Junyao Gao, Yanchen Liu, Yanan Sun‡, Yinhao Tang, Yanhong Zeng, Kai Chen*, Cairong Zhao*
(* corresponding authors, ‡ project leader)
From Tongji University and Shanghai AI lab.
In this paper, we show that, a good style representation is crucial and sufficient for generalized style transfer without test-time tuning.
We achieve this through constructing a style-aware encoder and a well-organized style dataset called StyleGallery.
With dedicated design for style learning, this style-aware encoder is trained to extract expressive style representation with decoupling training strategy, and StyleGallery enables the generalization ability.
We further employ a content-fusion encoder to enhance image-driven style transfer.
We highlight that, our approach, named StyleShot, is simple yet effective in mimicking various desired styles, i.e., 3D, flat, abstract or even fine-grained styles, without test-time tuning. Rigorous experiments validate that, StyleShot achieves superior performance across a wide range of styles compared to existing state-of-the-art methods.
- [2024/8/29] 🔥 Thanks to @neverbiasu's contribution. StyleShot is now available on ComfyUI.
- [2024/7/5] 🔥 We release online demo in HuggingFace.
- [2024/7/3] 🔥 We release StyleShot_lineart, a version taking the lineart of content image as control.
- [2024/7/2] 🔥 We release the paper.
- [2024/7/1] 🔥 We release the code, checkpoint, project page and online demo.
# install styleshot
git clone https://github.com/Jeoyal/StyleShot.git
cd StyleShot
# create conda env
conda create -n styleshot python==3.8
conda activate styleshot
pip install -r requirements.txt
# download the models
git lfs install
git clone https://huggingface.co/Gaojunyao/StyleShot
git clone https://huggingface.co/Gaojunyao/StyleShot_lineart
you can download our pretrained weight from here. To run the demo, you should also download the following models:
For inference, you should download the pretrained weight and prepare your own reference style image or content image.
# run text-driven style transfer demo
python styleshot_text_driven_demo.py --style "{style_image_path}" --prompt "{prompt}" --output "{save_path}"
# run image-driven style transfer demo
python styleshot_image_driven_demo.py --style "{style_image_path}" --content "{content_image_path}" --preprocessor "Contour" --prompt "{prompt}" --output "{save_path}"
# integrate styleshot with controlnet and t2i-adapter
python styleshot_t2i-adapter_demo.py --style "{style_image_path}" --condition "{condtion_image_path}" --prompt "{prompt}" --output "{save_path}"
python styleshot_controlnet_demo.py --style "{style_image_path}" --condition "{condtion_image_path}" --prompt "{prompt}" --output "{save_path}"
- styleshot_text_driven_demo: text-driven style transfer with reference style image and text prompt.
Text-driven style transfer visualization
- styleshot_image_driven_demo: image-driven style transfer with reference style image and content image.
Image style transfer visualization
- styleshot_controlnet_demo, styleshot_t2i-adapter_demo: integration with controlnet and t2i-adapter.
We employ a two-stage training strategy to train our StyleShot for better integration of content and style. For training data, you can use our training dataset StyleGallery or make your own dataset into a json file.
# training stage-1, only training the style component.
accelerate launch --num_processes 8 --multi_gpu --mixed_precision "fp16" \
tutorial_train_styleshot_stage_1.py \
--pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5/" \
--image_encoder_path="{image_encoder_path}" \
--image_json_file="{data.json}" \
--image_root_path="{image_path}" \
--mixed_precision="fp16" \
--resolution=512 \
--train_batch_size=16 \
--dataloader_num_workers=4 \
--learning_rate=1e-04 \
--weight_decay=0.01 \
--output_dir="{output_dir}" \
--save_steps=10000
# training stage-2, only training the content component.
accelerate launch --num_processes 8 --multi_gpu --mixed_precision "fp16" \
tutorial_train_styleshot_stage_2.py \
--pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5/" \
--pretrained_ip_adapter_path="./pretrained_weight/ip.bin" \
--pretrained_style_encoder_path="./pretrained_weight/style_aware_encoder.bin" \
--image_encoder_path="{image_encoder_path}" \
--image_json_file="{data.json}" \
--image_root_path="{image_path}" \
--mixed_precision="fp16" \
--resolution=512 \
--train_batch_size=16 \
--dataloader_num_workers=4 \
--learning_rate=1e-04 \
--weight_decay=0.01 \
--output_dir="{output_dir}" \
--save_steps=10000
We have carefully curated a style-balanced dataset, called StyleGallery, with extensive diverse image styles drawn from publicly available datasets for training our StyleShot. To prepare our dataset StyleGallery, please refer to tutorial, or download json file from here.
To address the lack of a benchmark in reference-based stylized generation, we establish a style evaluation benchmark containing 40 content images and 73 distinct styles across 490 reference images.
This project strives to positively impact the domain of AI-driven image generation. Users are granted the freedom to create images using this tool, but they are expected to comply with local laws and utilize it in a responsible manner. The developers do not assume any responsibility for potential misuse by users.
If you find StyleShot useful for your research and applications, please cite using this BibTeX:
@article{gao2024styleshot,
title={StyleShot: A Snapshot on Any Style},
author={Junyao, Gao and Yanchen, Liu and Yanan, Sun and Yinhao, Tang and Yanhong, Zeng and Kai, Chen and Cairong, Zhao},
booktitle={arXiv preprint arxiv:2407.01414},
year={2024}
}The code is built upon IP-Adapter.