Wonder3D

Wonder3D reconstructs a textured 3D mesh from a single image in roughly two to three minutes, and the method behind that speed made it a CVPR 2024 highlight. A cross-domain diffusion model, fine-tuned from Stable Diffusion on renderings from the Objaverse dataset, generates six consistent views of the subject as paired color images and normal maps in a single pass. A domain switcher with cross-domain attention keeps the color and normal predictions aligned with each other, while multi-view attention keeps the six viewpoints consistent, which is exactly where most single-image lifting approaches fall apart. A geometry-aware normal fusion stage then extracts the final mesh from those maps, so the pipeline skips the slow per-shape score distillation optimization used by earlier systems while preserving fine geometric detail. The open source research code runs on PyTorch with pretrained weights hosted on Hugging Face, includes a Gradio interface for interactive use, and performs inference on a single CUDA GPU.