Wonder3D++: Cross-Domain Diffusion for High-Fidelity 3D Generation From a Single Image

Area of Science:

• Computer Vision
• 3D Reconstruction
• Deep Learning

Background:

• Score Distillation Sampling (SDS) methods offer potential for 3D geometry recovery from 2D priors but are slow and inconsistent.
• Direct 3D generation methods are fast but yield low-quality results lacking detail.

Purpose of the Study:

• To develop an efficient and high-fidelity method for single-view 3D reconstruction.
• To address the limitations of existing 3D generation techniques, focusing on quality, consistency, and speed.

Main Methods:

• Introduced Wonder3D++, a cross-domain diffusion model generating multi-view normal maps and color images.
• Employed a multi-view cross-domain attention mechanism for enhanced generation consistency.
• Developed a cascaded 3D mesh extraction algorithm for rapid, high-quality surface generation.

Main Results:

• Achieved high-fidelity textured mesh generation from single-view images.
• Demonstrated robust generalization capabilities across various inputs.
• Showcased significant efficiency improvements, with mesh extraction in approximately 3 minutes.

Conclusions:

• Wonder3D++ offers a superior approach to single-view 3D reconstruction, balancing quality, consistency, and efficiency.
• The proposed method sets a new benchmark for generating detailed 3D models from limited input data.