pith. sign in

arxiv: 2511.18719 · v4 · pith:PI53PQ3Vnew · submitted 2025-11-24 · 💻 cs.CV

Seeing What Matters: Visual Preference Policy Optimization for Visual Generation

Ziqi Ni , Yuanzhi Liang , Rui Li , Yi Zhou , Haibin Huang , Chi Zhang , Xuelong Li This is my paper

Pith reviewed 2026-05-21 18:38 UTC · model grok-4.3

classification 💻 cs.CV
keywords visual preference policy optimizationGRPO variantperceptual structuring modulereinforcement learning for generationhuman preference alignmentimage and video generationadvantage mapslocalized artifacts
0
0 comments X

The pith

ViPO turns single scalar rewards into pixel-level advantage maps to guide visual generation toward perceptually important regions.

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

The paper introduces Visual Preference Policy Optimization (ViPO) to address a limitation in existing reinforcement learning methods for visual generators. Standard Group Relative Policy Optimization uses one reward number for each whole image or video, which overlooks where problems like artifacts actually occur. ViPO adds a Perceptual Structuring Module that takes pretrained vision backbones and produces detailed spatial and temporal maps showing which parts of the output deserve more training focus. These maps convert the coarse reward into localized advantage signals while keeping the original training procedure stable. Experiments show this leads to stronger matches with human preferences on both familiar and new data for images and videos.

Core claim

ViPO is a GRPO variant that employs a Perceptual Structuring Module to lift scalar feedback into structured, pixel-level advantages by constructing spatially and temporally aware advantage maps with pretrained vision backbones, redistributing optimization pressure toward perceptually important regions while preserving GRPO stability and yielding better in-domain alignment plus out-of-domain generalization on image and video benchmarks.

What carries the argument

The Perceptual Structuring Module, which uses pretrained vision backbones to build spatially and temporally aware advantage maps that redistribute optimization signals to perceptually important regions.

If this is right

  • ViPO improves alignment with human-preference rewards on in-domain image and video benchmarks.
  • The method enhances generalization on out-of-domain evaluations compared to standard GRPO.
  • ViPO remains architecture-agnostic and fully compatible with existing GRPO training pipelines.
  • The approach provides a more expressive learning signal for correcting localized artifacts in visual outputs.

Where Pith is reading between the lines

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

  • The same map-based redistribution idea could be tested in other structured generation tasks such as audio waveforms or 3D scenes where local quality matters.
  • Replacing the pretrained backbones with task-specific fine-tuned ones might further reduce any domain mismatch in the advantage maps.
  • If the maps prove robust, training pipelines could incorporate them into reward models that operate directly on partial generations rather than final outputs.

Load-bearing premise

The Perceptual Structuring Module that uses pretrained vision backbones can reliably construct spatially and temporally aware advantage maps that correctly identify and prioritize perceptually important regions without introducing biases.

What would settle it

Running the same training setup but replacing the constructed advantage maps with uniform random values and observing no performance gain over vanilla GRPO would show the maps are not delivering the claimed benefit.

Figures

Figures reproduced from arXiv: 2511.18719 by Chi Zhang, Haibin Huang, Rui Li, Xuelong Li, Yi Zhou, Yuanzhi Liang, Ziqi Ni.

Figure 1
Figure 1. Figure 1: Brief illustration of our work. Existing GRPO for vi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on Flux. Each group of results is arranged from left to right as follows: outputs from Flux, DanceGRPO, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison on Wan2.1. Each demo group is arranged top-to-bottom as follows: the result from Wan2.1, the output [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison under the redness reward across training [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: More qualitative comparison results for Flux. Each group of images, from left to right, shows the output from Flux, DanceGRPO, [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: More qualitative comparison of video generation. For each group of sequences, the rows correspond to outputs from Wan2.1, [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of allocation maps. (a) Allocation maps [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of results obtained with different ViPO [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: More comparison of results using the redness reward. [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
read the original abstract

