Zero123++ is an open-source diffusion model from SUDO-AI-3D that turns a single image into a set of consistent multi-view images, a key building block for image-to-3D reconstruction. Given one photo or render of an object, it generates six aligned views with consistent geometry and appearance, which downstream 3D pipelines can use to reconstruct a mesh. The repository, which includes the model code and inference scripts, has around 2,100 GitHub stars; the last code update landed in 2024 with the v1.2 release.
Zero123++ improves on the original Zero123 line by generating all views jointly in a single tiled diffusion pass, which keeps the views mutually consistent instead of generating them independently. Version 1.2 handles camera intrinsics more deliberately, making the model robust to a wider range of input fields of view and croppings, and adds better local conditioning detail so textures survive the view synthesis. The project page links a technical report (arXiv 2310.15110), Hugging Face demo spaces, a Google Colab notebook and a Replicate demo, so you can try the model before installing anything.
Free. The code is released under the Apache-2.0 license; run it on your own GPU or use the free community demos.
3D and AIGC researchers, game and e-commerce prototypers, and anyone building image-to-3D workflows who needs consistent novel views from a single input image.
In practice, Zero123++ is rarely used alone: its six-view output is typically fed into a reconstruction method such as NeuS, InstantMesh or a Gaussian-splatting pipeline to produce a usable 3D asset. Because the model generates all views jointly with fixed relative camera poses, the outputs drop into those pipelines with less alignment work than independently generated views, which is the main reason the model became a standard component of open image-to-3D toolchains.