Accepted to AI for Content Creation (AI4CC) Workshop at CVPR 2024
Silky Singh, Surgan Jandial, Simra Shahid, Abhinav Java.
Media and Data Science Research (MDSR), Adobe
Project Page: arXiv
Create a conda environment using the provided environment.yml file:
conda env create -f environment.yml
conda activate least
Download SAM's vit-h checkpoint and place it here: segment-anything/checkpoints/sam_vit_h_4b8939.pth exactly following the name convention.
A working notebook is provided here: local_style_transfer.ipynb. To run the notebook using the environment least:
conda install -c anaconda ipykernel
python -m ipykernel install --user --name=least
Given a path to an image and a style description, our method LEAST attempts to constrain the stylization process to the target region in the image, while maintaining the content and structure of the rest of the image.
We collected a set of 25 natural images to perform evaluation of our work against the baselines. The dataset is provided in the dataset directory. Please note that the copyrights exist with the owners of these images.
This repository is heavily based on CLIPstyler, LLaVA and Segment Anything. We thank all the respective authors for open-sourcing their amazing work!
If you find our work useful, please consider citing:
@misc{singh2024least,
title={LEAST: "Local" text-conditioned image style transfer},
author={Silky Singh and Surgan Jandial and Simra Shahid and Abhinav Java},
year={2024},
eprint={2405.16330},
archivePrefix={arXiv},
primaryClass={cs.CV}
}