Reinforcement learning (RL) has become a powerful tool for post-training visual generative models, with Group Relative Policy Optimization (GRPO) increasingly used to align generators with human preferences. However, existing GRPO pipelines rely on a single scalar reward per sample, treating each image or video as a holistic entity and ignoring the rich spatial and temporal structure of visual content. This coarse supervision hinders the correction of localized artifacts and the modeling of fine-grained perceptual cues. We introduce Visual Preference Policy Optimization (ViPO), a GRPO variant that lifts scalar feedback into structured, pixel-level advantages. ViPO employs a Perceptual Structuring Module that uses pretrained vision backbones to construct spatially and temporally aware advantage maps, redistributing optimization pressure toward perceptually important regions while preserving the stability of standard GRPO. Across both image and video benchmarks, ViPO consistently outperforms vanilla GRPO, improving in-domain alignment with human-preference rewards and enhancing generalization on out-of-domain evaluations. The method is architecture-agnostic, lightweight, and fully compatible with existing GRPO training pipelines, providing a more expressive and informative learning signal for visual generation.

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 / 1 minor

Summary. The manuscript proposes Visual Preference Policy Optimization (ViPO), a variant of Group Relative Policy Optimization (GRPO) for post-training visual generative models. It introduces a Perceptual Structuring Module that employs pretrained vision backbones to lift scalar human-preference rewards into spatially and temporally aware pixel-level advantage maps. These maps are intended to redistribute optimization pressure toward perceptually important regions while maintaining GRPO stability. The authors claim that ViPO consistently outperforms vanilla GRPO across image and video benchmarks, yielding better in-domain alignment and improved out-of-domain generalization. The method is presented as architecture-agnostic and lightweight.

Significance. If the empirical claims hold after addressing the noted concerns, ViPO could offer a practical, plug-in improvement to GRPO-based alignment pipelines by incorporating perceptual structure without major architectural changes. This would be particularly relevant for reducing localized artifacts in image and video generation and for enhancing generalization, building directly on existing RLHF-style methods in the field.

major comments (2)
  1. [Method] Method section (Perceptual Structuring Module description): The central claim that the module produces reliable, human-preference-relevant advantage maps rests on the assumption that pretrained vision backbones (e.g., CLIP, DINO, or video equivalents) do not introduce systematic biases from their training distributions. No ablation studies are described that swap backbones, compare against non-semantic baselines (such as uniform or random maps), or hold the GRPO pipeline fixed while varying only the structuring component. Without such isolation, it remains possible that reported gains arise from implicit regularization effects rather than the claimed spatially/temporally aware redistribution of optimization pressure. This directly affects the validity of the outperformance and generalization results.
  2. [Experiments] Experiments section: The abstract states that ViPO 'consistently outperforms vanilla GRPO' on both in-domain and out-of-domain evaluations, yet the provided description supplies no quantitative metrics, error bars, statistical significance tests, or detailed ablation tables. To support the load-bearing claim of improved alignment and generalization, the manuscript must include specific results (e.g., reward scores, FID, or human preference win rates) with controls that isolate the Perceptual Structuring Module's contribution.
minor comments (1)
  1. [Abstract] The abstract would be strengthened by briefly noting the specific pretrained backbones used and the key quantitative improvements observed, to allow readers to immediately gauge the scale of the reported gains.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below with clarifications and commit to revisions that directly strengthen the empirical support for ViPO.

read point-by-point responses

  1. Referee: [Method] Method section (Perceptual Structuring Module description): The central claim that the module produces reliable, human-preference-relevant advantage maps rests on the assumption that pretrained vision backbones (e.g., CLIP, DINO, or video equivalents) do not introduce systematic biases from their training distributions. No ablation studies are described that swap backbones, compare against non-semantic baselines (such as uniform or random maps), or hold the GRPO pipeline fixed while varying only the structuring component. Without such isolation, it remains possible that reported gains arise from implicit regularization effects rather than the claimed spatially/temporally aware redistribution of optimization pressure. This directly affects the validity of the outperformance and generalization results.

