What is MyTimeMachine
MyTimeMachine (MyTM) is an advanced personalized facial age transformation technology that, based on approximately 50 personal photos spanning a time range of 20 to 40 years, trains an adapter network to personalize a pretrained global aging model. MyTimeMachine can achieve high-quality age regression and age progression effects while preserving individual identity features. MyTimeMachine can be extended to video, generating aging effects with high identity preservation and temporal consistency, outperforming the existing state of the art.
Main Features of MyTimeMachine
- Personalized Age Transformation: Performs age transformation on a person's face based on a collection of personal photos, including age regression (de-aging) and age progression (aging).
- Identity Preservation: While performing age transformation, preserves the person's facial identity features, ensuring that the transformed image matches the facial features of the original person.
- High-Quality Image Generation: Based on advanced deep learning technology, generates high-resolution, realistic facial images.
- Video Extension: In addition to processing static images, extends to video, enabling dynamic changes in a person's age in videos while maintaining temporal consistency.
- Adaptability: Adapts to different age ranges, performs well within the age range covered by the training data, and optimizes performance for extrapolation to unseen ages.
Technical Principles of MyTimeMachine
- Adapter Network: Introduces an adapter network that can combine personalized aging features and global aging features, and generates aged images based on StyleGAN2.
- Loss Functions: For the personalized adapter network, introduces three loss functions:
- Personalized Aging Loss: Ensures that the aged image is similar in identity features to reference images of similar age in the personal photo collection.
- Extrapolation Regularization: Controls aging effects beyond the training age range using global priors.
- Adaptive w-norm Regularization: Addresses the inversion-editability trade-off of StyleGAN, ensuring shape and texture aging changes while preserving identity.
- Global Aging Prior: Based on a pretrained global aging model, the model learns the aging patterns of the general population.
- Personal Photo Collection: Users need to provide approximately 50 personal photos, with the photos spanning a certain age range, used to train the adapter network to learn personalized aging features.
- Extension to Video: Based on face swapping technology, applies personalized aging effects to video, generating temporally consistent aging videos.
MyTimeMachine Project Address
- Project Official Website：mytimemachine.github.io
- arXiv Technical Paper：https://arxiv.org/pdf/2411.14521
Application Scenarios of MyTimeMachine
- Film and Television Production: In films and TV series, creates age change effects for characters, such as making actors younger or older, to suit the needs of the plot.
- Advertising and Entertainment: In advertisements, shows the effects of products over time, such as skincare or health products.
- Forensics and Criminal Investigation: Help law enforcement agencies identify and track suspects or missing persons who have not been seen for many years through age-progressed images.
- Historical recreation: Recreate the images of historical figures at different age stages, used in education or historical documentaries.
- Personal entertainment: Users experience what they will look like at a certain age in the future, used for entertainment or special commemorations.