This repository contains the reference source code for the paper ConsisLoRA: Enhancing Content and Style Consistency for LoRA-based Style Transfer.

• 2025/03/24: We release the inference code and LoRA checkpoints (see here).

• Release the inference code.
• Release the training code.

This code was tested with Python 3.11, Pytorch 2.1 and Diffusers 0.31.

git clone https://github.com/000linlin/ConsisLoRA.git
cd consislora

conda create -n consislora python=3.11
conda activate consislora

pip install -r requirements.txt

Waiting for training code release.

• For style transfer, run:

python inference.py \
    --prompt "a [c] in the style of [v]" \
    --content_image_lora_path "path/to/content" \
    --style_image_lora_path "path/to/style" \
    --lora_scaling 1. 1. \
    --guidance_scale 7.5 \
    --output_dir "inference-images" \
    --num_images_per_prompt 1 \
    --num_steps 30 

Note that some additional parameters can be set for two guidance (see Section 4.3 of our paper).

1. --content_guidance_scale, --style_guidance_scale for controlling the strength of two guidance. Turning on the guidance will increase the inference time.
2. --add_positive_content_prompt, --add_negative_content_prompt is positive and negative prompts for content guidance, respectively. e.g, you can set a [c] and a [v] for them.
3. --add_positive_style_prompt, --add_negative_style_prompt is positive and negative prompts for style guidance, respectively. e.g, you can set in the style of [v] and in the style of [c] for them.

• For using the content LoRA separately, run:

python inference.py \
    --prompt "a [c] in pixel art style" \
    --content_image_lora_path "path/to/content" \
    --lora_scaling 1. 0.

• For using the style LoRA separately, run:

python inference.py \
    --prompt "a dog in the style of [v]" \
    --style_image_lora_path "path/to/style" \
    --lora_scaling 0. 1.

See the inference_demo notebook for more details on how to generate stylized images.

For more results, please visit our Project page.

Our code mainly bases on B-LoRA and diffusers. A huge thank you to the authors for their valuable contributions.

If you use this code, please consider citing our paper:

@article{chen2025consislora,
        title={ConsisLoRA: Enhancing Content and Style Consistency for LoRA-based Style Transfer},
        author={Bolin Chen, Baoquan Zhao, Haoran Xie, Yi Cai, Qing Li and Xudong Mao},
        journal={arXiv preprint arXiv:2503.10614},
        year={2025}
}