    Authors: We agree that isolating the contribution of the Perceptual Structuring Module is important for validating the core claim. The current manuscript does not include the requested ablations (backbone swaps or non-semantic baselines such as uniform/random maps with GRPO held fixed). In the revised version we will add these experiments, comparing multiple backbones (CLIP, DINO, and video equivalents) against uniform and random advantage maps while keeping the rest of the GRPO pipeline identical. This will provide direct evidence that performance differences arise from the spatially and temporally aware redistribution rather than incidental regularization. revision: yes

  2. Referee: [Experiments] Experiments section: The abstract states that ViPO 'consistently outperforms vanilla GRPO' on both in-domain and out-of-domain evaluations, yet the provided description supplies no quantitative metrics, error bars, statistical significance tests, or detailed ablation tables. To support the load-bearing claim of improved alignment and generalization, the manuscript must include specific results (e.g., reward scores, FID, or human preference win rates) with controls that isolate the Perceptual Structuring Module's contribution.

    Authors: The Experiments section of the full manuscript reports quantitative results, including human-preference win rates, FID scores, and out-of-domain generalization metrics that support the abstract claim. However, we acknowledge that error bars, statistical significance tests, and more granular ablation tables isolating the Perceptual Structuring Module are not sufficiently detailed. In the revision we will expand these sections to include error bars across runs, paired statistical tests, and explicit controls that vary only the structuring module while reporting reward scores, FID, and win rates. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper introduces ViPO as an extension to standard GRPO by adding an independent Perceptual Structuring Module that leverages external pretrained vision backbones to generate pixel-level advantage maps. This structuring step is not defined in terms of the target rewards or outcomes, nor does it reduce any claimed prediction or advantage to a fitted parameter by construction. No self-citations are invoked as load-bearing for uniqueness or ansatz choices in the provided description, and performance gains are framed as empirical results on in-domain and out-of-domain benchmarks rather than mathematical identities. The derivation chain remains self-contained with external components supplying the perceptual signal.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Only the abstract is available, so the ledger is necessarily incomplete and limited to elements explicitly named in the summary text.

axioms (1)
  • domain assumption Pretrained vision backbones can be used to construct reliable spatially and temporally aware advantage maps
    The Perceptual Structuring Module depends on this to redistribute optimization pressure.
invented entities (1)
  • Perceptual Structuring Module no independent evidence
    purpose: To lift scalar feedback into structured pixel-level advantages for visual content
    New component introduced to enable spatially and temporally aware optimization in GRPO.

pith-pipeline@v0.9.0 · 5740 in / 1280 out tokens · 47324 ms · 2026-05-21T18:38:20.633141+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. CreFlow: Corrective Reflow for Sparse-Reward Embodied Video Diffusion RL

    cs.CV 2026-05 conditional novelty 7.0

    CreFlow combines LTL compositional rewards with credit-aware NFT and corrective reflow losses in online RL to improve embodied video diffusion models, raising downstream task success by 23.8 percentage points on eight...

  2. Learning to Credit the Right Steps: Objective-aware Process Optimization for Visual Generation

    cs.CV 2026-04 unverdicted novelty 7.0

    OTCA improves GRPO training for visual generation by estimating step importance in trajectories and adaptively weighting multiple reward objectives.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages · cited by 2 Pith papers · 15 internal anchors

  1. [1]

    Training Diffusion Models with Reinforcement Learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforce- ment learning.arXiv preprint arXiv:2305.13301, 2023. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    Neural mechanisms of selective visual attention.Annual review of neuroscience, 18 (1):193–222, 1995

    Robert Desimone and John Duncan. Neural mechanisms of selective visual attention.Annual review of neuroscience, 18 (1):193–222, 1995. 4

    work page 1995

  3. [3]

    Optimizing ddpm sampling with shortcut fine-tuning.arXiv preprint arXiv:2301.13362,

    Ying Fan and Kangwook Lee. Optimizing ddpm sampling with shortcut fine-tuning.arXiv preprint arXiv:2301.13362,

    work page arXiv

  4. [4]

    Dpok: Reinforcement learning for fine-tuning text-to-image diffu- sion models.Advances in Neural Information Processing Systems, 36:79858–79885, 2023

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Moham- mad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffu- sion models.Advances in Neural Information Processing Systems, 36:79858–79885, 2023. 1, 2

    work page 2023

  5. [5]

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025. 1, 3

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    Deep saliency mod- els learn low-, mid-, and high-level features to predict scene attention.Scientific reports, 11(1):18434, 2021

    Taylor R Hayes and John M Henderson. Deep saliency mod- els learn low-, mid-, and high-level features to predict scene attention.Scientific reports, 11(1):18434, 2021. 2

    work page 2021

  7. [7]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 5, 1

    work page 2016

  8. [8]

