← News

ResearchSources: arXiv · 20 Nov 2026

Researchers target where diffusion models inject source features during image edits

An arXiv paper proposes adaptive masks and learned block selection for FLUX-based image editing, aiming to ease the trade-off between preserving a source image and applying the requested edit.

In plain words

Editing models that follow text prompts often copy parts of the original image into the new one, and copying too much or too little hurts the result. The paper's method decides where that copying happens, using masks and a small learned set of model blocks chosen in advance.

What was announced

The authors argue that prompt-driven FLUX editing methods, which inject source features into target-conditioned generation, face an inherent trade-off between source preservation and the desired edit. Their approach controls where injection occurs and which transformer blocks are used. Adaptive masks spatially gate feature injection based on differences between source- and target-conditioned block outputs at matched inversion states, while REINFORCE selects a Top-K set of injection blocks under preservation or editability objectives; the chosen blocks are fixed at inference and need no online search.

Reported results are limited in scope. On PIE-Bench with FLUX.1-dev, the authors say spatial masking controls the alignment-preservation trade-off. On 60 validation examples, a learned nine-block set achieved lower whole-image LPIPS than the 18-block RF-Edit set within the same masked pipeline, with slightly lower alignment. Among 50 random nine-block sets it had the lowest LPIPS but not the highest reward, and separate sweeps found different block preferences for preservation and alignment.

What it does not show

  • Evidence is limited to the paper's abstract; the full article was not reviewed, and the reported comparisons come from the authors' own evaluation on small validation sets (60 examples, 50 random sets) rather than independent confirmation.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.