PaperlayerView exact revision on arXiv

Source-linked AI summary

Lifespan Age Transformation Synthesis

Roy Or-El, Soumyadip Sengupta, Ohad Fried, Eli Shechtman, Ira Kemelmacher-Shlizerman

arXiv:2003.09764v2cs.CV

TL;DR

Existing aging methods often change texture without capturing head-shape transformations, making full-lifespan aging a challenging problem. The paper uses a multi-domain GAN and fixed age anchors to model age transformation, and reports stronger evaluations than several comparison methods.

  • Problem

    Many aging methods focus on texture rather than head-shape change, while full-lifespan transformation requires handling pronounced changes across large age gaps.

    Source: §1 · p. 1

  • Method

    A multi-domain GAN separates identity features from age codes and uses fixed age classes as anchors to approximate continuous age transformation.

    Source: §1 · p. 2

  • Results

    Qualitative and quantitative evaluations show the method outperforms state-of-the-art aging algorithms, multi-domain transfer, and latent-space traversal methods on face aging.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The framework produces age transformations for ages 0–70 and can generate ages not seen in training through latent-space interpolation.

    Source: §6 · p. 14

    Generalization is limited for extreme poses, removing glasses or thick beards during rejuvenation, and handling occluded faces.

    Source: §5.3 · p. 14

Abstract

from arXiv · show

We address the problem of single photo age progression and regression-the prediction of how a person might look in the future, or how they looked in the past. Most existing aging methods are limited to changing the texture, overlooking transformations in head shape that occur during the human aging and growth process. This limits the applicability of previous methods to aging of adults to slightly older adults, and application of those methods to photos of children does not produce quality results. We propose a novel multi-domain image-to-image generative adversarial network architecture, whose learned latent space models a continuous bi-directional aging process. The network is trained on the FFHQ dataset, which we labeled for ages, gender, and semantic segmentation. Fixed age classes are used as anchors to approximate continuous age transformation. Our framework can predict a full head portrait for ages 0-70 from a single photo, modifying both texture and shape of the head. We demonstrate results on a wide variety of photos and datasets, and show significant improvement over the state of the art.

1 Introduction

Existing aging methods often change facial texture without modeling head-shape changes, limiting full-lifespan transformations, especially across childhood. The paper proposes shape-and-texture age synthesis using a multi-domain GAN and fixed age classes as anchors.

1 Introduction

  • Large age gaps such as ages 1 to 15 pose a challenge because they involve pronounced changes in both head shape and facial texture.The paper frames learning shape and texture together as an open problem and aims for transformations across ages 0–70 from one portrait.Source: §1 · p. 1
  • The method approximates continuous lifespan aging with six fixed anchor classes spanning childhood, youth, adulthood, and ages 50–69.The classes are designed to capture geometric transformation while covering the full age span in latent space.Source: §1 · pp. 2–3
  • A multi-domain conditional GAN separates identity features from age codes, maps age codes into a shared latent space, and combines them to transform shape and texture.The paper describes the learned latent space as approximating continuous age transformations.Source: §1 · p. 3
  • Qualitative and quantitative evaluations show the method outperforms state-of-the-art aging algorithms, multi-domain transfer, and latent-space traversal methods on face aging.The paper identifies shape-and-texture lifespan transformations and a labeled FFHQ dataset as additional contributions.Source: §1 · p. 3
  • The authors identify potential ethical issues and bias as concerns and refer readers to a supplementary ethics-and-bias statement.Source: §1 · p. 3

2 Related Work

Prior facial-aging research spans handcrafted transformations and deep-learning methods, while many approaches focus on adults or treat age groups as distinct domains. The paper positions its approach as addressing correlated age domains and full-lifespan transformation.

2 Related Work

  • Earlier work used separate models for aging effects, average-face transitions, wavelets, dictionary learning, factor analysis, and face-parameter fitting.A child-aging approach modeled progression as cascaded flows between mean faces of age clusters.Source: §2 · p. 4
  • Deep-learning aging methods include recurrent, autoencoder-GAN, probabilistic, and other GAN approaches, but most focus on adult-to-elderly texture changes.The authors characterize their work as the first to propose full-lifespan aging from ages 0–70.Source: §2 · p. 4
  • Disentangled translation methods can be hard to scale or may transfer non-age attributes, while StarGAN- and STGAN-style methods assume distinct domains unlike correlated age groups.The authors say their architecture enables translation between correlated domains and a continuous traversable age latent space while maintaining identity and image quality.Source: §2 · pp. 4–5

3 Algorithm

The algorithm approximates continuous age transformation with anchor classes, separates age and identity representations, and trains a conditional generator with reconstruction, cycle, identity, age, and adversarial signals.

