This code is our implementation of the following paper:
Taishi Ito, Yuki Endo, Yoshihiro Kanamori: "SelfAge: Personalized Facial Age Transformation Using Self-reference Images"
Age transformation of facial images is a technique that edits age-related person's appearances while preserving the identity. Existing deep learning-based methods can reproduce natural age transformations; however, they only reproduce averaged transitions and fail to account for individual-specific appearances influenced by their life histories. In this paper, we propose the first diffusion model-based method for personalized age transformation. Our diffusion model takes a facial image and a target age as input and generates an age-edited face image as output. To reflect individual-specific features, we incorporate additional supervision using self-reference images, which are facial images of the same person at different ages. Specifically, we fine-tune a pretrained diffusion model for personalized adaptation using approximately 3 to 5 self-reference images. Additionally, we design an effective prompt to enhance the performance of age editing and identity preservation. Experiments demonstrate that our method achieves superior performance both quantitatively and qualitatively compared to existing methods.
Run the following code to install all conda packages.
conda env create -f environment/selfage_env.yml
Download following pretrained models from the paths and save them to the directory pretrained_model.
| Path | Description |
|---|---|
| DEX age classifier | VGG age classifier from DEX to estimate the age of the input image. Fine-tuned by yuval-alaluf on the FFHQ-Aging dataset. |
| LoRA weights | LoRA weights trained with self-reference images. |
Download following datasets from the paths and unzip them.
| Path | Description |
|---|---|
| Regularization set | Regularization set for use in our training. Originally created by sudban3089 and labeled with DEX age classifier by ourselves. |
| Self-reference and test set | Self-reference images for training and input images for testing. Obtained by preprocessing a part of AgeDB dataset. |
Training personalized model in the paper can be done by running the following command:
accelerate launch scripts/train.py \
--self_ref_data_dir=/path/to/self-ref_images \
--output_dir=/path/to/experiment \
--regularization_dir=/path/to/regularization_set \
--with_prior_preservation \
--prior_loss_weight=1.0 \
--contrast_weight=0.1 \
--instance_prompt="photo of sks person" \
--resolution=512 \
--train_batch_size=4 \
--gradient_accumulation_steps=1 \
--checkpointing_steps=1000 \
--learning_rate=1e-6 \
--lr_scheduler=constant \
--lr_warmup_steps=0 \
--max_train_steps=800 \
--validation_prompt="photo of sks person" \
--validation_epochs=50 \
--train_text_encoder \
--rank=16 \
--num_instance_images=3
Note:
- You can specify the base diffusion model for fine-tuning with
--base_model.
Age editing with personalized model in the paper can be done by running the following command:
python scripts/age_editing.py \
--data_path=/path/to/test_image_dir \
--gender=[female or male] \
--exp_dir=/path/to/experiment \
--personalized_path=/path/to/personalized_lora_model \
--target_age=0,10,20,30,40,50,60,70,80 \
--test_batch_size=4 \
--test_workers=4
Notes:
- You can specify a base diffusion model with
--base_model. Change it to match your personalized model if you chose a different one when you trained yourself. - Add
--side_by_sideto save all aging results for a given input side-by-side. - You don't have to specify gender, but doing so may lead to poorer identity preservation.
Please cite our paper if you find the code useful:
@article{ItoSelfage25,
author = {Taishi Ito and Yuki Endo and Yoshihiro Kanamori},
title = {SelfAge: Personalized Facial Age Transformation Using Self-reference Images},
journal = {},
volume = {},
number = {},
pages = {},
year = {2025},
publisher = {}
}
This code heavily borrows from the FADING and IDP repository.
