MAD: Makeup All-in-One with Cross-Domain Diffusion Model

A unified cross-domain diffusion model for various makeup tasks

Bo-Kai Ruan, Hong-Han Shuai

• Contact: Bo-Kai Ruan
• arXiv paper | Project Website

🚀 A. Installation

Step 1: Create Environment

• Ubuntu 22.04 with Python ≥ 3.10 (tested with GPU using CUDA 11.8)

conda create --name mad python=3.10 -y
conda activate mad

Step 2: Install Dependencies

conda install pytorch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 pytorch-cuda=11.8 -c pytorch -c nvidia -y
conda install xformers -c xformers -y
pip install -r requirements.txt

# Weights for landmarks
wget http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2
bzip2 -d shape_predictor_68_face_landmarks.dat.bz2 && mkdir weights && mv shape_predictor_68_face_landmarks.dat weights

Step 3: Prepare the Dataset

The following table provides download links for the datasets:

Dataset	Link
MT Dataset	all
BeautyFace Dataset	images, parsing map

We recommend unzipping and placing the datasets in the same folder with the following structure:

📦 data
┣ 📂 mtdataset
┃ ┣ 📂 images
┃ ┃ ┣ 📂 makeup
┃ ┃ ┗ 📂 non-makeup
┃ ┣ 📂 parsing
┃ ┃ ┣ 📂 makeup
┃ ┃ ┗ 📂 non-makeup
┣ 📂 beautyface
┃ ┣ 📂 images
┃ ┗ 📂 parsing
┗ ...

Run misc/convert_beauty_face.py to convert the parsing maps for the BeautyFace dataset:

python misc/convert_beauty_face.py --original data/beautyface/parsing --output data/beautyface/parsing

We also provide the labeling text dataset here.

📦 B. Usage

The pretrained weight is uploaded to Hugging Face 🤗.

B.1 Training a Model

• With our model

# Single GPU
python main.py --config configs/model_256_256.yaml

# Multi-GPU
accelerate launch --multi_gpu --num_processes={NUM_OF_GPU} main.py --config configs/model_256_256.yaml

• With stable diffusion

./script/train_text_to_image.sh

B.2 Beauty Filter or Makeup Removal

To use the beauty filter or perform makeup removal, create a .txt file listing the images. Here's an example:

makeup/xxxx1.jpg
makeup/xxxx2.jpg

Use the source-label and target-label arguments to choose between beauty filtering or makeup removal. 0 is for makeup images and 1 is for non-makeup images.

For makeup removal:

python generate_translation.py \
    --config configs/model_256_256.yaml \
    --save-folder removal_results \
    --source-root data/mtdataset/images \
    --source-list assets/mt_makeup.txt \
    --source-label 0 \
    --target-label 1 \
    --num-process {NUM_PROCESS} \
    --opts MODEL.PRETRAINED Justin900/MAD

B.3 Makeup Transfer

For makeup transfer, prepare two .txt files: one for source images and one for target images. Example:

# File 1           |   # File 2
makeup/xxxx1.jpg   |   non-makeup/xxxx1.jpg
makeup/xxxx2.jpg   |   non-makeup/xxxx2.jpg
...                |   ...

To apply makeup transfer:

python generate_transfer.py \
    --config configs/model_256_256.yaml \
    --save-folder transfer_result \
    --source-root data/mtdataset/images \
    --target-root data/beautyface/images \
    --source-list assets/nomakeup.txt \
    --target-list assets/beauty_makeup.txt \
    --source-label 1 \
    --target-label 0 \
    --num-process {NUM_PROCESS} \
    --inpainting \
    --cam \
    --opts MODEL.PRETRAINED Justin900/MAD

B.4 Text Modification

For text modification, prepare a JSON file:

[
  {"image": "xxx.jpg", "style": "makeup with xxx"}
  ...
]

python generate_text_editing.py \
    --save-folder text_editing_results \
    --source-root data/mtdataset/images \
    --source-list assets/text_editing.json \
    --num-process {NUM_PROCESS} \
    --model-path Justin900/MAD

🎨 C. Web UI (Beta)

Users can start the web UI by add access our designed UI with gradio from localhost:7860:

python app.py

Note: Please put the weights with the following way:

📦 {PROJECT_ROOT}
┣ 📂 makeup_checkpoint.pth  # For our model
┃ 📂 text_checkpoint.pth    # For SD model
┗ ...

Citation

@article{ruan2025mad,
  title={MAD: Makeup All-in-One with Cross-Domain Diffusion Model},
  author={Ruan, Bo-Kai and Shuai, Hong-Han},
  journal={arXiv preprint arXiv:2504.02545},
  year={2025}
}