Research map / Reinforcement learning / RL from feedback
Seeing What Matters: Visual Preference Policy Optimization for Visual Generation
Ziqi Ni, Yuanzhi Liang, Rui Li, Yi Zhou, Haibin Huang, Chi Zhang, Xuelong Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace scalar GRPO advantage with perceptual allocation-map advantages (DINO) | DanceGRPO on FLUX.1-dev · HPD test | HPSv2.1: 0.3203 -> 0.3321 | Improved |
| add ViPO to DanceGRPO | FLUX.1-dev · HPD test | ImageReward (OOD): 1.0392 -> 1.1883 | Improved |
| DanceGRPO training alone | FLUX.1-dev · HPD test | ImageReward / PickScore: 1.1495 -> 1.0392 (ImageReward); 22.7038 -> 22.5962 (PickScore) | Hurt |
| use SAM backbone instead of DINO in PSM (ablation) | ViPO · HPD test | HPSv2.1: 0.3321 -> 0.3219 | Hurt |
| use ResNet backbone in PSM instead of DINO (ablation) | ViPO · HPD test | PickScore: 22.8305 -> 22.8492 | Improved |
| apply ViPO to video | Wan2.1-T2V-14B DanceGRPO · VidProM / VBench | VBench Total: 80.84 -> 81.70 | Improved |
| add ViPO to DanceGRPO on video | Wan2.1 · VBench | Quality: 69.68 -> 72.59 | Improved |
| use semantically structured allocation map instead of uniform all-ones map (ablation) | ViPO · HPD test | HPSv2.1: 0.3043 -> 0.3321 | Improved |
| apply allocation map to advantage rather than reward (ablation) | ViPO · HPD test | ImageReward: 1.0058 -> 1.1883 | Improved |
| variance-weighted rather than average aggregation of components (ablation) | ViPO · HPD test | ImageReward: 1.1318 -> 1.1883 | Improved |
| vary retained PCA components K from 1 to 5 (ablation) | ViPO · HPD test | HPSv2.1/PickScore/ImageReward: K=3: 0.3321/22.8305/1.1883; K=4 HPS 0.3337 | Mixed |
| add Gaussian smoothing sigma=1 vs none (ablation) | ViPO · HPD test | ImageReward: 1.1618 -> 1.1883 | Improved |
Closest papers by meaning
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.89)
- DrPO: Drifting Preference Optimization for One-Step Generative Models (NeurIPS 2026 · similarity 0.88)
- RoVRM: A Robust Visual Reward Model Optimized via Auxiliary Textual Preference Data (AAAI 2025 · similarity 0.88)
- MIA-DPO: Multi-Image Augmented Direct Preference Optimization For Large Vision-Language Models (ICLR 2025 · similarity 0.88)
- MaskSense: Confronting the Visual Exploration Trap in Masked Image Generation (NeurIPS 2026 · similarity 0.88)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
