Faithful Image Editing via Degraded Representations

TMLR Paper8855 Authors

Abstract: Rectified flow and diffusion-based models currently represent the state-of-the-art in image editing, leveraging powerful pre-trained generative priors to produce visually compelling modifications. Despite their impressive capabilities, maintaining faithfulness to the source image -- preserving structure and photometric characteristics while satisfying a target prompt -- remains a persistent challenge in this domain. Direct traversal between source and target distributions in rectified flow frameworks offers a promising direction for improving fidelity. However, identifying trajectories that are both semantically effective and strictly structure-preserving remains an open problem. In this work, we propose an optimization- and inversion-free image editing framework that is, in principle, agnostic to the underlying generative backbone. Our central insight is to operate within a carefully designed degraded representation space that constrains editing trajectories and suppresses unintended collateral modifications to the target. We first establish the existence of such degraded representations for generative-prior-based editing and then develop a principled method to project editing trajectories onto this space. The resulting method, Editing via Degraded Representations (EDR), systematically eliminates unfaithful trajectory deviations while preserving the flexibility required to satisfy the target text prompt. Extensive quantitative and qualitative evaluations demonstrate that EDR achieves precise, high-quality edits with superior fidelity, establishing a new state-of-the-art in faithful image editing. Code will be released upon acceptance.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Alain_Durmus1

Submission Number: 8855