Pi3X is a feed-forward neural network for visual geometry reconstruction from unordered sets of images. Hosted on Hugging Face by user yyfz233, it serves as an enhanced version of the π³ model presented in the paper with arXiv identifier 2507.13347. The model addresses the challenge of reconstructing geometry without requiring a fixed reference view by using a fully permutation-equivariant architecture that remains robust to the ordering of input images.
It predicts affine-invariant camera poses and scale-invariant local point maps. This design enables state-of-the-art performance in 3D reconstruction tasks according to the model card. The repository provides model weights in Safetensors format along with support for PyTorch model hub integration.
Pi3X focuses on improved flexibility and reconstruction quality, described as delivering smoother results in an engineering update. It belongs to the class of permutation-equivariant models for visual geometry learning. The project includes a dedicated page at yyfz.github.io/pi3, a GitHub repository at github.com/yyfz/Pi3, and a live demo hosted as a Hugging Face Space.
The model is released under the cc-by-nc-4.0 license. It is delivered as downloadable weights on the Hugging Face platform for use in research and development environments that support PyTorch and Safetensors.
Pi3X is an Other AI product. It focuses on reconstructing 3D geometry and camera poses from unordered sets of images without a fixed reference view. It is built as an open-source project for computer vision researchers. Pi3X is open source under the BSD-3-Clause license. The product ships for the web and API.
Behind Pi3X is yyfz233, and the product first shipped in 2025. Development happens publicly on GitHub with 2.1k stars and 2 commits in the last 90 days. Key capabilities include 3D Reconstruction, Camera Pose Estimation, and Permutation Equivariance.
Latest indexed changes and source events
Other apps tracked under the same category.