3.1 Overview

  • The method approximates continuous aging with six anchor classes and uses a conditional generator with separate identity and age paths.For ages between anchors, it linearly interpolates the neighboring age latent codes.Source: §3.1 · p. 5

3.2 Framework

  • A single generator combines identity features with age representations mapped into a unified latent space to produce the target-age image.The framework preprocesses images by removing background and clothing with semantic masks; its mapping network learns a smooth, interpolable age space.Source: §3.2 · p. 6

3.3 Training Scheme

  • Training samples source and target age clusters and performs target generation, source reconstruction, and a cycle back to the source age.These passes provide signals for adversarial, self-reconstruction, cycle-consistency, identity-feature, and age-vector losses.Source: §3.3 · pp. 7–8
  • The age-vector loss penalizes differences between age-encoder outputs and the sampled source and target age vectors.Source: §3.3 · p. 8

3.4 Implementation details

  • Separate male and female models were trained for 400 epochs, and the authors report that code and pretrained models are publicly available.Training used batches of 12 on four GeForce RTX 2080 Ti GPUs.Source: §3.4 · p. 8

4 Dataset

FFHQ-Aging adds age and gender annotations and other facial attributes to FFHQ images, providing a dataset for training and evaluation of age transformation.

4 Dataset

  • FFHQ-Aging annotates all 70,000 FFHQ images for gender and age cluster using crowd-sourced judgments.The dataset defines ten age clusters spanning 0–2 through 70+ years.Source: §4 · pp. 8–9
  • The dataset includes 19-label face semantic maps, head-pose angles, glasses type, and left- and right-eye occlusion scores.The semantic maps were extracted using a DeepLabV3 network trained on CelebAMask-HQ.Source: §4 · p. 9
  • Training uses six selected age clusters after filtering by annotation confidence, pose, glasses, and eye occlusion.The resulting training set contains 14,232 male and 14,066 female images, with separate test images.Source: §4 · p. 9

5 Evaluation

Across comparisons, the framework is reported to outperform age-transformation and multi-domain baselines on image quality, shape change, and user preference, while FaceApp lacks head-shape changes and extreme-age outputs. It still struggles with extreme poses, glasses, thick beards, and occlusions.

5.1 Comparison with Commercial Apps

5.2 Comparison with Age transformation methods

  • Against PyGAN and S2GAN, our single generator produced more photorealistic outputs across multiple classes and wider age transformations with more pronounced aging features.Source: §5.2 · p. 10
  • Our method outperformed both IPCGAN models in image quality and shape deformation, with sharper, more detailed, realistic outputs across age classes.Source: §5.2 · p. 10, Figure 5 · p. 11
  • 16 out of 20 test photos favored our result over PyGAN; PyGAN preserved identity more often (19 out of 20 cases, compared to 13), but mean age difference was 23.1 years versus 6.9.Source: Table 1 · p. 11, §5.2 · p. 11
  • 120 out of 150 cases favored our results over IPCGAN; its mean age difference was 22.6 years versus 11.3 for ours.Source: Table 2 · p. 12
  • StarGAN and STGAN could not consistently transform head shape across age classes, while our method modeled head growth and produced better-quality, more photorealistic images.Source: §5.2 · p. 12, Figure 6 · p. 12
  • InterFaceGAN traversal could lose identity, gender, or ethnicity, and required different λ values across inputs, unlike our consistent age outputs.Source: Figure 7 · p. 13, §5.2 · p. 14

5.3 Limitations

  • The network struggles with extreme poses, removing glasses or thick beards, and occluded faces; the authors suspect encoder downsampling and latent identity loss contribute.Source: §5.3 · p. 14, Figure 8 · p. 14

6 Conclusions

The framework produces age transformations across ages 0–70, changing both head shape and texture and interpolating to ages unseen during training. The paper also introduces a facial dataset and cautions that outputs are estimates, not ground truth.

6 Conclusions

  • The algorithm produces age transformations for ages 0–70, changing both head shape and texture and interpolating to ages unseen during training.The authors report state-of-the-art results and introduce a facial dataset for vision research.Source: §6 · p. 14

Appendix

  • The appendix includes additional continuous-transformation results, ablations, generalization evaluations, and user studies comparing the method with STGAN and IPCGAN.The listed user studies cover all six trained age clusters.Source: Appendix · p. 18

A.1 Intended Use

  • The method is intended mainly for art and entertainment, while outputs may be inaccurate or biased and should receive trained-professional review rather than be treated as ground truth.The paper also mentions approximating missing people’s appearance as a possible critical use.Source: §A.1 · p. 18
  • The authors prohibit using results as grounds for detention, arrest, or other legal evidence.Source: §A.1 · p. 18

A.2 Algorithm and Data Bias

The authors describe measures intended to preserve identity and reduce dataset bias, while acknowledging a gender-spectrum restriction and the possibility of other unconsidered biases.

A.2 Algorithm and Data Bias

  • The design seeks to preserve identity and minimize dataset-bias influence through local-structure preservation and identity-focused training losses.Source: §A.2 · p. 18
  • Cycle loss encourages reconstruction from aged outputs, while self-reconstruction loss trains the network to reproduce inputs when source and target age classes match.Source: §A.2 · p. 18
  • Because FFHQ has gender imbalance within age classes, the authors train separate male and female models to reduce gender-bias effects in outputs.Source: §A.2 · p. 18
  • The male/female model design restricts the algorithm from simulating aging across the entire gender spectrum, which the authors identify as future work.Source: §A.2 · p. 19
  • The authors acknowledge that the network may still introduce biases they did not anticipate during design.Source: §A.2 · p. 19

B Networks Architecture

The generator combines identity and age encoders, a mapping network, and a decoder, alongside a discriminator. The architecture uses distinct components to represent identity, age, and generated image features.

