Reproduce “TeCH: Text-guided Reconstruction of Lifelike Clothed Humans” 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

Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Jiaxiang Tang, Deng Cai, Justus Thies

TeCH reconstructs high-fidelity, lifelike 3D clothed human models from a single image by integrating personalized text-to-image diffusion models to infer detailed geometry and consistent textures, particularly for unseen regions.

Key insights

  • Personalized T2I diffusion models can effectively 'imagine' and provide 'multi-view supervision' for unseen 3D regions, compensating for the lack of explicit data in single-image reconstruction.
  • Combining specific semantic prompts (e.g., clothing style, color) with a unique subject-specific identifier token (from DreamBooth) captures both general attributes and fine-grained individual appearance, leading to more lifelike results.
  • A decoupled, two-stage optimization process (geometry followed by texture) on a hybrid DMTet representation prevents gradient entanglement and consistently yields superior 3D reconstruction quality.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.