/TiNO-Edit

TiNO-Edit: Timestep and Noise Optimization for Robust Diffusion-Based Image Editing (CVPR 2024)

Primary LanguagePythonApache License 2.0Apache-2.0

TiNO-Edit: Timestep and Noise Optimization for Robust Diffusion-Based Image Editing (CVPR 2024)

[Arxiv] [Poster] [Youtube]

TiNO-Edit is an image editing algorithm built on top of Stable Diffusion (SD) by optimizing noise and timesteps in the SD latent space.

Capabilities

Figure 1 Figure 2

LatentCLIPvis & LatentVGG

LatentCLIPvis and LatentVGG are the SD latent space equivalents of the CLIP vision model and VGG. To train these models, see latentclip and latentvgg respectively. We will provide an explanation of our code soon. We also provide pretrained checkpoints here.

Method (Code coming soon ...)

Method overview

Pseudocode

BibTeX

@inproceedings{chen2024tino,
  title={TiNO-Edit: Timestep and Noise Optimization for Robust Diffusion-Based Image Editing},
  author={Chen, Sherry X and Vaxman, Yaron and Ben Baruch, Elad and Asulin, David and Moreshet, Aviad and Lien, Kuo-Chin and Sra, Misha and Sen, Pradeep},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={6337--6346},
  year={2024}
}