Reproduce “UniCycleFlow: Bidirectional Unpaired Image Translation with a Shared Rectified Flow” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xianhao Zhou, Jianghao Wu, Shaoting Zhang, Guotai Wang
UniCycleFlow introduces a rectified flow framework for bidirectional unpaired image translation, employing a single time-conditioned velocity field to represent both forward and reverse transformations. The method achieved an average FID of 54.2 with multi-step integration across ten diverse image translation tasks, surpassing previous state-of-the-art approaches.
Key insights
- A single, shared, time-conditioned velocity field can effectively model coherent bidirectional image translation, intrinsically coupling forward and reverse transformations.
- Meaningful trajectories for rectified flow in unpaired settings can be constructed by learning deterministic, source-conditioned endpoints and matching their distributions adversarially.
- Comprehensive regularization at both intermediate states (self-flow matching) and across perceptual feature space (RPV) is crucial for ensuring smooth, consistent, and perceptually high-quality continuous transformations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.