I$^2$SB: Image-to-Image Schrödinger Bridge

Authors: Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos Theodorou, Weili Nie, Anima Anandkumar

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate I2SB in solving various image restoration tasks, including inpainting, super-resolution, deblurring, and JPEG restoration on Image Net 256 256 and show that I2SB surpasses standard conditional diffusion models with more interpretable generative processes.
Researcher Affiliation Collaboration 1Georgia Institute of Technology 2NVIDIA 3California Institute of Technology.
Pseudocode Yes Algorithm 1 Training
Open Source Code Yes Project page and codes: https://i2sb.github.io/.
Open Datasets Yes on Image Net 256 256 (Deng et al., 2009)
Dataset Splits Yes Following the baselines (Saharia et al., 2022; Song et al., 2022), we report super-resolution results on the full Image Net validation set and report the remaining results on a 10k validation subset.
Hardware Specification No The inference time is measured on a V100 16G. (This is mentioned only for inference time in a specific table, not as the general hardware used for training or all experiments, so it's not a full specification.)
Software Dependencies No Official Pytorch implementation of our I2SB can be found in https://github.com/NVlabs/I2SB. (This refers to the implementation language/framework but doesn't give specific version numbers for dependencies).
Experiment Setup Yes By default, we use 1000 sampling time steps for all tasks with quadratic discretization (Song et al., 2020a).