AsyncDSB: Schedule-Asynchronous Diffusion Schrödinger Bridge for Image Inpainting

Authors: Zihao Han, Baoquan Zhang, Lisai Zhang, Shanshan Feng, Kenghong Lin, Guotao Liang, Yunming Ye, Joeq , Kolaye

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on real-world datasets show that our Async DSB achieves superior performance, especially on FID with around 3% 14% improvement over state-of-the-art baseline methods. Extensive experiments conducted on two datasets (i.e., Celeb A-HQ and Places2) demonstrate the superior performance of the proposed Async DSB method in comparison to various state-of-the-art image inpainting methods.
Researcher Affiliation Collaboration 1 Harbin Institute of Technology, Shenzhen; 2 Shen Zhen Si Far Co., Ltd.; 3 Centre for Frontier AI Research; 4 Shengshu AI. EMAIL, EMAIL, Lisai EMAIL, EMAIL, EMAIL
Pseudocode No The paper describes the methodology using textual explanations and mathematical equations, but it does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement about the release of source code for the described methodology, nor does it include any links to code repositories.
Open Datasets Yes We validate our method and various baselines over two typical dataset: 1) Celeb A-HQ is a face dataset. We split this dataset by following standard dataset split (Liu et al. 2015). 2) Places2 is a scene image dataset, which split into 1.8 million training images and 36500 validation images, from which we randomly select 6000 images as test.
Dataset Splits Yes We split this dataset by following standard dataset split (Liu et al. 2015). Places2 is a scene image dataset, which split into 1.8 million training images and 36500 validation images, from which we randomly select 6000 images as test.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments, such as GPU or CPU models.
Software Dependencies No The paper does not specify any particular software dependencies or their version numbers that are needed to replicate the experiment.
Experiment Setup Yes Following (Liu et al. 2023), T = 1000 is used. where λadv and λF M are hyper-parameters; where τ max and τ min are two hyper-parameters, which controls the asynchronous range of pixel schedule; In Figure 6, we conduct a parameter analysis on Places 2 from τ min = 0.001T to τ min = 0.9T and τ max = 0.001T to τ min = 0.9T.