Splicing ViT Features for Semantic Appearance Transfer (CVPR 2022 - Oral)

[Project Page]

Splice is a method for semantic appearance transfer, as described in Splicing ViT Features for Semantic Appearance Transfer (link to paper).

Given two input images—a source structure image and a target appearance image–our method generates a new image in which the structure of the source image is preserved, while the visual appearance of the target image is transferred in a semantically aware manner. That is, objects in the structure image are “painted” with the visual appearance of semantically related objects in the appearance image. Our method leverages a self-supervised, pre-trained ViT model as an external semantic prior. This allows us to train our generator only on a single input image pair, without any additional information (e.g., segmentation/correspondences), and without adversarial training. Thus, our framework can work across a variety of objects and scenes, and can generate high quality results in high resolution (e.g., HD).

Getting Started

Installation

git clone https://github.com/omerbt/Splice.git
pip install -r requirements.txt

Run examples

Run the following command to start training

python train.py --dataroot datasets/splicing/cows

Intermediate results will be saved to <dataroot>/out/output.png during optimization. The frequency of saving intermediate results is indicated in the save_epoch_freq flag of the configuration.

Sample Results

Citation

@inproceedings{tumanyan2022splicing,
  title={Splicing ViT Features for Semantic Appearance Transfer},
  author={Tumanyan, Narek and Bar-Tal, Omer and Bagon, Shai and Dekel, Tali},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={10748--10757},
  year={2022}
}