Generalizing Consistency Models for Fast Image-to-Image Translation

Sankarshana Venugopal, Mohammad Mostafavi, Jonghyun Choi

Keywords: Consistency Models, Diffusion Bridge Models, Image-to-Image Translation

TL;DR: We propose a new framework for Consistency-based Image-to-Image Translation, improving the quality and efficiency of the generation process.

Abstract:

Diffusion models have shown strong performance in image-to-image (I2I) translation, combining high-fidelity generation with scalability to large-scale datasets.
However, state-of-the-art models like Diffusion Bridge Models (DBMs) suffer from slow sampling speeds, requiring dozens to hundreds of expensive model evaluations. We address this limitation by extending Consistency Models (CMs), originally developed for noise-to-image (N2I) generation, to the I2I setting. We propose Consistency Bridge Models (CBMs), a new framework that enables few-step I2I translation from arbitrary source images without relying on pretrained diffusion models. CBMs inherit the efficiency of CMs while generalizing their theory to arbitrary non-Gaussian prior distributions. Evaluating on multiple datasets and image resolutions, we show that CBMs outperform prior work, reducing forward evaluations by up to 88\%, and improving FID scores by up to 71\%, offering an efficient framework for high-quality I2I translation.

Primary Area: generative models

Submission Number: 1925