AsyncDSB: Schedule-Asynchronous Diffusion Schrödinger Bridge for Image Inpainting

• DOI: 10.1609/aaai.v39i3.32349
• Publication Date: 2025-04-11T11:59:41Z
• Authors: Zihao Han; Baoquan Zhang; Lisai Zhang; Shanshan Feng; Kenghong Lin; Guotao Liang; Yunming Ye; Joeq; Kolaye
• Keywords: Inpainting; Bridge (graph theory)

Image inpainting is an important image generation task, which aims to restore corrupted from partial visible area. Recently, diffusion Schrödinger bridge methods effectively tackle this task by modeling the translation between and target images as a process along noising schedule path. Although these have shown superior performance, in paper, we find that 1) existing suffer schedule-restoration mismatching issue, i.e., theoretical practical restoration processes usually exist large discrepancy, theoretically results not fully leveraged for restoring images; 2) key reason causing such issue of all pixels are actually asynchronous but set synchronous noise them, shares same schedule. To end, propose schedule-Asynchronous Diffusion Bridge (AsyncDSB) inpainting. Our insight preferentially scheduling with high frequency (i.e., gradients) then low small gradients). Based on insight, given image, first train network predict its gradient map Then, regard predicted prior design simple yet effective pixel-asynchronous strategy enhance bridge. Thanks at pixels, temporal interdependence can be characterized high-quality Experiments real-world datasets show our AsyncDSB achieves especially FID around 3% ∼ 14% improvement over state-of-the-art baseline methods.