Visual Concept Translator (VCT) aims to achieve image tranlation with one-shot image guidance. Given only one reference image, VCT can automatically learn its dominant concepts and integrate them into input source image. The following examples show its performance.
For each image group, the upper-left image is the source image, the lower-left image is the reference image, and the right part is the translated image. The VCT can be applied in many general image-to-image and style transfer tasks.
To set up the environment, please run
conda create -n vct python=3.8
conda activate vct
pip install -r requirements.txt
We test our method on both Nvidia A30 and A100 GPU. However, it should work in any GPU with 24G memory.
To use the VCT to image-to-image tasks, please run
accelerate launch main.py \
--concept_image_dir="./examples/concept_image" \
--content_image_dir="./examples/content_image" \
--pretrained_model_name_or_path="/put/your/downloaded/huggingface/model"
--output_image_path="./outputs" \
--initializer_token="girl" \
--max_train_steps=500 \
--concept_embedding_num=3 \
--cross_attention_injection_ratio=0.2 \
--self_attention_injection_ratio=0.9 \
--use_l1
Please put your one-shot concept image into concept_image_dir, and any number of content images into content_image_dir. The translated images will be saved in output_image_path.
To avoid the loading error or repeated downloads, it is recommended to download the pre-trained huggingface model such as
stable-diffusion-v1-5 to local. Then put the downloaded path into pretrained_model_name_or_path.
The initializer_token is used as the beginning of concept embeddings. The max_train_steps defines the training steps. For different concept, the optimal training step is also different, so you can adjust the max_train_steps to generate better results (always between 100 to 1000).
Inspired by prompt-to-prompt, the VCT also applies the self-attention and cross-attention injection. Larger self_attention_injection_ratio or cross_attention_injection_ratio means more source contents preserved and less target concepts transferred. If you think the current results are not desired, please adjust these two parameters to achieve more content preservation or concept translation.
If this code is useful for your work, please cite our paper:
@article{cheng2023general,
title={General Image-to-Image Translation with One-Shot Image Guidance},
author={Cheng, B. and Liu, Z. and Peng, Y. and Lin, Y.},
journal={arXiv preprint arXiv:2307.14352},
year={2023}
}