Repeatedly editing an image turns small, almost invisible changes into a compounding problem: every output becomes the next input. The key insight here is that preservation need not reshape the whole generation process—steering the later part of denoising can limit unintended changes while leaving earlier model behavior largely intact.
Key Findings
- Across FLUX, SenseNova, and Qwen model families, the authors report substantially slower drift over long editing sequences, with base-model text-to-image and single-image editing performance largely retained.
- The controlled Add & Preserve benchmark separates preservation from task completion: edits are applied within a marked region while image changes outside it are measured independently. This helps distinguish genuine drift reduction from simply refusing requested edits.
- On-policy training uses the editor’s own denoising rollouts and unchanged source content as a preservation reference, avoiding the need for curated ground-truth editing pairs.
Methodology
A trainable copy of each frozen base model is refined to stay close to the base model’s early denoising dynamics and learn preservation-oriented corrections toward the trajectory tail. The steering onset matters: the reported best preservation often comes from beginning corrections before the very last steps. This is a training-time refinement, not a switch that activates a separate model only during inference.
Who Should Care?
Great fit for researchers building image editors intended for iterative, long-horizon use, where preserving untouched regions matters as much as following each new instruction. Look elsewhere if you need a plug-in inference-time fix, or if your priority is single-turn editing alone: the method requires backbone-specific final-stage training, and its results are tied to the evaluated models and benchmarks.