Idiap/AgeSynth-Latent-Directions

AgeSynth – Latent Age Direction for Aging / De-Aging on StyleGAN 2

A 512-D latent direction (normal vector) in StyleGAN2’s W space that enables smooth, age manipulation of human faces. The vector was obtained via Support-Vector Regression (SVR) trained to predict chronological age from latent codes projected from UTKFace images. Traversing the vector produces synthetic age progression/regression. This latent direction was computed on the official StyleGAN2 1024x1024 model.

Overview

• Training: AgeSynth – Latent Age Direction was trained on UTKFace (age-annotated face dataset)
• Backbone: AgeSynth – Latent Age Direction is adapted from StyleGAN2-FFHQ (official NVIDIA release)
  • Source model: https://github.com/NVlabs/stylegan2
  • License: Nvidia Source Code License-NC
• Parameters: 512 (direction vector) + SVR support vectors (float32)
• Task: Latent-space facial age editing (progression & regression)
• Framework: NumPy

License

Copyright (c) 2025 Idiap Research Institute

CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International) This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. To view a copy of this license, visit http://creativecommons.org/licenses/by-nc/4.0/ or send a letter to Creative Commons, PO Box 1866, Mountain View, CA 94042, USA.

Model output structure:

• File: latent_directions_regression.pkl
• Format: -dict with keys:
  
  • normal – (512,) numpy array, the normalised latent vector.
    
  • svm – trained Support-Vector Regressor predicting chronological age from latent scalars. The vector was obtained through Support-Vector Regression performed on StyleGAN projections of UTKFace images.

Minimal code to instantiate the model and perform inference:

from agesynth.synthetics_subset.latent_edit import LatentEdit
editor = LatentEdit(network_type="stylegan2")
editor.load_covariates_analysis("latent_directions_regression.pkl")
# w_latent: (18,512) StyleGAN2 latent of input image
w_edit = editor.single_sample_age_autmentation(w_latent, age_delta_scalar=15)  # age +15 years

Methodology

1. Data acquisition: A curated subset of the UTKFace dataset was aligned & projected into StyleGAN2 using a gradient-descent projector.
2. Scalar extraction: For each latent, the optimal linear coefficient along every dimension was estimated via SVR to predict the annotated chronological age.
3. Direction selection: The weight vector of the trained SVR yields a single latent direction that, when traversed, monotonically changes the predicted age while minimally affecting identity.

Intended Use

Use with the agesynth Python package:

from agesynth.synthetics_subset.latent_edit import LatentEdit
editor = LatentEdit(network_type="stylegan2")
editor.load_covariates_analysis("latent_directions_regression.pkl")
w_edit = editor.single_sample_age_autmentation(w_latent, age_delta_scalar)

Limitations & Biases

• Inherits demographic biases of UTKFace (age/ethnicity distribution).
• Vector is StyleGAN2-specific; not directly transferable to other GAN backbones.
• Extreme traversals (>50 years) may cause identity drift.

Citation

If you use these models, please cite:

    @InProceedings{luevano2024agesynth,
      title={Identity-Preserving Aging and De-Aging of Faces in the StyleGAN Latent Space},
      author={Luevano, Luis S. and Korshunov, Pavel and Marcel, S{\'e}bastien},
      booktitle = {International Joint Conference on Biometrics (IJCB 2025)},
      year = {2025},
      note = {Accepted for Publication in IJCB2025},
    }
    
    @InProceedings{geissbuhler2024syntheticfacedatasetsgeneration,
      title={Synthetic Face Datasets Generation via Latent Space Exploration from Brownian Identity Diffusion}, 
      author={David Geissb\"uhler and Hatef Otroshi Shahreza and Sébastien Marcel},
      booktitle = {The Forty-second International Conference on Machine Learning (ICML)},
      year =     {2025},
      publisher =    {PMLR},
    }

    @InProceedings{Colbois_IJCB_2021,
        author = {Colbois, Laurent and de Freitas Pereira, Tiago and Marcel, S{\'{e}}bastien},
        projects = {Idiap, Biometrics Center},
        title = {On the use of automatically generated synthetic image datasets for benchmarking face recognition},
        booktitle = {International Joint Conference on Biometrics (IJCB 2021)},
        year = {2021},
        note = {Accepted for Publication in IJCB2021},
        pdf = {http://publications.idiap.ch/downloads/papers/2021/Colbois_IJCB_2021.pdf}
    }

Third-Party Data

This model was derived from the UTKFace dataset. UTKFace images are licensed for non-commercial research purposes only. Any use of this latent direction must therefore remain non-commercial. See https://susanqq.github.io/UTKFace/ for full license terms.