[ECCV2024] TinyBeauty: Toward Tiny and High-quality Facial Makeup with Data Amplify Learning

Official PyTorch implementation of Toward Tiny and High-quality Facial Makeup with Data Amplify Learning [ECCV2024].

Our TinyBeauty effectively synthesizes stunning makeup styles with consistent content, enabling seamless video application.

Release

• [2024/07/17] 🔥 We release the Pretrained Lightweight Model and Demo for quick test.
• [2024/06/02] 🔥 We release the Source Code.
• [2024/04/03] 🔥 We launch the Project Page.
• [2024/03/22] 🔥 We release the Arxiv.

Usage

Demo

First you should make sure you have installed gradio.

cd Demo
python gradio_test.py 

Prepare data

Download sample image pair and makeup style template from here, and place it in the ./data folder.

Finetune

python SD_finetune.py 
    -m runwayml/stable-diffusion-v1-5
    -e h94/IP-Adapter
    -nonmakeup data/Finetune_Data/train
    -makeup data/Finetune_Data/train_purple
    -o "$LORA_MODEL_SAVE_PATH"

Inference

python SD_inference.py
    -m "$LORA_MODEL_SAVE_PATH"
    -s data/Finetune_Data/purple.png
    -d data/Finetune_Data/test
    -o res/test1

More Visual Results

High-quality Facial Makeup

Our DDA generates consistent makeup styles while retain the facial content and identity of the original image

Facial makeup results on high-resolution (1024*1024) images.

Comparison with Previous Works

Visual comparison of TineBeauty and competing methods on the FFHQ Dataset.

Visual comparison of TineBeauty and competing methods on the MT Dataset.

Visual comparison of TineBeauty and BeautyREC on challenging out-of-distribution examples

Cite

If you find TinyBeauty useful for your research and applications, please cite us using this BibTeX:

@misc{jin2024tiny,
      title={Toward Tiny and High-quality Facial Makeup with Data Amplify Learning}, 
      author={Qiaoqiao Jin and Xuanhong Chen and Meiguang Jin and Ying Cheng and Rui Shi and Yucheng Zheng and Yupeng Zhu and Bingbing Ni},
      year={2024},
      eprint={2403.15033},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}