Bootstrap Your Generator: Unpaired Visual Editing with Flow Matching

TL;DR

Bootstrap Your Generator (BYG) — an NVIDIA × Tel Aviv method, ICML 2026 — turns a pretrained flow-matching generator into a visual editor for stylization targets it was never shown paired data for. The project page demonstrates zero-shot edits into six unusual styles (GTA V, Minecraft, American comic, low-poly 3D, voxel, Lego) and claims to outperform supervised editors FLUX-Kontext and Qwen-Image-Edit — which were trained on millions of paired edits — despite using no paired training data on these styles [project page §Results]. The recipe is “unpaired editing with flow matching”: exploit the base generator’s own knowledge of the target style to bootstrap a transformation, rather than collect (source, edited) pairs.

Key claims

Visual editing into stylization targets can be performed without paired training data by bootstrapping off a pretrained flow-matching generator’s prior over the target style [project page §Overview].
BYG generalizes to six stylization targets — GTA V, Minecraft, American comic, low-poly 3D, voxel, Lego — none of which the method is trained on, and which are not represented in common editing benchmarks [project page §Results].
On these six unusual targets, BYG outperforms FLUX-Kontext and Qwen-Image-Edit despite those baselines being trained on millions of paired edits [project page §Results].

Method

The project page renders as a thin landing page; the full technical recipe is in the ICML 2026 manuscript. The headline framing is: given a pretrained flow-matching generator that already covers the target style distribution, learn an unpaired editing trajectory in its velocity field that maps source content into the target style while preserving content identity — no source→target paired supervision required. The six demonstrated targets (GTA V, Minecraft, comic, low-poly, voxel, Lego) are picked because they are deliberately outside the distribution of standard editing benchmarks like the ones FLUX-Kontext and Qwen-Image-Edit train against.

Results

Six-style head-to-head against FLUX-Kontext and Qwen-Image-Edit: BYG reported as outperforming both, despite using no paired training data on any of the six styles [project page §Results]. No numerical metrics are surfaced on the landing page; comparisons are shown qualitatively.
Accepted at ICML 2026 [arxiv listing metadata].

Why it’s interesting

The recipe fits the “synthetic / unpaired supervision can beat large paired corpora when the generator’s prior is strong enough” thesis that Synthetic Training Data has been tracking — particularly the L2P: Unlocking Latent Potential for Pixel Generation finding that an LDM-trained generator can transfer to a new objective using only its own samples, with zero paired collection. BYG is the editing-flavored sibling: paired editors like Scaling Instruction-Based Video Editing with a High-Quality Synthetic Dataset (Ditto / Editto) (Ditto/Editto) and Pico-Banana-400K: A Large-Scale Dataset for Text-Guided Image Editing have spent significant data engineering on million-pair corpora, and BYG’s claim is that those corpora may not generalize to styles the corpus doesn’t cover — whereas the generator’s own latent knowledge does. It also fits the Classifier-Free Guidance variants lineage where editing is increasingly recast as a training-time manipulation of a pretrained flow rather than a paired-data fit.