Research map / Image generation / Fast diffusion sampling
TD-Paint: Faster Diffusion Inpainting Through Time-Aware Pixel Conditioning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TD-Paint, a time-aware pixel-conditioning method that makes diffusion inpainting faster by better exploiting the sampling process, evaluated on CelebA-HQ, ImageNet1K and Places2 masks against RePaint, MCG, CoPaint, LDM, ControlNet, BLD, Uni-paint and PowerPaint.
Problem
Diffusion-based inpainting models sample slowly, and recent quality improvements have further increased sampling time; the paper observes these models make minimal use of the known pixels during generation.
Approach
Time-aware conditioning on the known pixels across diffusion timesteps so the model uses observed content more effectively, allowing fewer sampling steps while keeping inpainting quality.
Key results
- On CelebA-HQ Wide masks, TD-Paint reaches LPIPS 0.0554 and SSIM 0.873 versus RePaint-20's 0.0666 and 0.864.
- On CelebA-HQ Altern. Lines, TD-Paint achieves the best LPIPS 0.0096 and SSIM 0.971 among compared methods.
- On ImageNet1K Half masks, TD-Paint attains LPIPS 0.257 vs BLD's 0.356, while BLD is best on Wide masks (LPIPS 0.050 vs TD-Paint 0.099).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TD-Paint instead of the strongest listed baseline | CelebA-HQ inpainting · Wide | LPIPS: 0.01384996622800827 -> 0.05542798712849617 | Hurt |
| Use TD-Paint instead of the strongest listed baseline | CelebA-HQ inpainting · Narrow | LPIPS: 0.016606109216809273 -> 0.02750529907643795 | Hurt |
| Use TD-Paint instead of the strongest listed baseline | CelebA-HQ inpainting · Super-Resolve 2x | LPIPS: 0.032878048717975616 -> 0.01711963303387165 | Helped |
| Use TD-Paint instead of the strongest listed baseline | ImageNet1K inpainting · Wide | LPIPS: 0.049798160791397095 -> 0.09947434812784196 | Hurt |
| Use TD-Paint instead of the strongest listed baseline | ImageNet1K inpainting · Narrow | SSIM: 0.8350746035575867 -> 0.8636433482170105 | Helped |
| Use TD-Paint instead of the strongest listed baseline | ImageNet1K inpainting · Super-Resolve 2x | SSIM: 0.8183684349060059 -> 0.6480721235275269 | Hurt |
| Use TD-Paint instead of the strongest listed baseline | Places2 inpainting · Wide | LPIPS: 0.03526424244046211 -> 0.11154896020889282 | Hurt |
| Use TD-Paint instead of the strongest listed baseline | Places2 inpainting · Narrow | SSIM: 0.8860372304916382 -> 0.8652424812316895 | Hurt |
| Use TD-Paint instead of the strongest listed baseline | Places2 inpainting · Super-Resolve 2x | LPIPS: 0.07137319445610046 -> 0.1295531690120697 | Helped |
Surprising findings
- BLD achieves the best LPIPS on CelebA-HQ Wide and Narrow masks, outperforming TD-Paint on those settings.
- LDM and ControlNet perform very poorly on Super-Resolve 2x masks (LPIPS above 0.5 on CelebA-HQ).
Limitations
- TD-Paint does not outperform all baselines on every mask type (e.g., BLD is better on Wide and Narrow masks for some metrics).
- Super-resolution masks yield weaker SSIM for TD-Paint on ImageNet1K than CoPaint.
Experimental setup
Datasets: CelebA-HQ ImageNet1K Places2
Benchmarks: CelebA-HQ ImageNet1K Places2
Metrics: LPIPS SSIM
Baselines: RePaint RePaint-20 MCG CoPaint LDM ControlNet BLD Uni-paint PowerPaint
Builds on: RePaint ControlNet
Closest papers by meaning
- RAD: Region-Aware Diffusion Models for Image Inpainting (CVPR 2025 · similarity 0.89)
- Image Inpainting via Tractable Steering of Diffusion Models (ICLR 2024 · similarity 0.89)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.88)
- Video Diffusion Models Are Strong Video Inpainter (AAAI 2025 · similarity 0.87)
- Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps Image Generation (AAAI 2025 · similarity 0.87)
- HD-Painter: High-Resolution and Prompt-Faithful Text-Guided Image Inpainting with Diffusion Models (ICLR 2025 · similarity 0.87)
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.