Research map / Image generation / Text-to-image diffusion
Learning What to Trust: Bayesian Prior-Guided Optimization for Visual Generation
Ruiying Liu, Yuanzhi Liang, Haibin Huang, Tianshu Yu, Chi Zhang
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 |
|---|---|---|---|
| add RAS and CRT (BPGO) to GRPO | Wan2.1-1.3B T2V · T2V VideoAlign-TA | VideoAlign-TA: 0.8984 -> 1.1193 | Improved |
| add RAS and CRT (BPGO) to GRPO | Wan2.2-14B I2V · I2V VideoAlign-TA | VideoAlign-TA: 0.2307 -> 1.0589 | Improved |
| add RAS and CRT (BPGO) vs GRPO on T2I | FLUX · T2I | HPSv2: 0.2564 -> 0.2434 | Hurt |
| add RAS and CRT (BPGO) vs GRPO on T2I | FLUX · T2I | ImageReward: 1.0607 -> 1.2136 | Improved |
| add RAS and CRT (BPGO) to GRPO | Wan2.1-1.3B T2V · T2V VideoAlign-overall | VideoAlign-overall: -0.5411 -> -0.0478 (base 0.0939) | Mixed |
| add reliability-adaptive scaling (RAS) alone to GRPO (ablation) | Wan2.1-1.3B T2V · T2V VideoAlign-TA | VideoAlign-TA: 0.8984 -> 1.2327 | Improved |
| add contrastive reward transformation (CRT) alone to GRPO on I2V (ablation) | Wan2.2-14B I2V · I2V | VideoAlign-TA: 0.2307 -> 0.2162 | Hurt |
| combine CRT with RAS rather than RAS alone (ablation) | BPGO T2V · T2V | VideoAlign-TA: 1.2327 -> 1.1193 | Mixed |
| combine CRT with RAS rather than CRT alone on I2V (ablation) | BPGO I2V · I2V | VideoCLIP-XL: 2.0682 -> 2.6855 | Improved |
| raise RAS scale parameter from 0.5 to 0.9 (ablation) | BPGO T2V · T2V | VideoAlign-TA: 1.2327 -> 1.0334 | Hurt |
| lower RAS scale parameter from 0.5 to 0.1 (ablation) | BPGO T2V · T2V | VideoAlign-TA: 1.2327 -> 1.1866 | Hurt |
Closest papers by meaning
- Seeing What Matters: Visual Preference Policy Optimization for Visual Generation (CVPR 2026 · similarity 0.90)
- MaskSense: Confronting the Visual Exploration Trap in Masked Image Generation (NeurIPS 2026 · similarity 0.88)
- Ctrl-U: Robust Conditional Image Generation via Uncertainty-aware Reward Modeling (ICLR 2025 · similarity 0.88)
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.88)
- Reward-Instruct: A Reward-Centric Approach to Fast Photo-Realistic Image Generation (NeurIPS 2025 · similarity 0.87)
- Co-GRPO: Co-Optimized Group Relative Policy Optimization for Masked Image Generation (NeurIPS 2026 · similarity 0.86)
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.
