Research map / Reinforcement learning / RL from feedback
MaskFocus: Focusing Policy Optimization on Critical Steps for Masked Image Generation
Guohui Zhang, Hu Yu, Xiaoxiao Ma, Yaning Pan, Hang Xu, Jie Huang, Feng Zhao
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 |
|---|---|---|---|
| select critical steps by information gain instead of random early steps (ablation) | MaskFocus · GenEval | accuracy: 0.72 -> 0.76 | Improved |
| select critical steps by information gain (ablation) | MaskFocus · DEQA | score: 4.34 -> 4.39 | Improved |
| add dynamic routing sampling (ablation) | MaskFocus · GenEval | accuracy: 0.74 -> 0.76 | Improved |
| add dynamic routing sampling (ablation) | MaskFocus · PickScore | score: 22.31 -> 22.39 | Improved |
| use entropy-based sampling for all samples instead of dynamic routing (ablation) | MaskFocus · RL training curves | training stability | Hurt |
| use MaskGIT sampling instead of dynamic routing (ablation) | MaskFocus · RL training curves | reward improvement | Hurt |
| randomly select steps within first 40% of steps instead of critical steps (ablation) | MaskFocus · RL training curves | late-stage performance | Hurt |
| re-mask with trajectory masks instead of regenerating masks (ablation) | MaskFocus · RL training curves | training performance | Hurt |
| train with MaskFocus instead of MaskGRPO | Meissonic · GenEval | overall accuracy: 0.73 -> 0.76 | Improved |
| train with MaskFocus instead of MaskGRPO | Meissonic · DrawBench PickScore | PickScore: 22.34 -> 22.39 | No clear effect |
| use moderate CFG (5) rather than higher or lower in RL training (ablation) | MaskFocus · RL training curves | training performance | Improved |
Closest papers by meaning
- MaskSense: Confronting the Visual Exploration Trap in Masked Image Generation (NeurIPS 2026 · similarity 0.93)
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.91)
- Co-GRPO: Co-Optimized Group Relative Policy Optimization for Masked Image Generation (NeurIPS 2026 · similarity 0.89)
- Fixed-Point Masked Generative Modeling (NeurIPS 2026 · similarity 0.88)
- Plug-and-Play Context Feature Reuse for Efficient Masked Generation (NeurIPS 2025 · similarity 0.88)
- Unlocking the Capabilities of Masked Generative Models for Image Synthesis via Self-Guidance (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.
