Wonder3D is an open-source research project that reconstructs a textured 3D object from a single image. It uses a cross-domain diffusion model to generate consistent multi-view images and normal maps of the object, then fuses them into a 3D mesh. Released under the MIT license and first presented at CVPR 2024, the repository has earned more than 5,400 stars on GitHub, with code activity continuing into 2025.
Single-image 3D reconstruction is hard because one photo hides most of the object. Wonder3D approaches it generatively: a diffusion model, fine-tuned to work across both the colour and geometry domains, imagines the object from several viewpoints at once, producing aligned multi-view RGB images and normal maps that stay consistent with each other. A reconstruction step then merges these views into a textured mesh using an optimisation that leverages the normal maps for clean geometry. This cross-domain design was the paper’s key idea โ generating appearance and normals together reduces the inconsistencies that plague pipelines which predict views independently, and it made Wonder3D a widely used baseline in the image-to-3D literature. The repository includes the model code, pretrained weights and instructions for running inference on your own images, and its outputs can be imported into standard 3D tools for games, AR and prototyping.
Free. Wonder3D is open-source research code under the MIT license; you run it on your own hardware at no cost.
Researchers and developers working on 3D generation, and creators who want to prototype 3D assets from a single photo or illustration without manual modelling.