Pippo (full name Pippo: High-Resolution Multi-View Humans from a Single Image) is a generative model research project jointly proposed by Meta Reality Labs, the University of Toronto, the Vector Institute, and UC Berkeley, and has been accepted as a Highlight paper at CVPR 2025. The project code is hosted on GitHub (facebookresearch/pippo).
Purpose Pippo aims to solve the problem of generating high-resolution, multi-view human body views from a single casually taken portrait photo. Its core capability is generating 1K-resolution "dense turnaround videos," that is, multi-view results captured around a person for a full circle. It does not require additional inputs, such as a fitted parametric human body model or the camera parameters of the input image, which distinguishes it from many multi-view generation methods that rely on preprocessing steps.
Main capabilities and technical features Pippo is a multi-view diffusion transformer. According to official information, the project provides model training configurations at different resolutions (128, 512, 1024), as well as a tiny model configuration that can be trained on smaller GPUs. In addition, the codebase provides several reusable technical components: ControlMLP for injecting spatial control into the diffusion Transformer, Attention Biasing for running 5x longer sequences during inference, and the Re-projection Error evaluation metric. The official team also plans to update code cleanups for pippo.py and dit.py in the future, as well as inference scripts for pretrained models.
How to use Users can use it by cloning the repository and adding it to the path. For training, example training can be started based on a small sample of the Ava-256 dataset. The samples are provided as npy files and need to be downloaded using a Hugging Face login token. If using a custom dataset, data needs to be prepared in the format of the Ava-256 samples. The generation methods for the more difficult Plücker Rays and Spatial Anchor images are also given in an officially provided gist. In addition, the re-projection error between generated images and realimages can be calculated via the command line.
Applicable scenarios This project is mainly aimed at computer vision and generative AI research, especially researchers and developers in directions such as multi-view human generation, 3D human reconstruction, and diffusion models. Since it provides training configurations and a custom data pipeline, it is also suitable for teams that wish to conduct experiments and further development on this basis.
Limitations It should be noted that the official information does not provide a directly usable inference script for pretrained models (planned to be added in the future), nor does it mention specific information such as model inference speed, VRAM requirements, or commercial licensing. Using a custom dataset requires implementing preprocessing fields such as Plücker Rays and Spatial Anchor yourself. The specific usage conditions of the training configurations and sample data should refer to the LICENSE file.
