Reproduce “IA-FaceS: A Bidirectional Method for Semantic Face Editing” 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
Wenjing Huang, Shikui Tu, Lei Xu
IA-FaceS, developed by researchers at Shanghai Jiao Tong University, introduces a bidirectional deep network that reconciles high-fidelity face reconstruction with flexible, disentangled semantic manipulation. The method achieves superior editing accuracy and identity preservation, notably enabling the independent editing of a single eye without affecting the other or requiring explicit visual guidance.
Key insights
- Separating the latent space into an overall 'face icon' and component-specific embeddings effectively resolves the long-standing trade-off between reconstruction fidelity and manipulation flexibility.
- Localized style modulation via the CAM module allows for unprecedented disentanglement, enabling the first latent space method to semantically edit a single eye independently without external visual guidance.
- Utilizing fixed, coarse masks on intermediate feature tensors for component embedding extraction bypasses the need for real-time semantic segmentation while still achieving component-level disentanglement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.