Research map / Image generation / Fast diffusion sampling
RePaint: Inpainting Using Denoising Diffusion Probabilistic Models
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Condition an unconditional DDPM by replacing known pixels during reverse diffusion | RePaint · CelebA-HQ and ImageNet, thin masks | LPIPS: CelebA-HQ Super-Resolve 2×: 0.177 -> 0.029; Altern. Lines: 0.083 -> 0.009. ImageNet Super-Resolve 2×: 0.272 -> 0.183; Al | Helped |
| Use RePaint rather than LaMa for wide masks | RePaint · CelebA-HQ and ImageNet Wide | LPIPS: CelebA-HQ: 0.045 -> 0.059; ImageNet: 0.105 -> 0.134 | Hurt |
| Use RePaint rather than LaMa for narrow masks | RePaint · CelebA-HQ and ImageNet Narrow | LPIPS: CelebA-HQ: 0.028 -> 0.028; ImageNet: 0.061 -> 0.064 | Mixed |
| Use RePaint rather than LaMa for large-area Half and Expand masks | RePaint · CelebA-HQ and ImageNet Half and Expand | LPIPS: CelebA-HQ Half: 0.138 -> 0.165; Expand: 0.342 -> 0.435; ImageNet Half: 0.254 -> 0.304; Expand: 0.534 -> 0.629 | Hurt |
| Add repeated forward-backward resampling rather than only standard reverse diffusion (ablation) | RePaint · ImageNet validation, LaMa Wide mask | LPIPS: 0.168 -> 0.134 | Helped |
| Increase resampling count from r=1 to r=2, r=3, or r=4 (ablation) | RePaint · ImageNet validation, LaMa Wide mask | LPIPS: 0.168 -> 0.148 -> 0.142 -> 0.134 | Helped |
| Increase diffusion steps by slowing down instead of resampling (ablation) | RePaint · ImageNet validation, LaMa Wide mask | LPIPS: 0.168; 0.167; 0.179; 0.161 | No clear effect |
| Use larger jump length j=10 rather than j=1 (ablation) | RePaint · CelebA validation, LaMa Wide mask | LPIPS: r=5: 0.075 -> 0.073; r=10: 0.088 -> 0.068; r=15: 0.065 -> 0.065 | Mixed |
| Increase jump length from j=1 to j=5 (ablation) | RePaint · CelebA validation, LaMa Wide mask | LPIPS: r=5: 0.075 -> 0.072; r=10: 0.088 -> 0.073; r=15: 0.065 -> 0.063 | Mixed |
| Increase resampling count from r=5 to r=10 (ablation) | RePaint · CelebA validation, LaMa Wide mask | LPIPS: j=1: 0.075 -> 0.088; j=5: 0.072 -> 0.073; j=10: 0.073 -> 0.068 | Mixed |
| Use RePaint resampling instead of the SDEdit alternative sampling schedule (ablation) | RePaint · ImageNet, Places2, and CelebA-HQ; six mask types | LPIPS: ImageNet Super-Res.: 0.3902 -> 0.1831; Places2 Super-Res.: 0.2712 -> 0.0991; CelebA-HQ Super-Res.: 0.1132 -> 0.0291 | Helped |
Closest papers by meaning
- Image Inpainting via Iteratively Decoupled Probabilistic Modeling (ICLR 2024 · similarity 0.90)
- Paint by Inpaint: Learning to Add Image Objects by Removing Them First (CVPR 2025 · similarity 0.89)
- Image Inpainting via Tractable Steering of Diffusion Models (ICLR 2024 · similarity 0.89)
- RAD: Region-Aware Diffusion Models for Image Inpainting (CVPR 2025 · similarity 0.89)
- InpDiffusion: Image Inpainting Localization via Conditional Diffusion Models (AAAI 2025 · similarity 0.89)
- Towards Enhanced Image Inpainting: Mitigating Unwanted Object Insertion and Preserving Color Consistency (CVPR 2025 · similarity 0.88)
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.