FireRedTeam 开源 LayerDiffuse 的 Flux 版实现

This repo is a Flux version implementation of LayerDiffuse (LayerDiffuse).

We train a new transparent vae to adapt to Flux and train a lora to finetune Flux to generate transparent images.

Installation

• Clone this repository.

git clone https://github.com/RedAIGC/Flux-version-LayerDiffuse.git
cd Flux-version-LayerDiffuse
conda create -n flux_layer python=3.10 -y
conda activate flux_layer
pip install torch==2.3.0 torchvision==0.18.0 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

• download weights

huggingface-cli download --resume-download --local-dir ./models RedAIGC/Flux-version-LayerDiffuse 

• Make sure that you have the following directory structure.

Flux-version-LayerDiffuse/
| - imgs
| - lib_layerdiffuse
| - models
    | - layerlora.safetensors
    | - TransparentVAE.pth
| ...

Flux Transparent T2I

demo

python demo_t2i.py --ckpt_path /your/path/to/FLUX.1_dev --prompt /your/prompt

examples

Examples: top to bottom: flux, flux-layer(ours), sdxl, sdxl-layer(LayerDiffuse)

Flux Transparent I2I

python demo_i2i.py --ckpt_path /your/path/to/FLUX.1_dev --image "./imgs/causal_cut.png" --prompt /your/prompt

Prompt: "a handsome man with curly hair, high quality"

Strength: 0.9

Input (Transparent image)	Output (Transparent image)

Acknowledgements

Thanks lllyasviel for their great work LayerDiffuse You can use style RGBA image to finetune the network to get better results.

Contact

If you have any questions about the code, please do not hesitate to contact us!

Email: sunshuang1@xiaohongshu.com, xiangqiang1601@163.com