Function‐Aware Regional Injection for Training‐Free Image Editing
Abstract Training‐free image editing with MMDiT‐based flow models typically preserves source content by injecting source features into the target generation branch. However, achieving high‐quality text‐guided editing remains challenging because different editing tasks require different balances between content preservation and modification. To address this, we propose Function‐Aware Regional Injection (FARI), a framework that analyzes the functional roles across MMDiT transformer blocks to determine where and how to inject source features. Through analysis, we identify four key functional components: (1) spatial anchoring blocks that preserve object positions and layout, (2) pre‐activation blocks where editing‐related semantic changes begin to emerge, (3) semantic blocks where editing‐related semantic changes develop into explicit content modifications, and (4) late refinement blocks that consolidate visual information into a coherent final output. Based on these insights, FARI applies selective feature transfer in the anchoring and refinement blocks, while incorporating all source features in the pre‐activation and semantic blocks to support larger content changes. This design enables interpretable control over preservation‐modification trade‐offs and provides editing policies tailored to different edit instructions (e.g., changing object, adjusting color). Experiments on a newly curated benchmark and PIE‐Bench demonstrate that FARI achieves favorable trade‐offs across all five edit instruction types by balancing preservation and editability, rather than relying on a single globally optimal injection policy.
Authors
- Xiao‐Ming Fu (ORCID: https://orcid.org/0000-0001-8479-0107)
- Siyuan Xing (ORCID: https://orcid.org/0000-0003-3874-6521)
Institutions
- University of Science and Technology of China (CN)
Publication Details
- Journal
- Computer Graphics Forum
- Published
- 2026-11-13
- DOI
- https://doi.org/10.1111/cgf.70614
- Primary Topic
- Generative Adversarial Networks and Image Synthesis
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- National Natural Science Foundation of China
- Ministry of Education of the People's Republic of China