Decoupling Source Preservation from Target Generation in Next-Scale Autoregressive Image Editing via Fixed-History Contrast

Next-scale autoregressive models have emerged as a new paradigm for visual generation, inspiring growing efforts to extend them to training-free image editing. Effective editing requires preserving source content unrelated to the intended change and generating new content aligned with the target prompt. Existing methods often derive preservation signals from source–target prediction contrasts during autoregressive generation. Because these contrasts are evaluated along an evolving edited history, preservation decisions at later scales may depend on earlier changes, potentially causing unintended changes or insufficient edits. To overcome this limitation, we propose DecoEdit, a training-free editing framework that decouples the determination of source preservation from the evolving target generation process. DecoEdit achieves this through fixed-history contrast, which computes source- and target-conditioned prediction contrasts under a fixed source history obtained from the encoded source image. For source preservation, DecoEdit converts these contrasts into preservation evidence to identify source content to retain, and constructs a preservation bias that controls the strength of source retention. For target generation, DecoEdit uses this evidence to modulate the preservation bias and injects the modulated bias into the evolving target-conditioned process to produce the requested edit. Experiments on Infinity-2B and Infinity-8B demonstrate that DecoEdit achieves a better preservation-editability trade-off than existing methods. The anonymous code is attached in the supplementary material.