MyTimeMachine: Personalized Facial Age Transformation

Luchao Qi, Jiaye Wu, Bang Gong, Annie N. Wang, David W. Jacobs, Roni Sengupta

TL;DR

MyTimeMachine (MyTM) is a novel framework for personalized facial age transformation that moves beyond "generic" aging. By augmenting a global aging prior with a small collection of an individual's personal photos (as few as 10–50), the system generates re-aged images that maintain high identity fidelity and temporal consistency, outperforming state-of-the-art global and diffusion-based models.

What Problem Does This Paper Address?

Facial aging is a highly individualized process influenced by genetics, ethnicity, and lifestyle. While existing "global" aging models (like SAM or CUSP) can produce realistic-looking older or younger faces, they often fail to capture the specific way you will age. They tend to produce an "average" version of an aged face rather than a personalized one.

The challenge in personalizing these models lies in data scarcity. Most people don't have thousands of high-quality photos spanning 80 years. Naive attempts to fine-tune global models on small personal photo collections (which often have biased poses or lighting) usually lead to catastrophic overfitting, where the model loses its ability to generalize to new poses or ages outside the training set.

Methods

MyTimeMachine builds upon the SAM (Style-based Age Manipulation) architecture, which uses a pre-trained StyleGAN2 backbone. The core innovation is a lightweight Adapter Network that sits between the global age encoder and the GAN decoder.

Instead of replacing the global aging knowledge, the Adapter Network learns to produce a "personalized offset" ($\Delta W^+_{tgt}$) in the StyleGAN latent space. This offset nudges the global aging trajectory to match the individual’s unique characteristics. To train this adapter effectively on limited data, the authors introduce three specific loss functions:

  1. Personalized Aging Loss: Uses a facial recognition network (ArcFace) to ensure the generated face's identity features match reference images of the person at a similar target age.
  2. Extrapolation Regularization: Prevents the model from "breaking" when asked to predict ages far outside the training range (e.g., predicting age 10 when the training data only starts at age 30) by encouraging the output to stay close to the reliable global SAM output.
  3. Adaptive w-norm Regularization: Manages the "inversion-editability trade-off." When the target age is close to the input age, it prioritizes reconstruction accuracy; as the target age diverges, it allows the latent code to move toward the center of the distribution to facilitate significant structural changes (like bone growth or skin sagging).

For video, the authors avoid the flickering common in per-frame processing by re-aging a single keyframe and using a face-swapping pipeline (Inswapper) to propagate the re-aged face across the rest of the sequence.

Datasets

The authors curated a new longitudinal facial aging dataset featuring 12 celebrities (e.g., Al Pacino, Oprah Winfrey, Robert De Niro) with roughly 50 images each, spanning 20–40 years. They also evaluated the model on non-celebrity "in-the-wild" data sourced from YouTube "age transformation" time-lapse videos. The global prior was pre-trained on the standard FFHQ and FFHQ-Aging datasets.

Key Insights and Ideas

  • The "Nudge" vs. "Overhaul" Strategy: The most significant insight is that you shouldn't retrain the whole model for personalization. By using an Adapter Network to provide a latent offset, the model retains the robust "structural" aging knowledge of the global prior while layering on the "textural" and "identity" specifics of the individual.
  • Managing the Latent Space: The authors recognized that as you age a face significantly, you must move the latent code closer to the "average" face ($W$) to allow for the necessary deformations. Staying too close to the original image's specific latent code (high inversion accuracy) prevents the model from making the dramatic changes required for 40+ year age gaps.
  • Experience Replay for Aging: Using the global model as a "teacher" during extrapolation (ages not seen in personal photos) acts as a vital anchor, preventing the personalized model from producing "frozen" or nonsensical results at extreme ages.

Key Results

MyTM demonstrated superior performance in both age regression (de-aging) and age progression (aging):

  • Identity Preservation: In age regression tasks, MyTM achieved an 11.7% improvement in identity similarity ($ID_{sim}$) over the previous best-performing model, FADING.
  • Age Accuracy: The model achieved a significantly lower Mean Absolute Error ($Age_{MAE}$) compared to personalized diffusion baselines (e.g., 7.7 vs. 25.9 for FADING + Dreambooth), proving it doesn't just overfit to the training images.
  • User Preference: In pairwise human evaluations, users preferred MyTM results over SAM, AgeTransGAN, and FADING across all tasks, citing better resemblance to the actual person at the target age.
  • Efficiency: The video pipeline completes in under 5 minutes on a single A6000 GPU, whereas competing methods like STIT can take over 3 hours.

Limitations and Future Work

  • Accessories and Hair: The model still struggles with accessories like glasses, which aren't handled consistently by the underlying e4e encoder. It also has difficulty transitioning hair to white for very old target ages ($\geq 80$), a common limitation in GAN-based aging.
  • Artifacts: The authors noted occasional "red-eye" artifacts when generating very old faces.
  • Ethical Concerns: As with any high-fidelity facial manipulation technology, the authors acknowledge the potential for misuse in creating deepfakes and emphasize that the tool is intended for research and creative industries (like VFX).

Future work may involve integrating more sophisticated hair-editing modules and improving the robustness of the latent encoder for diverse accessories.

Summary updated

Abstract

Facial aging is a complex process, highly dependent on multiple factors like gender, ethnicity, lifestyle, etc., making it extremely challenging to learn a global aging prior to predict aging for any individual accurately. Existing techniques often produce realistic and plausible aging results, but the re-aged images often do not resemble the person's appearance at the target age and thus need personalization. In many practical applications of virtual aging, e.g. VFX in movies and TV shows, access to a personal photo collection of the user depicting aging in a small time interval (20$\sim$40 years) is often available. However, naive attempts to personalize global aging techniques on personal photo collections often fail. Thus, we propose MyTimeMachine (MyTM), which combines a global aging prior with a personal photo collection (using as few as 50 images) to learn a personalized age transformation. We introduce a novel Adapter Network that combines personalized aging features with global aging features and generates a re-aged image with StyleGAN2. We also introduce three loss functions to personalize the Adapter Network with personalized aging loss, extrapolation regularization, and adaptive w-norm regularization. Our approach can also be extended to videos, achieving high-quality, identity-preserving, and temporally consistent aging effects that resemble actual appearances at target ages, demonstrating its superiority over state-of-the-art approaches.
BibTeX
@article{qi2025mytimemachine,
  author = {Luchao Qi and Jiaye Wu and Bang Gong and Annie N. Wang and David W. Jacobs and Roni Sengupta},
  title = {{MyTimeMachine: Personalized Facial Age Transformation}},
  journal = {ACM Trans. Graphics},
  year = {2025},
  url = {https://openreview.net/forum?id=6laYmPLb4F},
}

Loading related content...

Figure