pixel-talk/pear

PEAR is a research project that reconstructs a detailed 3D model of a human body, including the face and hands, from a single photo or video frame. Creating an accurate 3D human mesh, a digital wireframe of a person's body, face, and hands together, has traditionally been slow or required multiple cameras. PEAR does it in real time, at 100 frames per second, from one image, and the accompanying paper is set to appear at SIGGRAPH 2026, a major computer graphics research conference. The way it works is by taking a 2D image and predicting the parameters of standard 3D human body models, statistical models researchers use to represent body shape and pose mathematically. It processes the image so its predictions align tightly with what the pixels actually show, rather than producing a generic 3D figure that may not match the person's real proportions or pose. The result is a 3D mesh that covers the full body, face, and hands at the same time. Setting it up involves cloning the repository, installing Python and PyTorch dependencies, and downloading several separate body and face model files such as SMPL, SMPLX, and FLAME from their own research sites, since these cannot be redistributed directly. Pretrained PEAR model weights download automatically. The project also includes code and a small sample dataset for training the model yourself, though the full training dataset is not yet public. You would use PEAR if you are a computer vision researcher working on avatar creation, motion capture, or virtual try on, or any application that needs to understand the 3D structure of people in images or video. It requires a compatible GPU setup.