Latent Schrodinger bridge for fast image-to-image learning

Summarise with VIRAT
Authors: Jerald Wee Hao Tan
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Image-to-image translation has traditionally been framed as the challenge of mapping a source distribution to a target distribution while maintaining structural consistency. While Denoising Diffusion Probabilistic Models (DDPMs) have recently defined the state-of-the-art in generative realism, they remain constrained by an "information destruction" paradigm. In tasks like colourization, standard diffusion models unnecessarily degrade the source image into Gaussian noise, requiring hundreds of computationally expensive sampling steps to reconstruct the original signal with added chromatic information. This thesis proposes a Latent Schrödinger Bridge (LSB) framework designed for fast, high-fidelity image-to-image learning, with a primary focus on automated colourization. Unlike standard diffusion, the Schrödinger Bridge identifies the optimal stochastic transport path directly between the grayscale and color distributions, eliminating the need for a pure-noise intermediate state. By operating within a compressed, low-dimensional latent manifold, the proposed model significantly reduces the "curse of dimensionality" and memory overhead associated with pixel-space models. Experimental results on the Imagenet dataset demonstrate that the LSB frameowrk achieves perceptual quality close to standard diffusion models like Palette, while only requiring 1 epoch of training. This research confirms that by optimizing the transport path and the underlying latent manifold, professional-grade image restoration becomes feasible on consumer-grade hardware.

Fields of Study & Subjects

Generative Adversarial Networks and Image Synthesis Image Enhancement Techniques Image and Signal Denoising Methods

Publication details

Journal / source
DR-NTU (Nanyang Technological University)
Publication date
1 January 2026
Language
EN