Research map / Image generation / Text-to-image diffusion
DiffDoctor: Diagnosing Image Diffusion Models Before Treating
Yiyang Wang, Xi Chen, Xiaogang Xu, Sihui Ji, Yu Liu, Yujun Shen, Hengshuang Zhao
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 real photos as negatives to detector training (ablation) | RichHF* detector · detector benchmark | MSE (percent): 1.601 -> 1.167 | Improved |
| add hard cases to detector training (ablation) | RichHF* detector · detector benchmark | MSE (percent): 1.167 -> 0.504 | Improved |
| add 1M pseudo-labeled images to detector training (ablation) | RichHF* detector · detector benchmark | MSE (percent): 0.504 -> 0.337 | Improved |
| add 1M pseudo-labeled images to detector training (ablation) | RichHF* detector · detector benchmark | KL (false negative): 0.981 -> 1.004 | Hurt |
| use balanced best detector instead of naive RichHF* detector for treating (ablation) | DiffDoctor · FLUX.1 | ImageReward: -0.884 -> 1.183 | Improved |
| use threshold mask instead of all pixels (ablation) | DiffDoctor · FLUX.1 | ImageReward: 1.161 -> 1.183 | Improved |
| use threshold mask instead of max pixel (ablation) | DiffDoctor · FLUX.1 | ImageReward: 1.159 -> 1.183 | Improved |
| add offline regularization at epoch 3 (ablation) | DiffDoctor · FLUX.1 | ImageReward: 1.183 -> 1.167 | No clear effect |
| add offline regularization at epoch 10 (ablation) | DiffDoctor · FLUX.1 | ImageReward: 0.904 -> 1.180 | Improved |
| add offline regularization at epoch 10 (ablation) | DiffDoctor · FLUX.1 | CLIP-T: 34.952 -> 35.842 | Improved |
| treat FLUX.1 with DiffDoctor | FLUX.1 · FLUX.1 | mean artifact frequency: 82.66% -> 22.00% | Improved |
| treat SDXL with DiffDoctor | SDXL · SDXL | mean artifact frequency: 55.33% -> 27.50% | Improved |
Closest papers by meaning
- Temporal Score Analysis for Understanding and Correcting Diffusion Artifacts (CVPR 2025 · similarity 0.89)
- Diffusion Rejection Sampling (ICML 2024 · similarity 0.87)
- Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · similarity 0.87)
- Image Copy Detection for Diffusion Models (NeurIPS 2024 · similarity 0.87)
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.86)
- Diffusing Differentiable Representations (NeurIPS 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.
