Reproduce “Efficient Few-shot Identity Preserving Attribute Editing for 3D-aware Deep Generative Models” 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
Vishal Vinod
Vishal Vinod at UC San Diego introduces an efficient few-shot approach for identity-preserving attribute editing in 3D-aware deep generative models. The method achieves multi-view consistent facial transformations using minimal synthetic data, outperforming 2D editing baselines with a FID score of 39.91 and strong identity preservation.
Key insights
- The strong prior of a pre-trained 3D-aware GAN can map even unrealistic synthetic input images into its realistic latent manifold during inversion, enabling effective few-shot data generation.
- Disentangled attribute edit directions can be precisely estimated in the W+ latent space using only a small number of synthetic examples, allowing for robust sequential edits that preserve identity and 3D consistency.
- Pivotal Tuning Inversion (PTI) significantly enhances the inversion of out-of-distribution images, enabling high-fidelity 3D-aware editing of real-world faces without requiring explicit camera pose input.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.