Research map / Image generation / Fast diffusion sampling
Region-Aware Test-Time Scaling for Compositional Image Generation
Mingzhu Shen, Peng Ye, Xinyin Ma, Gongfan Fang, Christos-Savvas Bouganis, Yiren Zhao, Xinchao Wang
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 |
|---|---|---|---|
| add RAS (region-aware generation) at a single sample | FLUX.1-dev · GenEval | overall score: 0.67 -> 0.77 | Improved |
| add RAS instead of noise scaling at 32 samples | FLUX.1-dev · GenEval | overall score: 0.85 (N=32 noise) vs 0.85 (RAS N=2) | Improved |
| add RAS with 4 samples versus prompt scaling with 32 | FLUX.1-dev · GenEval | overall score: 0.87 -> 0.88 | Improved |
| add RAS to ReflectionFlow | ReflectionFlow · GenEval | overall score: 0.91 -> 0.92 | Improved |
| add RAS to SANA-1.0-1.6B with 8 samples versus 20 noise samples | SANA-1.0-1.6B · GenEval | overall score: 0.80 -> 0.81 | Improved |
| add regional guidance (RAS) to FLUX at N=1 | FLUX.1-dev · GenEval Position | accuracy: 0.21 -> 0.67 | Improved |
| add LLM rewrite to vanilla FLUX (ablation) | FLUX.1-dev · GenEval Position | accuracy: 0.21 -> 0.40 | Improved |
| add LLM planning and RAS after rewrite (ablation) | FLUX.1-dev · GenEval Position | accuracy: 0.40 -> 0.60 | Improved |
| add Best-of-2 selection on top of RAS (ablation) | FLUX.1-dev · GenEval Position | accuracy: 0.60 -> 0.77 | Improved |
| add a single region prompt to noise-only scaling (ablation) | FLUX.1-dev · GenEval Position | prompt-level success rate: 82% -> 89% | Improved |
| balance budget across noise/global/region (2/4/4) versus all noise (ablation) | FLUX.1-dev · GenEval Position | prompt-level success rate: 82% -> 92% | Improved |
| use GPT-4o instead of Qwen2-VL-7B as planner (ablation) | RAS · GenEval Position | layout parsing accuracy: 52% -> 90% | Improved |
Closest papers by meaning
- Scale Where It Matters: Training-Free Localized Scaling for Diffusion Models (NeurIPS 2026 · similarity 0.89)
- TTS-VAR: A Test-Time Scaling Framework for Visual Auto-Regressive Generation (NeurIPS 2025 · similarity 0.86)
- Improving Compositional Generation with Diffusion Models Using Lift Scores (ICML 2025 · similarity 0.86)
- From Scale to Speed: Adaptive Test-Time Scaling for Image Editing (CVPR 2026 · similarity 0.85)
- SANA 1.5: Efficient Scaling of Training-Time and Inference-Time Compute in Linear Diffusion Transformer (ICML 2025 · similarity 0.85)
- Coarse-to-Fine Compositional Diffusion for Long-Horizon Planning (NeurIPS 2026 · similarity 0.85)
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.
