    work page

  59. [59]

    • Unnatural twisting, deformation, merging, or blurring of objects

    Severe Visual Anomalies (Any occurrence leads to an immediate low-quality rejec- tion): • Motions violating physical laws (unnatural movement/jittering/floating of objects without applied forces). • Unnatural twisting, deformation, merging, or blurring of objects. • Appearance of mosaics or digital glitches. • Sudden visual jumps, PPT-like fading, or disc...

    work page

  60. [60]

    • Sudden appearance of mosaics in the frame

    Obvious Visual Issues: • Abnormal surface textures (suddenly floating, warping, or distorting). • Sudden appearance of mosaics in the frame. • Unnatural material representation (e.g., sudden changes in leather or metal textures). • Unnatural blurring or over-sharpening of object edges. • Illogical dynamic changes in patterns, motifs, or text. • Shadows an...

    work page

  61. [61]

    • Unnatural scene transitions

    Visual Coherence Issues: • Discontinuous object motion trajectories. • Unnatural scene transitions. • Camera shaking or jittering

    work page

  62. [62]

    • When visual quality is comparable, prioritize the video with natural camera movements and dynamic visual changes

    Dynamic Performance Evaluation: • Visual changes should be natural and smooth; avoid selecting essentially static videos. • When visual quality is comparable, prioritize the video with natural camera movements and dynamic visual changes. Please strictly inspect each video for the above issues, paying primary attention to Severe Visual Anomalies. If any se...

    work page

  63. [63]

    Video 1 has a lower degree of visual distortion/anomalies

    work page

  64. [64]

    Video 2 has a lower degree of visual distortion/anomalies

    work page

  65. [65]

    Both videos have similar degrees of distortion, but Video 1 features larger/better camera movement

    work page

  66. [66]

    C.3 Evaluation Performance Evaluated on strictly aligned internal human-preference test sets, our specialized ranking models demonstrate exceptional precision and recall

    Both videos have similar degrees of distortion, but Video 2 features larger/better camera movement. C.3 Evaluation Performance Evaluated on strictly aligned internal human-preference test sets, our specialized ranking models demonstrate exceptional precision and recall. TheHuman-Centric Ranking Modelachieves accuracy/recall rates of 91.8% and 94.2% across...

    work page