Reproduce “Revisiting Avatar-As-Image: High-Fidelity Registration is All You Need” 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
Margaret Kostyrko, Yuxuan Xue, Garvita Tiwari, Gerard Pons-Moll
AvaImg is an optimization-based pipeline for converting single-frame clothed human scans into SMPL(-X)+D meshes with aligned UV texture and displacement maps. Across six datasets, it lowers clothing–body penetration while preserving scan fidelity, reaches 2.62 mm bidirectional Chamfer error for surface registration, and can pass its maps through a frozen FLUX VAE with limited degradation, supporting Avatar-As-Image as a source of training-quality 3D data.
Key insights
- Scan proximity alone is insufficient for valid body fitting: combining it with explicit inside/outside reasoning substantially reduces penetration without forcing the estimated body to shrink away from the true shape.
- Low-resolution deformation captures global clothing structure, while subdivision and fourth-power edge coupling improve recovery of fine folds, garment boundaries, and footwear details.
- The resulting UV maps are sufficiently image-like to be encoded and decoded by a frozen FLUX VAE, although this demonstrates representation compatibility rather than a complete diffusion-based avatar generator.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.