Research map / Image generation / Text-to-image diffusion
Synthetic Curriculum Reinforces Compositional Text-to-Image Generation
Shijian Wang, Runhao Fu, Siyi Zhao, Qingqin Zhan, Xingjian Wang, Jiarui Jin, Yuan Lu, Hanqian Wu, Cunjian Chen
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 |
|---|---|---|---|
| apply CompGen curriculum RL to Stable-Diffusion-1.5 | Stable-Diffusion-1.5 · 5 benchmarks avg | avg score: 54.90% -> 66.62% | Improved |
| apply CompGen curriculum RL to SimpleAR | SimpleAR-0.5B · 5 benchmarks avg | avg score: 63.66% -> 71.27% | Improved |
| use stronger MLLM reward model (LLaVA-v1.6-13B vs InstructBLIP) (ablation) | CompGen reward · SD-1.5 5 benchmarks avg | avg score: 57.22% -> 66.62% | Improved |
| use CompGen fine-grained scene-graph questions with curriculum instead of plain VQAScore reward (ablation) | SD-1.5 RL · 5 benchmarks avg | avg score: 61.44% -> 66.62% | Improved |
| use multiplicative scene-graph difficulty instead of additive (ablation) | CompGen difficulty measure · SD-1.5, 10K prompts | avg score: 62.06% (approx) -> 66.62%; paper states +4.56 points | Improved |
| use multiplicative difficulty instead of additive on T2I-CompBench (ablation) | CompGen difficulty measure · T2I-CompBench | score: 37.00% -> 37.68% | No clear effect |
| add random curriculum sampling vs no curriculum (ablation) | CompGen GRPO · GenEval SD-1.5 | accuracy: 42% -> 52.1% | Improved |
| use Gaussian scheduling instead of random sampling (ablation) | CompGen GRPO · GenEval SD-1.5 | accuracy: 52.1% -> 54.6% | Improved |
| train beyond ~500 steps (ablation) | CompGen GRPO · GenEval SD-1.5 | accuracy | Hurt |
| train only on skew-easy or skew-hard difficulty distributions (ablation) | CompGen · difficulty distribution | compositional generalization | Hurt |
Closest papers by meaning
- Sketch Then Paint: Hierarchical Reinforcement Learning for Diffusion Multi-Modal Large Language Models (NeurIPS 2026 · similarity 0.87)
- IterComp: Iterative Composition-Aware Feedback Learning from Model Gallery for Text-to-Image Generation (ICLR 2025 · similarity 0.87)
- Progressive Compositionality in Text-to-Image Generative Models (ICLR 2025 · similarity 0.87)
- Curriculum Direct Preference Optimization for Diffusion and Consistency Models (CVPR 2025 · similarity 0.86)
- Breaking the SFT Plateau: Multimodal Structured Reinforcement Learning for Chart-to-Code Generation (ICLR 2026 · similarity 0.86)
- PixArt-$\alpha$: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis (ICLR 2024 · 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.
