Ekta Gavas, Sudipta Banerjee, Chinmay Hegde and Nasir Memon
Official implementation of paper "DiffClean: Diffusion-based Makeup Removal for Accurate Age Estimation".
Accurate age verification can protect underage users from unauthorized access to online platforms and e-commerce sites that provide age-restricted services. However, accurate age estimation can be confounded by several factors, including facial makeup that can induce changes to alter perceived identity and age to fool both humans and machines. In this work, we propose DiffClean which erases makeup traces using a text-guided diffusion model to defend against makeup attacks without requiring any reference image unlike prior work. DiffClean improves age estimation (minor vs. adult accuracy by 5.8%) and face verification (TMR by 5.1% at FMR=0.01%) compared to images with makeup. Our method is: (1) robust across digitally simulated and real-world makeup styles with high visual fidelity, (2) can be easily integrated as a pre-processing module in existing age and identity verification frameworks, and (3) advances the state-of-the art in terms of biometric and perceptual utility.
git clone https://github.com/Ektagavas/DiffClean.gitcd DiffClean
# use anaconda to build environment
conda env create -f requirements.yml
conda activate diffclean
# install packages
pip install git+https://github.com/openai/CLIP.git- The model checkpoints required for the execution of DiffClean can be downloaded from here. Please unzip the file and move the checkpoints to appropriate locations as shown below.
mkdir pretrained
mv model_ir_se50.pth pretrained/
mv shape_predictor_68_face_landmarks.dat pretrained/
mkdir checkpoint
mv diffclean_ssr_age.pth checkpoint/
mv diffclean_clip_age.pth checkpoint/
mv ssrnet_finetuned.pth checkpoint/- Please download the target FR models, MT-dataset and target images here. Unzip the assets.zip file in
DiffClean/assets.
We finetuned our models on MT-dataset (following the same instructions as DiffAM). To further finetune the checkpoint on UTKFace, please download the UTKFace dataset. Generate no-makeup - makeup pairs using EleGANt.
Update the dataset root folder in configs/paths_config.py
Create train/test csvs with the following format and update the paths in datasets/utkface_dataset.py:
sample_train.csv
makeup_name,nomakeup_name,gt_age,makeup_age,nomakeup_age
0001_makeup.jpg,0001_nomakeup.jpg,23,20,24
0002_makeup.jpg,0002_nomakeup.jpg,28,25,28In order to finetune the models on your own dataset, prepare the dataset with makeup/no makeup pairs in above format.
You can run makeup removal on single image or an entire folder with one of the following commands:
# single image
python main.py --edit_one_image_MR --config UTK.yml --exp ./runs/{EXP_NAME} --n_iter 1 --t_0 80 --n_inv_step 40 \
--n_train_step 6 --n_test_step 6 --img_path {IMG_PATH} --model_path checkpoint/{MODEL_NAME} --schedule cosine
# folder with images
python main.py --edit_dir_MR --config UTK.yml --exp ./runs/{EXP_NAME} --n_iter 1 --t_0 80 --n_inv_step 40 \
--n_train_step 6 --n_test_step 6 --img_path {FOLDER_PATH} --model_path checkpoint/{MODEL_NAME} --schedule cosineIMG_PATH: Path to an image to test.FOLDER_PATH: Path to a folder with images.MODEL_NAME: Path to model checkpoint to test
To get age predictions, setup MiVOLO following the instructions in the repo MiVOLO
To obtain baseline results, follow the instructions from respective repos - BeautyGAN, LADN, PSGAN, DiffAM, Clip2Protect, MAD. Please note that for reference-based baselines like BeautyGAN, PSGAN++, LADN, we randomly selected six non-makeup images as references following DeBeauty. Also, we use input image as the target image and 'no makeup' text prompt to generate makeup-removed version with Clip2Protect.
- If you want to fine-tune DiffClean for makeup removal on your dataset, please run the following commands:
SSRNet-Age Loss
python main.py --makeup_removal --config UTK.yml --exp ./runs/diffclean_ssr --do_train 1 --do_test 1 \
--n_train_img 200 --n_test_img 100 --n_iter 5 --t_0 80 --n_inv_step 40 --n_train_step 6 --n_test_step 6 \
--lr_clip_finetune 4e-6 --sch_gamma 1.1 --align_face 1 --schedule cosine --MR_age_loss_w 0.5 \
--model_path checkpoint/diffclean_ssr_age.pthCLIP-Age Loss
python main.py --makeup_removal --config UTK.yml --exp ./runs/diffclean_clip --do_train 1 --do_test 1 \
--n_train_img 200 --n_test_img 100 --n_iter 5 --t_0 80 --n_inv_step 40 --n_train_step 6 --n_test_step 6 \
--lr_clip_finetune 4e-6 --sch_gamma 1.1 --align_face 1 --schedule cosine --MR_age_loss_w 5 \
--model_path checkpoint/diffclean_clip_age.pthAbove commands use the provided finetuned SSR-Net checkpoint. In case you want to fine-tune SSR-Net, please follow the instructions here and replace the SSR-Net checkpoint path in losses/age_loss.py.
To run the demo with the DiffClean checkpoints, run the following commands:
pip install gradio
python app.pyNote: To use age estimation feature in the Gradio UI, MiVOLO must be set up first.
We recognize that our application focus is a sensitive topic, particularly since any modeling or algorithmic errors may adversely impact a vulnerable demographic (minors/teenagers). We strongly advocate for the ethical use of DIFFCLEAN only to assist with facial analytics (not for the purposes of malicious image editing). The authors have not systematically evaluated potential data leakage or memorization issues, including those that may arise from fine-tuned models or downstream applications. This code is provided for academic and research purposes in connection with the paper "DiffClean: Diffusion-based Makeup Removal for Accurate Age Estimation". Commercial use is not intended or supported. Please cite the paper when using this code.
@article{gavas2025diffclean,
title={DiffClean: Diffusion-based Makeup Removal for Accurate Age Estimation},
author={Gavas, Ekta Balkrishna and Banerjee, Sudipta and Hegde, Chinmay and Memon, Nasir},
journal={arXiv preprint arXiv:2507.13292},
year={2025}
}Our code structure is based on DiffAM.
If you have any questions, please contact at eg4131@nyu.edu