SuperCarver: Texture-Consistent 3D Geometry Super-Resolution for High-Fidelity Surface Detail Generation.

Area of Science:

• Computer Graphics
• Artificial Intelligence
• 3D Modeling

Background:

• Traditional 3D asset creation is labor-intensive, relying heavily on manual sculpting.
• AI advancements enable 3D content generation from prompts, but realistic surface detail synthesis remains challenging.
• Enhancing existing low-quality 3D meshes, rather than generating new ones, is an open research problem.

Purpose of the Study:

• Introduce SuperCarver, a novel pipeline for 3D geometry super-resolution.
• Enable texture-consistent surface detail enhancement for coarse 3D meshes.
• Address the challenge of improving geometry fidelity in existing 3D assets.

Main Methods:

• Render textured meshes from multiple viewpoints.
• Employ a deterministic prior-guided normal diffusion model trained on normal map renderings.
• Utilize a noise-resistant inverse rendering scheme with deformable distance fields for mesh updates.

Main Results:

• SuperCarver generates realistic and expressive surface details consistent with texture appearance.
• The pipeline effectively upgrades historical low-quality 3D assets.
• Demonstrates a significant reduction in the workload for sculpting high-poly meshes.

Conclusions:

• SuperCarver offers a powerful solution for enhancing 3D mesh detail and quality.
• The method successfully bridges the gap between coarse geometry and high-fidelity surface details.
• Presents a viable approach for both asset restoration and efficient high-poly model creation.