Reproduce “Face Time Traveller : Travel Through Ages Without Losing Identity” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Purbayan Kar, Ayush Ghadiya, Vishal Chudasama, Pankaj Wasnik, C.V. Jawahar
Face Time Traveller (FaceTT), developed by Sony Research India and IIIT Hyderabad, introduces a diffusion-based framework for high-fidelity, identity-consistent face aging across wide age ranges. It leverages novel strategies for prompt refinement, latent space inversion, and dynamic attention control, achieving a 15% lower Mean Absolute Error in age accuracy and up to 0.80 Cyclic Identity Similarity compared to prior methods while significantly improving generation speed.
Key insights
- Providing attribute-aware prompts that differentiate between intrinsic and extrinsic aging factors significantly enhances a diffusion model's ability to generate nuanced and context-aware facial age transformations.
- A tuning-free angular inversion method can effectively and efficiently bridge the gap between real images and the diffusion model's latent space, preserving crucial identity details without costly iterative optimization.
- Intelligent, dynamic modulation of attention mechanisms (cross-attention for semantics, self-attention for structure) across denoising phases is crucial for simultaneously achieving pronounced aging effects and strong identity preservation in diffusion-based image editing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.