DriftRect: Selective Trajectory Rectification for Faithful Flow-Based Image Editing

Flow-based image editors can realize strong semantic changes with pretrained generative transport, yet the same editing dynamics may also move content that should remain unchanged. Source-specific inversion provides useful information about the input image, but directly following a reconstruction trajectory can couple source recovery with edit execution and suppress the requested change. We therefore formulate faithful flow editing as selective trajectory rectification: retain the native source–target edit transport as the primary actuator and selectively correct collateral drift around it. We introduce DRIFTRECT, a training-free framework that first partitions the latent into an edit core, a transition/release region, and a protected exterior from source–target responses. It then combines late exterior damping with progressive spatial and channel–time release, while an inversion-derived source reference supplies bounded image-specific geometric and channel cues. Together, these components regulate where, when, and how strongly preservation pressure is applied without constructing a replacement editing direction. On a fixed 28-pair controlled diagnostic set, DRIFTRECT improves four complementary preservation metrics by 2.8–6.7% over FlowEdit; on the full 700-image PIE-Bench at seed 0, the corresponding gains are 2.0–3.5%. A complete second-seed replication retains favorable gains on all four measures after averaging seeds within image. In both settings, the paired CLIP-T interval spans zero. Controlled mechanism studies further show that inversion-derived source information provides a complementary source-specific fidelity margin when routed through the bounded controller.