Implicit Bridge Consistency Distillation for One-Step Unpaired Image Translation

Suhyeon Lee, Kwanyoung Kim, Jong Chul Ye

Keywords: image translation, consistency distillation, unpaired, one-step, diffusion models

Abstract:

Recently, diffusion models have been extensively studied as powerful generative tools for image translation. However, the existing diffusion model-based image translation approaches often suffer from several limitations: 1) slow inference due to iterative denoising, 2) the necessity for paired training data, or 3) constraints from learning only one-way translation paths. To mitigate these limitations, here we introduce a novel framework, called Implicit Bridge Consistency Distillation (IBCD), that extends consistency distillation with a diffusion implicit bridge model that connects PF-ODE trajectories from any distribution to another one. Moreover, to address the challenges associated with distillation errors from consistency distillation, we introduce two unique improvements: Distribution Matching for Consistency Distillation (DMCD) and distillation-difficulty adaptive weighting method. Experimental results confirm that IBCD for bidirectional translation can achieve state-of-the-art performance on benchmark datasets in just one step generation.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3440