Reproduce “Pose Manipulation with Identity Preservation” 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
Andrei-Timotei Ardelean, Lucian Mircea Sasu
Researchers from Transilvania University of Brasov introduce CainGAN, a generative model that synthesizes human face images in novel poses while preserving identity using only a few source images, without requiring per-subject fine-tuning. CainGAN achieves an FID of 24.92 with 8 source images, surpassing baselines like pix2pixHD and FSHM, and demonstrates improved identity preservation and perceptual realism even with a single input image.
Key insights
- Providing target landmark information to the embedder ("Targeted Embedder") facilitates earlier feature alignment and improves performance, especially when multiple source images are available.
- Weighting individual spatial embeddings by a "responsibility" tensor during combination effectively leverages the most certain features from diverse source inputs, enhancing identity capture.
- A gradual increase in the identity discriminator's importance factor during training is crucial for stabilizing the GAN and effectively balancing the learning of realistic pose generation with strong identity preservation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.