    Videoscore: Building automatic metrics to simulate fine-grained human feedback for video genera- tion.arXiv preprint arXiv:2406.15252, 2024

    Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, et al. Videoscore: Building automatic metrics to simulate fine-grained human feedback for video genera- tion.arXiv preprint arXiv:2406.15252, 2024. 2

    work page arXiv 2024

  9. [9]

    TempFlow-GRPO: When Timing Matters for GRPO in Flow Models

    Xiaoxuan He, Siming Fu, Yuke Zhao, Wanli Li, Jian Yang, Dacheng Yin, Fengyun Rao, and Bo Zhang. Tempflow-grpo: When timing matters for grpo in flow models.arXiv preprint arXiv:2508.04324, 2025. 2

    work page internal anchor Pith review arXiv 2025

  10. [10]

    Meaning-based guidance of attention in scenes as revealed by meaning maps

    John M Henderson and Taylor R Hayes. Meaning-based guidance of attention in scenes as revealed by meaning maps. Nature Human Behaviour, 1:743–747, 2017. 2, 4

    work page 2017

  11. [11]

    Meaning guides attention in real-world scene images: Evidence from eye movements and meaning maps.Journal of vision, 18(6):10– 10, 2018

    John M Henderson and Taylor R Hayes. Meaning guides attention in real-world scene images: Evidence from eye movements and meaning maps.Journal of vision, 18(6):10– 10, 2018. 2

    work page 2018

  12. [12]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 1

    work page 2020

  13. [13]

    Vbench: Comprehensive bench- mark suite for video generative models

    Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive bench- mark suite for video generative models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807–21818, 2024. 4

    work page 2024

  14. [14]

    Computational modelling of visual attention.Nature reviews neuroscience, 2(3):194–203,

    Laurent Itti and Christof Koch. Computational modelling of visual attention.Nature reviews neuroscience, 2(3):194–203,

    work page

  15. [15]

    Salicon: Saliency in context

    Ming Jiang, Shengsheng Huang, Juanyong Duan, and Qi Zhao. Salicon: Saliency in context. InProceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 1072–1080, 2015. 2

    work page 2015

  16. [16]

    Perceptual losses for real-time style transfer and super-resolution

    Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European conference on computer vision, pages 694–711. Springer, 2016. 2

    work page 2016

  17. [17]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. InProceedings of the IEEE/CVF international confer- ence on computer vision, pages 4015–4026, 2023. 5, 1

    work page 2023

  18. [18]

    Pick-a-pic: An open dataset of user preferences for text-to-image generation.Ad- vances in neural information processing systems, 36:36652– 36663, 2023

    Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Ma- tiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation.Ad- vances in neural information processing systems, 36:36652– 36663, 2023. 2, 4

    work page 2023

  19. [19]

    Flux.https://github.com/ black-forest-labs/flux, 2024

    Black Forest Labs. Flux.https://github.com/ black-forest-labs/flux, 2024. 4

    work page 2024

  20. [20]

    MixGRPO: Unlocking Flow-based GRPO Efficiency with Mixed ODE-SDE

    Junzhe Li, Yutao Cui, Tao Huang, Yinping Ma, Chun Fan, Miles Yang, and Zhao Zhong. Mixgrpo: Unlocking flow- based grpo efficiency with mixed ode-sde.arXiv preprint arXiv:2507.21802, 2025. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  21. [21]

    Integrating reinforcement learning with vi- sual generative models: Foundations and advances.arXiv preprint arXiv:2508.10316, 2025

    Yuanzhi Liang, Yijie Fang, Rui Li, Ziqi Ni, Ruijie Su, and Chi Zhang. Integrating reinforcement learning with vi- sual generative models: Foundations and advances.arXiv preprint arXiv:2508.10316, 2025. 1

    work page arXiv 2025

  22. [22]

    Flow Matching for Generative Modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling.arXiv preprint arXiv:2210.02747, 2022. 1, 2

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  23. [23]

    Improving Video Generation with Human Feedback

    Jie Liu, Gongye Liu, Jiajun Liang, Ziyang Yuan, Xiaokun Liu, Mingwu Zheng, Xiele Wu, Qiulin Wang, Wenyu Qin, Menghan Xia, et al. Improving video generation with human feedback.arXiv preprint arXiv:2501.13918, 2025. 2, 4

