Idiap/AgeSynth-Age-To-Scalar

AgeSynth – Age Scalar ⇄ Chronological Age Regression Models

This directory stores regression models that convert the StyleGAN age scalar obtained when traversing the AgeSynth age-direction in StyleGAN2’s W space into an estimated chronological age (and vice-versa), depending on the original subject's age. The mappings were produced by fitting linear and polynomial regressions on a curated subset of images from the UTKFace dataset, after projecting each image into StyleGAN2 and measuring the scalar value along the learned age direction.

Overview

• Training: AgeSynth – Age Scalar was trained on UTKFace (age-annotated face dataset)
• Backbone: AgeSynth – Age Scalar is adapted from StyleGAN2-FFHQ (official NVIDIA release)
  • Source model: https://github.com/NVlabs/stylegan2
  • License: Nvidia Source Code License-NC
• Parameters:
  • Linear fits: 2 coefficients per age bracket
  • 5-th order polynomial fit: 6 coefficients per age bracket
• Task: Obtain a scalar mapping of Latent-space scalar ⇄ age regression. These models are used extensively in agesynth/age_vector.py from our original code base at https://gitlab.idiap.ch/biometric/code.ijcb2025.agesynth
• 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:

• Files:
  • age_fit_linear_pos.pkl : for linear fittings in the 'aging' direction.
  • age_fit_linear_neg.pkl : for linear fittings in de 'de-aging' direction.
  • age_fit_5_dict.pkl : for polynomial fittings in the 'aging' and 'de-aging' direction.
• Format:
  • dict with keys corresponding to age brackets of the original subject's age, with lower bound inclusive and upper bound exclusive. The available keys are: ['below-8', '8-to-13', '13-to-18', 18-to-25', '25-to-35', '35-to-45','45-to-55', '55-to-65', 'above-65']
  • Each value is a NumPy array containing the polynomial or linear coefficients (highest order first) that map latent scalar → chronological age for that bracket.

Minimal code to instantiate the model and perform inference:

Example for estimating a subject's age from a scalar offset in the latent space

import pickle
import numpy as np

# Load 5th-order polynomial model for the 25-35 age bracket
#original_subject_age = 30

with open("age_fit_5_dict.pkl", "rb") as f:
    model_dict = pickle.load(f)
coeffs = model_dict["25-to-35"]  # numpy array of length 6 (highest order first)

# 
scalar_origin = 0.0 #corresponds to 30 years old in the 25-to-35 age group

# Predict chronological age in years
age_pred = np.polyval(coeffs, scalar_origin)
print(f"Predicted age: {age_pred:.1f} years")

Available Models

File	Model type	Intended direction	Brief description
age_fit_linear_pos.pkl	Ordinary Least Squares – linear	Ageing (+)	Linear mapping for positive age deltas
age_fit_linear_neg.pkl	Ordinary Least Squares – linear	De-ageing (–)	Linear mapping for traversing the age direction in the negative direction
age_fit_5_dict.pkl	5-th order polynomial	Both	Captures non-linear relationship between latent scalar and real age

Training Data & Procedure

• Curated UTKFace image lists (covering the 0–116 year span) were projected into StyleGAN2’s w space.
• The latent scalars were regressed against the chronological ages to obtain the linear and 5-th order polynomial mappings.
• Model selection was validated on a held-out split.

Usage

1. Obtain the StyleGAN age scalar for a face (dot-product of the latent with the AgeSynth direction).
2. Pick the corresponding age bracket (e.g. 25-to-35).
3. Retrieve the coefficients from the dictionary and evaluate the polynomial / linear function with NumPy to get the predicted age in years.

No additional libraries are required beyond NumPy and pickle for loading.

Limitations & Biases

• Trained on UTKFace; inherits demographic biases (age/ethnicity imbalance).
• Accuracy deteriorates outside the 0–116 y range.
• Not intended for biometric age estimation nor decision making.

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, Sébastien},
  booktitle={2025 IEEE International Joint Conference on Biometrics (IJCB)},
  year={2025}
}

Third-Party Data

This model was derived from the UTKFace dataset. UTKFace images are licensed for non-commercial research purposes only. Any use of these regression models must therefore remain non-commercial.

The UTKFace dataset license states (excerpt):

The UTKFace dataset is available for non-commercial research purposes only.
Copyright for the original images belongs to the respective owners.

Full license text available at https://susanqq.github.io/UTKFace/.