Single-Step Bidirectional Unpaired Image Translation Using Implicit Bridge Consistency Distillation

TL;DR

Unpaired image-to-image translation often relies on iterative diffusion sampling or adversarial training, which can be costly or unstable. This work introduces IBCD, a single-step bidirectional translator built on a diffusion implicit bridge that connects pre-trained domain models via a unified distillation framework, supplemented by Distribution Matching for Consistency Distillation (DMCD), adaptive weighting, and a cycle-consistency loss. The approach achieves state-of-the-art realism-faithfulness trade-offs in one generation step on toy and real datasets, while maintaining practical inference efficiency. These advances enable robust, low-latency unpaired translations across domains with limited paired data, broadening the applicability of diffusion-based image translation.

Abstract

Unpaired image-to-image translation has seen significant progress since the introduction of CycleGAN. However, methods based on diffusion models or Schrödinger bridges have yet to be widely adopted in real-world applications due to their iterative sampling nature. To address this challenge, we propose a novel framework, Implicit Bridge Consistency Distillation (IBCD), which enables single-step bidirectional unpaired translation without using adversarial loss. IBCD extends consistency distillation by using a diffusion implicit bridge model that connects PF-ODE trajectories between distributions. Additionally, we introduce two key improvements: 1) distribution matching for consistency distillation and 2) adaptive weighting method based on distillation difficulty. Experimental results demonstrate that IBCD achieves state-of-the-art performance on benchmark datasets in a single generation step. Project page available at https://hyn2028.github.io/project_page/IBCD/index.html

Figures (16)

• Figure 1: PSNR-FID trade-off comparison with baselines on the Cat$\rightarrow$Dog task. The size of the marker represents the NFE.
• Figure 2: (a) IBCD performs single-step bi-directional translation using a distillation framework that extends consistency distillation with a diffusion implicit bridge. (b) The IBCD framework bridges two distributions by connecting the PF-ODE paths of two pre-trained diffusion models through bidirectionally extended consistency distillation. To mitigate distillation errors, we introduce distribution matching for consistency distillation and a cycle translation loss.
• Figure 3: (a) Bidirectional translation results on a toy dataset, showing the contributions of each component. (b) Visualization of distillation difficulty $\mathcal{D}(\cdot, c_b)$ and its one-step approximation $\mathbb{E}_t[\Hat{\mathcal{D}}(\cdot, c_b)]$ for A$\rightarrow$B translation, with $g$ selected as a logarithm.
• Figure 4: Qualitative comparison of unpaired image-to-image translation tasks. Compared to other diffusion-based baselines, our model achieves more realistic and source-faithful translations in a single step. The numbers in parentheses represent inference NFE.
• Figure 5: IBCD ablation study results on Cat$\rightarrow$Dog task.