Refinement Is Inherently Editable: Training-Free Prompt-to-Prompt Image Editing with Generative Refinement Network
RefineEdit claims prompt-based edits can be made without training, masks, or attention controls by editing binary image codes during refinement.
The paper frames the problem as preserving untouched image content while making the requested change. Its method starts an editing branch from an intermediate source state, then compares source and edit branch bit probabilities to decide what can change. Bits selected for editing follow the refinement path; the rest copy the evolving source image. On PIE-Bench, the authors report stronger background preservation and higher CLIP scores than the evaluated methods. HF Daily Papers' note
The paper frames the problem as preserving untouched image content while making the requested change. Its method starts an editing branch from an intermediate source state, then compares source and edit branch bit probabilities to decide what can change. Bits selected for editing follow the refinement path; the rest copy the evolving source image. On PIE-Bench, the authors report stronger background preservation and higher CLIP scores than the evaluated methods. HF Daily Papers' note
score 5