What is sapiens2?

Sapiens2 is a family of vision transformer models from Meta's research team, pretrained on one billion images of humans at resolutions up to 1024 by 768 pixels. The paper behind it was accepted at ICLR 2026, and the models are built to handle several human focused computer vision tasks at once: estimating body pose, segmenting different body parts, predicting surface normals, generating pointmaps, and separating a person from the background, a task known as matting. The project ships several model sizes, ranging from about 0.1 billion parameters up to 5 billion parameters, plus a special 1 billion parameter variant trained at a much higher 4096 by 3072 resolution. Each size trades off compute cost against accuracy, so a user can pick a smaller model for speed or a larger one for better results. Using a pretrained model is meant to be simple: you only need PyTorch and the safetensors library to load one of the standalone backbone files and run a forward pass on an image, without needing the rest of the codebase. For the fuller set of tasks, you clone the repository, install it with pip in editable mode, and download the specific checkpoints you need for pose, segmentation, normals, pointmaps, or matting from a separate model zoo document. Each task has its own documentation covering both running inference and training a model from scratch. The project acknowledges building on other Meta research work including DINOv3, as well as the OpenMMLab and Accelerate open source projects. It requires Python 3.12 or newer and PyTorch 2.7 or newer. The README does not describe the license terms in detail, only pointing to a separate Sapiens2 License file, so anyone wanting to use the models for a specific purpose should read that file directly before relying on them.