A Text-to-3D Framework for Joint Generation of CG-Ready Humans and Compatible Garments.

Area of Science:

• Computer Graphics
• Artificial Intelligence
• 3D Modeling

Background:

• Generating 3D human avatars with custom clothing is complex and labor-intensive.
• Existing generative AI for 3D humans and clothing lacks integrated, accessible pipelines for CG-ready assets.
• Current methods struggle to produce outfits that are both visually accurate and compatible with physical simulations.

Purpose of the Study:

• To introduce Tailor, an integrated text-to-3D framework for generating high-fidelity, customizable 3D avatars with simulation-ready garments.
• To provide an accessible and efficient pipeline for creating computer graphics-ready 3D models.
• To enable downstream tasks such as physical simulation with generated avatars and outfits.

Main Methods:

• Utilizing a large language model for semantic parsing of text descriptions into avatar and garment parameters.
• Employing topology-preserving deformation with geometric losses for geometry-aware garment generation aligned to the body.
• Implementing a novel multi-view diffusion process for consistent and detailed garment texturing, supporting symmetry.

Main Results:

• Tailor generates high-fidelity, customizable 3D avatars with simulation-ready garments from text descriptions.
• The framework produces CG-ready models with polygonal meshes and strand representations suitable for conventional CG pipelines.
• Quantitative and qualitative evaluations show Tailor surpasses state-of-the-art methods in fidelity, usability, and diversity.

Conclusions:

• Tailor offers a significant advancement in accessible text-to-3D content creation for virtual humans and fashion.
• The framework successfully bridges the gap between generative AI capabilities and practical CG production needs.
• Tailor demonstrates superior performance in generating diverse, high-quality, and simulation-ready 3D avatars with fitted garments.