DreamBarbie: Text to Barbie-Style 3D Avatars

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Generating high-quality, disentangled 3D avatars from text is crucial for digital humans.
• Existing text-driven 3D avatar methods struggle with expressiveness, simulation, and fine-grained control.

Purpose of the Study:

• To propose DreamBarbie, a novel text-driven framework for generating animatable 3D avatars with separable components.
• To achieve the iconic "Barbie doll" aesthetic with simulation-ready garments and accessories.

Main Methods:

• Utilizes G-Shell for uniform modeling of watertight and non-watertight components.
• Employs SDF-based initialization and hole regularization for a 100x speedup and stable open topology.
• Optimizes disentangled 3D representations using specialized expert diffusion models and geometric losses.

Main Results:

• DreamBarbie outperforms existing methods in dressed human and outfit generation.
• Achieves high-fidelity outputs with mitigated geometric artifacts and texture conflicts.
• Demonstrates stable open topology generation without image input.

Conclusions:

• DreamBarbie offers a robust framework for text-driven 3D avatar generation.
• Enables diverse applications including apparel editing, animation, and physical simulation.
• Advances the creation of expressive and simulation-ready digital avatars.