Research map / Image generation / Text-to-image diffusion
Self-Evaluation Unlocks Any-Step Text-to-Image Generation
Xin Yu, Xiaojuan Qi, Zhengqi Li, Kai Zhang, Richard Zhang, Zhe Lin, Eli Shechtman, Tianyu Wang, Yotam Nitzan
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 energy-preserving target normalization (ablation) | Self-E (0.5B, 100k iterations) · GenEval, 4 steps | Overall: 0.6156 -> 0.6381 | Improved |
| add energy-preserving target normalization (ablation) | Self-E (0.5B, 100k iterations) · GenEval, 2 steps | Overall: 0.5555 -> 0.5439 | Hurt |
| add energy-preserving target normalization (ablation) | Self-E (0.5B, 100k iterations) · GenEval, 50 steps | Overall: 0.7018 -> 0.7160 | Improved |
| include auxiliary (fake score) term from the start of training (ablation) | Self-E (0.5B, 100k iterations) · GenEval, 2 steps | Overall: 0.5439 -> 0.3307 | Hurt |
| include auxiliary term from the start of training (ablation) | Self-E (0.5B, 100k iterations) · GenEval, 50 steps | Overall: 0.7160 -> 0.6166 | Hurt |
| re-introduce auxiliary term in later training stages (ablation) | Self-E (2B) · 2-step qualitative | artifact severity | Improved |
| use self-evaluation loss versus flow matching only | 0.5B model, 300k iterations · GenEval, 2 steps | Overall: 0.2523 -> 0.6097 | Improved |
| use self-evaluation loss versus flow matching only | 0.5B model, 300k iterations · GenEval, 50 steps | Overall: 0.7311 -> 0.7543 | Improved |
| use self-evaluation versus IMM native few-step training | 0.5B model, 300k iterations · GenEval, 4 steps | Overall: 0.5994 -> 0.7121 | Improved |
| increase inference steps | Self-E (2B) · GenEval | Overall: 0.753 -> 0.781 -> 0.785 -> 0.815 (2/4/8/50 steps) | Improved |
| compared to TiM (concurrent any-step model) at few steps | Self-E (2B) · GenEval, 2 steps | Overall: 0.6338 -> 0.7531 | Improved |
| Self-E vs FLUX.1-Dev at 50 steps | Self-E (2B) · GenEval, 50 steps | Overall: 0.7966 -> 0.8151 | Improved |
Closest papers by meaning
- Diffusion Self-Distillation for Zero-Shot Customized Image Generation (CVPR 2025 · similarity 0.87)
- InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (ICLR 2024 · similarity 0.86)
- Your Student is Better Than Expected: Adaptive Teacher-Student Collaboration for Text-Conditional Diffusion Models (CVPR 2024 · similarity 0.86)
- Transition Matching: Scalable and Flexible Generative Modeling (NeurIPS 2025 · similarity 0.86)
- OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation (NeurIPS 2025 · similarity 0.86)
- Revisiting text-to-image evaluation with Gecko: on metrics, prompts, and human rating (ICLR 2025 · 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.
