Top: Supervised training for image editing. Given a source image x, target image y, and editing instruction c, the target is noised and fed to the network along with x and c.
Bottom: We finetune a pretrained text-to-image model into an editing model without paired supervision. Given source x and instruction c, a frozen EMA copy generates a noisy pseudo-target via multi-step sampling. The trainable model then predicts the edit, supervised by: (1) a prior loss aligning the edit direction with the base T2I model, and (2) a cycle loss reconstructing x from the edit using the reverse instruction.
Our framework naturally extends to video editing by applying it to a text-to-video model (Wan2.2). Despite using no paired video data, our method significantly outperforms Ditto, a supervised baseline trained on one million video editing pairs, achieving 70.0% win rate for cartoon targets and 80.5% for photo-realistic, averaging 75.3% overall. On out-of-distribution 3D-CGI inputs, it wins 85% of comparisons.
User study results: our method is preferred in both editing directions.
We evaluate on six unusual stylization targets not represented in common editing benchmarks: GTA V, Minecraft, American comic, low-poly 3D, voxel, and Lego. Our method is not trained on any of these styles, yet it outperforms supervised baselines (FLUX-Kontext, Qwen-Image-Edit) trained on millions of paired edits.
On the GEdit-Bench general editing benchmark, our unpaired method is competitive with FLUX-Kontext across most categories, notably outperforming it on motion changes, human-centric edits, and style changes — categories where paired supervision may be limited.
Each component of our framework is critical. Removing gradient routing, cycle loss leads to stronger edits but degrades source preservation. Without bootstrapping, edits produce artifacts. Without the prior loss, the model collapses to identity mapping.
A key insight: one-step predictions used for cycle consistency are blurry and miss fine details, creating a train-test mismatch. Our gradient routing (adapted from Straight-Through Estimation) conditions the reverse pass on clean multi-step predictions while routing gradients through the one-step estimate, bridging this gap.
@article{tewel2026byg,
title={Bootstrap Your Generator: Unpaired Visual Editing with Flow Matching},
author={Tewel, Yoad and Atzmon, Yuval and Chechik, Gal and Wolf, Lior},
journal={arXiv preprint},
year={2026}
}