Reproduce “Bootstrap Your Generator: Unpaired Visual Editing with Flow Matching” on OpenResearch

About this paper

Yoad Tewel, Yuval Atzmon, Gal Chechik, Lior Wolf

Bootstrap Your Generator (ByG) introduces an unpaired training framework that enables flow-matching models to perform instruction-based image and video editing without requiring paired data or external reward models. The method achieves superior user preference with a 75.3% win rate in long-tail video editing compared to supervised baselines and demonstrates strong generalization to unseen styles.

Key insights

• Generative models can effectively learn complex visual editing capabilities solely from their intrinsic knowledge of visual content, reducing the need for costly paired datasets or external reward models.
• The combination of self-bootstrapping noisy inputs, semantic guidance from a frozen T2I backbone, and cycle consistency provides a robust framework for unpaired training of flow-matching editing models.
• Gradient routing through Straight-Through Estimation effectively bridges the gap between operating on noisy training states and requiring clean, high-quality images for downstream losses like cycle consistency, improving both conditioning and learning.

Summary by alphaXiv.