UniCycleFlow: Bidirectional Unpaired Image Translation with a Shared Rectified Flow

2026-08-07

AI Summary

This work addresses the absence of shared local transformation rules in unpaired bidirectional image translation by introducing, for the first time, the rectified flow framework to this task. It models the bidirectional mappings as forward and backward integrations of a single time-conditional velocity field, thereby unifying both translation directions through shared continuous dynamics. The method innovatively incorporates source-conditional deterministic endpoints, stop-gradient self-flow matching, discrete cycle closure, and path–velocity regularization, combined with adversarial distribution matching and cycle consistency constraints. Evaluated on ten benchmark tasks, the approach achieves the lowest FID scores on seven using only a single-step Euler integration, yielding an average FID of 55.1—significantly outperforming existing methods.

Abstract

Bidirectional unpaired image translation must preserve source-specific structure while learning coherent transformations in both directions without paired supervision. Existing methods typically employ two direction-specific generators or train separate one-way models. Even when linked by cycle consistency, such models constrain only the round-trip endpoint reconstruction, without requiring the two directions to obey a common local transformation rule. We propose UniCycleFlow, a rectified-flow framework that represents bidirectional translation as forward and reverse integration of a single time-conditioned velocity field. This formulation organizes both directions within the same continuous dynamics, rather than coupling otherwise separate endpoint mappings. A key challenge is that unpaired data provide no meaningful source--target coupling from which rectified-flow trajectories can be constructed. UniCycleFlow addresses this challenge by learning deterministic source-conditioned endpoints whose marginal distributions are adversarially matched to the opposite domains. The resulting paths are regularized by stop-gradient self-flow matching for intermediate velocity supervision, discrete cycle closure for forward--reverse consistency, and representation path-velocity regularization for controlling localized feature changes along the trajectory. Across ten translation directions, UniCycleFlow achieves the lowest FID on 7 of 10 tasks using a single Euler evaluation and obtains the best average FID of 55.1.

Problem

• unpaired image translation
• bidirectional translation
• cycle consistency
• rectified flow
• domain adaptation

Innovation

• rectified flow
• unpaired image translation
• bidirectional generation
• cycle consistency
• velocity field