Research map / Image generation / Fast diffusion sampling
Diffusion for Natural Image Matting
Yihan Hu, Yiheng Lin, Wei Wang, Yao Zhao, Yunchao Wei, Humphrey Shi
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 UTI self-aligned training to the diffusion matting model (ablation) | DiffMatte ViT-S · Composition-1k | SAD15: 23.68 -> 20.50 | Improved |
| use UTI self-aligned strategy instead of the DDP self-aligned strategy (ablation) | DiffMatte ViT-S · Composition-1k | SAD1: 21.11 -> 20.61 | Improved |
| add diffusion noise rather than iterating one-step ViTMatte with previous prediction as guidance (ablation) | ViTMatte · Composition-1k | SAD10: 24.84 -> 20.52 | Improved |
| decouple lightweight diffusion decoder from encoder (ViTS encoder) | DiffMatte · 2048x2048 inference | inference time per subsequent step: 78% less time, 60% less compute | Improved |
| increase DDIM steps from 1 to 10 (ResNet-34) (ablation) | DiffMatte-Res34 · Composition-1k | SAD: 31.28 -> 29.20 | Improved |
| increase DDIM steps from 1 to 10 (ViT-S) (ablation) | DiffMatte-ViTS · Composition-1k | SAD: 20.61 -> 20.52 | No clear effect |
| increase DDIM steps from 1 to 10 for ViT-B trained on Adobe only (ablation) | DiffMatte-ViTB · Distinctions-646 | SAD: 19.07 -> 19.19 | Hurt |
| train on Distinctions-646 instead of Adobe only (ablation) | DiffMatte-ViTB · AIM-500 | SAD (10 steps): 23.57 -> 16.73 | Improved |
| use linear instead of cosine or sigmoid noise schedule (ablation) | DiffMatte ViT-S · Composition-1k | SAD: 21.58 -> 20.52 (cosine to linear) | Improved |
| set input scaling to 0.2 rather than 1.0 (ablation) | DiffMatte ViT-S · Composition-1k | SAD: 22.19 -> 20.52 | Improved |
| increase diffusion decoder channels Nd from 16 to 32 (ablation) | DiffMatte ViT-S · Composition-1k | SAD: 21.63 -> 20.52 | Improved |
| increase Nd from 32 to 48 (ablation) | DiffMatte ViT-S · Composition-1k | SAD: 20.52 -> 20.29 | No clear effect |
Closest papers by meaning
- Image Matting without Matting-Specific Annotations via Eikonal Fields (NeurIPS 2026 · similarity 0.89)
- SDMatte++: Grafting High-Resolution Diffusion Models for Interactive Matting (TPAMI 2026 · similarity 0.89)
- DRIP: Unleashing Diffusion Priors for Joint Foreground and Alpha Prediction in Image Matting (NeurIPS 2024 · similarity 0.88)
- In-Context Matting (CVPR 2024 · similarity 0.87)
- Patched Denoising Diffusion Models For High-Resolution Image Synthesis (ICLR 2024 · similarity 0.87)
- Training Matting Models Without Alpha Labels (AAAI 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.
