Research map / Image generation / Image restoration
AsyncDSB: Schedule-Asynchronous Diffusion Schrödinger Bridge for Image Inpainting
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 AsyncDSB, a pixel-asynchronous noise schedule for diffusion Schrödinger bridge (I²SB) image inpainting, guided by a predicted gradient map so high-gradient pixels are restored earlier.
Problem
Existing diffusion Schrödinger bridge inpainting methods (I²SB) use one synchronous noise schedule for all pixels, causing a mismatch between the theoretical schedule and the actual restoration process.
Approach
Replace the global symmetric noise schedule with a per-pixel schedule β_{t,i,j} driven by a gradient map predicted by an adversarially trained gradient completion network, so high-frequency pixels are scheduled earlier than low-frequency ones.
Key results
- AsyncDSB reaches FID 1.9 on CelebA-HQ center mask vs 2.2 for I²SB (about 14% improvement), with 3–14% FID gains over I²SB across mask types.
- On Places2 center masks, AsyncDSB obtains FID 8.3 vs 8.6 for I²SB (about 4% improvement).
- Adding the gradient map as a UNet condition to I²SB (synchronous schedule) gives weaker results than AsyncDSB, e.g. half mask FID 5.8 vs 5.2 on CelebA-HQ.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the synchronous schedule with a gradient-guided pixel-asynchronous schedule (ablation) | I²SB+gradient · CelebA-HQ center, half, wide, and narrow | FID: 2.0 -> 1.9; 5.8 -> 5.2; 3.0 -> 2.8; 1.7 -> 1.6 | Helped |
| use smaller τmin with τmin < τmax rather than τmin == τmax (ablation) | AsyncDSB · Places2 center and wide masks | performance | Helped |
| set (τmin, τmax) to (0.2T, 0.5T) (ablation) | AsyncDSB · Places2 regular masks (center) | performance | Helped |
| set (τmin, τmax) to (0.001T, 0.4T) (ablation) | AsyncDSB · Places2 irregular masks (wide) | performance | Helped |
| tune τmin and τmax for each mask ratio (ablation) | AsyncDSB · Places2 mask ratios 1%~10%, 11%~20%, 21%~30%, and 31%~40% | FID: 7.6→7.5; 13.1→13.0; 17.6→17.5; 26.2→26.1 | Helped |
| use a pixel-asynchronous schedule instead of a pixel-synchronous schedule | AsyncDSB · CelebA-HQ center, half, wide, and narrow | FID: 2.2 -> 1.9; 5.4 -> 5.2; 2.9 -> 2.8; 1.7 -> 1.6 | Helped |
Surprising findings
- Optimal asynchrony range differs by mask type: large mask ratios favor larger tau_min and tau_max, i.e. earlier schedules.
Limitations
- Uses image gradient (frequency) as the only guidance for the asynchronous schedule; other schedule criteria are not explored.
- Improvements over I²SB are modest (3–14% FID) and IS gains are marginal.
- Mask-and-sample-specific schedules are left for future work; hyperparameters tau_min/tau_max must be tuned per mask setting.
Experimental setup
Datasets: CelebA-HQ Places2
Benchmarks: CelebA-HQ Places2
Metrics: FID IS
Baselines: MISF ZITS StrDiffusion RePaint DiffIR BBDM I²SB
Builds on: I²SB DDPM Schrödinger bridge edge prediction with adversarial learning
Closest papers by meaning
- UniDB: A Unified Diffusion Bridge Framework via Stochastic Optimal Control (ICML 2025 · similarity 0.86)
- RAD: Region-Aware Diffusion Models for Image Inpainting (CVPR 2025 · similarity 0.86)
- TD-Paint: Faster Diffusion Inpainting Through Time-Aware Pixel Conditioning (ICLR 2025 · similarity 0.86)
- Feature out! Let Raw Image as Your Condition for Blind Face Restoration (ICML 2025 · similarity 0.86)
- Image Restoration Through Generalized Ornstein-Uhlenbeck Bridge (ICML 2024 · similarity 0.86)
- Finding Local Diffusion Schrodinger Bridge using Kolmogorov-Arnold Network (CVPR 2025 · similarity 0.85)
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.