acceptodds
Desk rejected at ICLR 2027

Allocate Where It Matters: State-Adaptive Source Projection for Training-Free Image Editing

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Source-anchored image editors suppress unintended changes by projecting an editing trajectory toward a source reference. Existing methods mainly determine where preservation acts, but say little about how a limited correction budget should be distributed as the editing state evolves. We formulate source preservation as online fixed-budget correction allocation and introduce State-Adaptive Source Projection (ASP), a training-free mechanism that senses source-relative trajectory deviation, reallocates the available correction budget toward locations with greater current need, and feeds the corrected state back into the editing trajectory. Under matched editor loops, source references, supports, and realized budgets, ASP improves preservation by about 0.95 dB over Uniform Projection. Shuffling the same allocation values collapses to Uniform, reversing their state ordering degrades sharply, and freezing the first-step allocation also eliminates the gain, establishing the importance of spatial alignment, allocation direction, and continual state tracking. Integrated into the released SAM-Flow pipeline, ASP reduces Structure Distance by 18.8% and improves unedited-region PSNR by 0.70 dB on the 171 PIE-Bench++ cases with valid keep regions, while CLIP-based edit fidelity changes by at most 1.1%. On the FlowEdit Benchmark, background PSNR improves by 0.70 dB on 276/280 cases and background LPIPS decreases on 279/280 cases. ASP introduces no learned parameters, leaves peak memory unchanged, and adds only 0.61% editing-stage overhead.

Primary area:
generative models
Keywords:
text-guided image editing, training-free image editing, source preservation, inference-time control

Related papers

Discussion (0)

Sign in to comment.