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022. 1, 3

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  25. [25]

    DINOv2: Learning Robust Visual Features without Supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 5, 1

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  26. [26]

    Proximal Policy Optimization Algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Rad- ford, and Oleg Klimov. Proximal policy optimization algo- rithms.arXiv preprint arXiv:1707.06347, 2017. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  27. [27]

    Denoising Diffusion Implicit Models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 1

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  28. [28]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions.arXiv preprint arXiv:2011.13456, 2020. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  29. [29]

    Wan: Open and Advanced Large-Scale Video Generative Models

    Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianx- iao Yang, et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314, 2025. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    Emulating human-like adaptive vision for efficient and flexible machine visual perception.Nature Machine In- telligence, pages 1–19, 2025

    Yulin Wang, Yang Yue, Yang Yue, Huanqian Wang, Haojun Jiang, Yizeng Han, Zanlin Ni, Yifan Pu, Minglei Shi, Rui Lu, et al. Emulating human-like adaptive vision for efficient and flexible machine visual perception.Nature Machine In- telligence, pages 1–19, 2025. 2

    work page 2025

  31. [31]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis, 2023

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis, 2023. 4

    work page 2023

  32. [32]

    Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis.arXiv preprint arXiv:2306.09341,

    work page internal anchor Pith review Pith/arXiv arXiv

  33. [33]

    Imagere- ward: Learning and evaluating human preferences for text- to-image generation.Advances in Neural Information Pro- cessing Systems, 36:15903–15935, 2023

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere- ward: Learning and evaluating human preferences for text- to-image generation.Advances in Neural Information Pro- cessing Systems, 36:15903–15935, 2023. 2, 4

    work page 2023

  34. [34]

    VisionReward: Fine-Grained Multi-Dimensional Human Preference Learning for Image and Video Generation

    Jiazheng Xu, Yu Huang, Jiale Cheng, Yuanming Yang, Jiajun Xu, Yuan Wang, Wenbo Duan, Shen Yang, Qunlin Jin, Shu- run Li, et al. Visionreward: Fine-grained multi-dimensional human preference learning for image and video generation. arXiv preprint arXiv:2412.21059, 2024. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  35. [35]

    Show, attend and tell: Neural image caption gen- eration with visual attention

    Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption gen- eration with visual attention. InInternational conference on machine learning, pages 2048–2057. PMLR, 2015. 2

    work page 2048

  36. [36]

    DanceGRPO: Unleashing GRPO on Visual Generation

    Zeyue Xue, Jie Wu, Yu Gao, Fangyuan Kong, Lingting Zhu, Mengzhao Chen, Zhiheng Liu, Wei Liu, Qiushan Guo, Weilin Huang, et al. Dancegrpo: Unleashing grpo on visual generation.arXiv preprint arXiv:2505.07818, 2025. 1, 2, 5

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  37. [37]

    Identity- preserving text-to-video generation by frequency decompo- sition

    Shenghai Yuan, Jinfa Huang, Xianyi He, Yunyang Ge, Yu- jun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. Identity- preserving text-to-video generation by frequency decompo- sition. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 12978–12988, 2025. 4

    work page 2025

  38. [38]

    Vast 1.0: A unified framework for controllable and con- sistent video generation.arXiv preprint arXiv:2412.16677,

    Chi Zhang, Yuanzhi Liang, Xi Qiu, Fangqiu Yi, and Xuelong Li. Vast 1.0: A unified framework for controllable and con- sistent video generation.arXiv preprint arXiv:2412.16677,

    work page arXiv

  39. [39]

    Flow- grpo: Training flow matching models via online reinforce- ment learning

    Da Zhou, Yang Li, Qing Li, Yujia Yang, Jian Tang, Ye- long Shen, Xiang Li, Xinyang Wang, and Pan Zhou. Flow- grpo: Training flow matching models via online reinforce- ment learning. InProceedings of the International Confer- ence on Learning Representations (ICLR), 2024. 1, 2 Seeing What Matters: Visual Preference Policy Optimization for Visual Generation...

    work page 2024