Presentation Information
[N-1-38]Analysis of Non-target Region Changes and Internal Feature Updates in Prompt-to-Prompt Color Attribute Editing
◎△Yuto Ishita1, Kenya Jin'no1 (1. Tokyo City Univ.)
Keywords:
Image Generation,Image Editing,Diffusion Models,Prompt-to-Prompt,Cross-Attention
This study analyzes non-target region changes in Prompt-to-Prompt (P2P) color attribute editing and their relation to feature updates in U-Net cross-attention blocks. Using five two-object prompt pairs, only one object's color was edited. The edited object was FG and the non-target region BG. BGChange was computed as the mean L1 difference of normalized RGB values between source and edited images in BG. Varying self/cross replacement steps showed that shorter attention injection increased BGChange. We also compared the baseline with cross_0.0, where source cross-attention maps were not injected. In cross_0.0, BGChange increased by about 22–62% for all prompt pairs. To examine internal behavior, update norm maps were calculated from input-output feature differences in each cross-attention block. These maps were compared with image-change components projected onto the mean RGB change direction in FG by Spearman correlation. The baseline showed higher FG correlation and lower BG correlation. Results suggest that source cross-attention injection strengthens the correspondence between internal updates and the intended FG color-change component, suppressing non-target changes.