Computer Vision , AI

[One-page summary] ObjectStitch: Generative Object Compositing (CVPR 2023) by Song et al. 본문

Paper_review[short]

[One-page summary] ObjectStitch: Generative Object Compositing (CVPR 2023) by Song et al.

Elune001 2024. 1. 16. 01:31

● Summary: a single framework for Image composition (color harmonization, geometric correction, shadow generation) with no label

 

● Approach highlight

  • Self-supervised learning: segment object from the original image and mask that portion
  • Content adaptor for object identity: image-to-text embedding using CLIP embedding (to use a diffusion model designed for text embedding)

  • Diffusion with the masked image

● Main Results

 

● Discussion

  • Wouldn't artifacts appear at the boundary of the Mask area?
  • How would you generate shadows outside the Mask area?