[One-page summary] Zero-shot Image to Image Translation ( arxiv 2023) by Parmar et al.

● Summary: Zero-shot i mage translation using cross-attention map guidance

 

● Approach highlight

• Noise regularization for image inversion: to ensure Gaussian noise

• Cross-attention map guidance: Allows you to edit only the parts you want while maintaining the overall context of the original image

● Main Results

 

● Discussion

• Is it really a zero-shot setup? (using CLIPBLIP)