Research map / Image generation / Text-to-image diffusion
RubricRL: Simple Generalizable Rewards for Text-to-Image Generation
Xuelu Feng, Yunsheng Li, Ziyu Wan, Zixuan Gao, Junsong Yuan, Dongdong Chen, Chunming Qiao
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/composite rewards with rubric reward | Phi3-3.8B AR GRPO · GenEval | overall: 0.8080 (X-Omni) -> 0.8468; SFT 0.7624 | Improved |
| apply RubricRL via DiffusionNFT | Qwen-Image 4B · GenEval | overall: 0.8406 -> 0.9091 | Improved |
| apply RubricRL via DiffusionNFT | FLUX.1 Dev · GenEval | overall: 0.8121 -> 0.8969 | Improved |
| apply RubricRL | Qwen2.5-0.5B · GenEval counting | counting: 0.2500 -> 0.2469 | Hurt |
| apply RubricRL | Qwen2.5-0.5B · DPG-Bench Other | Other: 84.27 -> 68.00 | Hurt |
| apply RubricRL | Phi3-3.8B / Qwen2.5-0.5B · DPG-Bench | overall: 81.25 -> 86.07; 78.02 -> 81.43 | Improved |
| add hybrid rollout sampling (top-K + random) instead of none (ablation) | RubricRL · GenEval | overall: 0.8349 -> 0.8468 | Improved |
| use FFKC-1D or DAPO sampling instead of vanilla (ablation) | RubricRL · GenEval | overall: 0.8349 -> 0.8353 / 0.8333 | No clear effect |
| normalize advantage over retained N rollouts instead of all N' (ablation) | RubricRL · GenEval | overall: 0.8337 -> 0.8468 | Improved |
| use binary instead of continuous rubric scores (ablation) | RubricRL · GenEval | overall: 0.8170 -> 0.8468 | Improved |
| increase number of rubrics M from 4 to 10 (ablation) | RubricRL · GenEval | overall: 0.7866 -> 0.8468 | Improved |
| increase M from 10 to 12 (ablation) | RubricRL · GenEval | overall: 0.8468 -> 0.8427 | No clear effect |
Closest papers by meaning
- Auto-Rubric as Reward: From Implicit Preference to Explicit Generative Criteria (NeurIPS 2026 · similarity 0.91)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.90)
- Subject-driven Text-to-Image Generation via Preference-based Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
- Focus-N-Fix: Region-Aware Fine-Tuning for Text-to-Image Generation (CVPR 2025 · similarity 0.89)
- Preference Adaptive and Sequential Text-to-Image Generation (ICML 2025 · similarity 0.89)
- Learning to Generate via Understanding: Understanding-Driven Intrinsic Rewarding for Unified Multimodal Models (CVPR 2026 · similarity 0.89)
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.