B Networks Architecture

  • The generator comprises an identity encoder, mapping network, and decoder, with a separate age encoder and discriminator.Source: B Networks Architecture · p. 19
  • The identity encoder uses an initial 7×7 convolution, two stride-2 downsampling convolutions, and four residual blocks to produce identity features.Convolution layers use Pixel-norm and ReLU; the authors found Pixel-norm produced fewer artifacts than Instance-norm.Source: B Networks Architecture · p. 19
  • An eight-layer mapping network maps a 50×n age-code vector to a 256-element age latent code.Source: B Networks Architecture · p. 19
  • The decoder uses six styled convolution blocks and modulated convolutions, with its final 1×1 convolution mapping pixel features to RGB values.The authors report that replacing convolution-plus-AdaIN blocks and using Pixel-norm helped reduce droplet artifacts.Source: B Networks Architecture · p. 19
  • The age encoder converts the input image into a 50×n-channel feature map and globally pools it to produce an age code.It uses a 7×7 convolution, four stride-2 convolutions, and a final 1×1 convolution.Source: B Networks Architecture · p. 19
  • The discriminator follows the StyleGAN architecture, downsampling after every other 3×3 convolution block before producing multiple-class discrimination outputs.Source: B Networks Architecture · p. 20

C FFHQ-Aging Dataset Details

The appendix describes FFHQ-Aging distributions and image alignment, then reports anchor-class ablations and interpolated outputs between age classes.

C FFHQ-Aging Dataset Details

  • The dataset figures show male and female age distributions before cleaning and the number of training images per age class after pruning.Source: C FFHQ-Aging Dataset Details · p. 21
  • Image alignment mirrors and blurs image boundaries, then uses eye and mouth landmarks to select an oriented crop.The crop is enlarged to include the full head and neck, using a y′ scale factor of 4.4 rather than FFHQ’s 3.6.Source: C FFHQ-Aging Dataset Details · p. 21
  • Increasing anchor classes from two to three to six greatly improves high-quality age transformations across the full lifespan.Source: Figure 10 · p. 22

D Continuous Age Transformations

E Ablation Studies

The ablations test how age-class anchors and age-code structure support continuous lifespan transformations, and whether the learned age space permits interpolation.

E Ablation Studies

  • Additional age-class anchors are crucial for reliable and plausible full-lifespan transformations.The study compared models using two or three anchor classes and generated outputs by interpolating missing classes.Source: E Ablation Studies · p. 22
  • Using 50 elements per age class permits ages outside anchors, while added Gaussian noise further improves interpolated image quality.The ablations also identify multiple elements per class and noise as important for continuity and quality.Source: E Ablation Studies · p. 22, Figure 11 · p. 23
  • Similarity between trained 3–6 outputs and outputs interpolated from 0–2 and 7–9 suggests age is spanned quasi-linearly in Wage.The comparison interpolates the 3–6 class as the midpoint between the 0–2 and 7–9 age latent vectors.Source: E Ablation Studies · p. 23, Figure 12 · p. 23

E.1 Generalization Ability

The experiments test age-space interpolation and identity-feature generalization to untrained age classes. Interpolated outputs resemble trained outputs, and the method produces high-quality results for unseen face structures.

E.1 Generalization Ability

  • Similarity between trained 3–6 outputs and outputs interpolated from 0–2 and 7–9 suggests approximate age-space linearity and generation beyond trained classes.The 3–6 class was interpolated as the midpoint of the 0–2 and 7–9 classes.Source: §E.1 · p. 23
  • Figure 13's masking artifacts result from segmentation, not from the method.Source: Figure 13 · p. 24
  • For untrained classes 10–14, 20–29, 40–49, and 70+, the method produces high-quality results for unseen face structures.These are the four remaining untrained classes in FFHQ-Aging.Source: §E.1 · pp. 23–24

E.2 User Studies

User studies compared the method with IPCGAN and STGAN retrained on the authors' data. Users preferred the method overall, while the studies also assessed identity and age accuracy.

E.2 User Studies

  • The user interface assessed age, identity, and overall quality.Source: §E.2 · p. 24, Figure 14 · p. 25
  • Users preferred the method in 95% of cases versus IPCGAN and 82% versus STGAN, both retrained on the authors' data.The reported counts were 142 of 150 versus IPCGAN and 123 of 150 versus STGAN.Source: Table 8 · p. 25, Table 9 · p. 25, §E.2 · p. 25
  • The method preserved identity better than both baselines, with age accuracy extremely close to IPCGAN and better than STGAN.Source: Table 8 · p. 25, Table 9 · p. 25, §E.2 · p. 25
  • The studies used 25 randomly selected photos per age class across six classes, repeated each question three times, and collected 2,250 answers per study.Source: §E.2 · p. 25
Loading 2003.09764v2…