Reproduce “One Shot Face Swapping on Megapixels” 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
Yuhao Zhu, Qi Li, Jian Wang, Chengzhong Xu, Zhenan Sun
MegaFS introduces the first method for one-shot face swapping at megapixel resolution (1024x1024), leveraging a modular architecture to achieve high-fidelity identity transfer and attribute preservation. This work provides a new megapixel-level face swapping database to support research in DeepFake detection.
Key insights
- The extended W++ latent space, generated by the Hierarchical Representation Face Encoder (HieRFE), is superior to W+ for comprehensive face representation and high-fidelity reconstruction, especially for difficult cases like glasses and lighting.
- The proposed Face Transfer Module (FTM) effectively achieves a balanced identity transfer and target attribute preservation without explicit feature disentanglement, providing visually pleasing swapped faces.
- A modular, "divide and conquer" architecture leveraging a fixed StyleGAN2 generator enables stable training and efficient GPU memory usage, crucial for scaling generative models to megapixel resolutions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.