Research map / Image generation / Text-to-image diffusion
From Reflection to Perfection: Scaling Inference-Time Optimization for Text-to-Image Diffusion Models via Reflection Tuning
Le Zhuo, Liangbing Zhao, Sayak Paul, Yue Liao, Renrui Zhang, Yi Xin, Peng Gao, Mohamed Elhoseiny, Hongsheng Li
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 noise-level scaling (32 samples) | FLUX.1-dev · GenEval | Overall: 0.67 -> 0.85 | Improved |
| add prompt-level scaling (ablation) | FLUX.1-dev + noise scaling · GenEval | Overall: 0.85 -> 0.87 | Improved |
| add reflection-level scaling (ablation) | FLUX.1-dev + noise & prompt scaling · GenEval | Overall: 0.87 -> 0.91 | Improved |
| add reflection-level scaling (ablation) | FLUX.1-dev + noise & prompt scaling · GenEval | Position: 0.80 -> 0.93 | Improved |
| add reflection-level scaling | FLUX.1-dev + noise scaling · GenEval | Attribute binding: 0.78 -> 0.72 | Hurt |
| use sequential refinement (N1 M16) instead of parallel (N16 M1) (ablation) | ReflectionFlow · GenEval (GPT-4o verifier, budget 16) | Overall: 0.74 -> 0.78 | Improved |
| increase refinement depth with fixed budget (N8M2 to N2M8) (ablation) | ReflectionFlow · GenEval (GPT-4o verifier, budget 16) | Overall: 0.76 -> 0.78 | Improved |
| use SANA verifier instead of GPT-4o verifier (ablation) | ReflectionFlow · GenEval | Overall: 0.91 | Improved |
| use oracle verifier (upper bound) instead of SANA verifier (ablation) | ReflectionFlow · GenEval | Overall: 0.91 -> 0.98 | Improved |
| apply reflection to hard prompts | ReflectionFlow · GenEval hard prompts | correctness: 0.10 -> 0.81 | Improved |
| apply reflection to medium prompts | ReflectionFlow · GenEval medium prompts | correctness: 0.55 -> 0.85 | Improved |
| apply reflection to easy prompts | ReflectionFlow · GenEval easy prompts | correctness: 0.95 -> 0.97 | No clear effect |
Closest papers by meaning
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.90)
- Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models (ICLR 2024 · similarity 0.90)
- SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation (CVPR 2024 · similarity 0.90)
- Zigzag Diffusion Sampling: Diffusion Models Can Self-Improve via Self-Reflection (ICLR 2025 · similarity 0.90)
- Scaling Diffusion Transformers Efficiently via $\mu$P (NeurIPS 2025 · similarity 0.89)
- SANA 1.5: Efficient Scaling of Training-Time and Inference-Time Compute in Linear Diffusion Transformer (ICML 2025 · 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.
