Implicit Style-Content Separation using B-LoRA

This repository contains the official implementation of the B-LoRA method, which enables implicit style-content separation of a single input image for various image stylization tasks. B-LoRA leverages the power of Stable Diffusion XL (SDXL) and Low-Rank Adaptation (LoRA) to disentangle the style and content components of an image, facilitating applications such as image style transfer, text-based image stylization, and consistent style generation.

🔧 21.5.2024: Important Update 🔧

There were some issues with the new versions of diffusers and PEFT that caused the fine-tuning process to not converge as quickly as desired. In the meantime, we have uploaded the original training script that we used in the paper.

Please note that we used a previous version of diffusers (0.25.0) and did not use PEFT.

Getting Started

Prerequisites

• Python 3.11.6+
• PyTorch 2.1.1+
• Other dependencies (specified in requirements.txt)

Installation

1. Clone this repository:

   git clone https://github.com/yardenfren1996/B-LoRA.git
   cd B-LoRA
   

2. Install the required dependencies:

   pip install -r requirements.txt
   

   (for windows 10 here)

Usage

1. Training B-LoRAs

   To train the B-LoRAs for a given input image, run:

   accelerate launch train_dreambooth_b-lora_sdxl.py \
    --pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" \
    --instance_data_dir="<path/to/example_images>" \
    --output_dir="<path/to/output_dir>" \
    --instance_prompt="<prompt>" \
    --resolution=1024 \
    --rank=64 \
    --train_batch_size=1 \
    --learning_rate=5e-5 \
    --lr_scheduler="constant" \
    --lr_warmup_steps=0 \
    --max_train_steps=1000 \
    --checkpointing_steps=500 \
    --seed="0" \
    --gradient_checkpointing \
    --use_8bit_adam \
    --mixed_precision="fp16"
   

This will optimize the B-LoRA weights for the content and style and store them in output_dir. Parameters that need to replace instance_data_dir, output_dir, instance_prompt (in our paper we use A [v])

2. Inference

   For image stylization based on a reference style image (1), run:

   python inference.py --prompt="A <c> in <s> style" --content_B_LoRA="<path/to/content_B-LoRA>" --style_B_LoRA="<path/to/style_B-LoRA>" --output_path="<path/to/output_dir>"
   

   This will generate new images with the content of the first B-LoRA and the style of the second B-LoRA. Note that you need to replace c and s in the prompt according to the optimization prompt.

   For text-based image stylization (2), run:

   python inference.py --prompt="A <c> made of gold"" --content_B_LoRA="<path/to/content_B-LoRA>" --output_path="<path/to/output_dir>"
   

   This will generate new images with the content of the given B-LoRA and the style specified by the text prompt.

   For consistent style generation (3), run:

   python inference.py --prompt="A backpack in <s> style" --style_B_LoRA="<path/to/style_B-LoRA>" --output_path="<path/to/output_dir>"
   

   This will generate new images with the specified content and the style of the given B-LoRA.

   Several additional parameters that you can set in the inference.py file include:

   1. --content_alpha, --style_alpha for controlling the strength of the adapters.
   2. --num_images_per_prompt for specifying the number of output images.

   (For a111 and comfy see this issue)

Citation

If you use B-LoRA in your research, please cite the following paper:

@misc{frenkel2024implicit,
      title={Implicit Style-Content Separation using B-LoRA}, 
      author={Yarden Frenkel and Yael Vinker and Ariel Shamir and Daniel Cohen-Or},
      year={2024},
      eprint={2403.14572},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

License

This project is licensed under the MIT License.

Contact

If you have any questions or suggestions, please feel free to open an issue or contact the authors at yardenfren@gmail.com.