Bootstrap Your Generator: Unpaired Visual Editing with Flow Matching

2026-06-02

AI Summary

Existing image and video editing methods rely heavily on large-scale paired data, limiting their generalization in data-scarce scenarios. This work proposes ByG, a framework that trains editing models without any paired data by leveraging semantic instruction cues extracted from a frozen foundational generative model, enforcing cycle-consistency constraints to preserve structural integrity, and introducing a gradient routing mechanism to bridge the gap between training and inference. Relying solely on self-supervised signals, ByG unifies flow matching, unpaired training, and semantic guidance. It substantially outperforms supervised baselines that require millions of paired samples across both image and video editing tasks, with its superiority confirmed through comprehensive user studies and quantitative evaluations.

Abstract

Modern generative models possess a deep understanding of visual content, yet training them for image editing typically requires massive datasets of paired examples. This limits scalability, especially for video editing where collecting paired data is prohibitively expensive. We propose Bootstrap Your Generator (ByG), a general framework for unpaired training of flow matching editing models. It leverages the base model's knowledge without any external signal. Our approach pairs instruction-following cues extracted from the frozen model with cycle-consistency for structure preservation. To make this tractable, we propose to route gradients from downstream losses over clean predictions to noisy training states. We demonstrate state-of-the-art results on challenging data-scarce image and video editing scenarios. Extensive evaluations and user studies show that our method effectively generalizes to unseen domains and outperforms supervised baselines trained on millions of samples. Analysis reveals that our gradient routing bridges the train-inference gap, and extracting semantic cues from a base model provides a robust training signal that obviates the need for external reward models.

Problem

• unpaired visual editing
• flow matching
• video editing
• data scarcity
• image editing

Innovation

• unpaired editing
• flow matching
• gradient routing
• cycle-consistency
• instruction-following generation