hliang2/splatshot

SplatShot is a research tool from Rice University and Samsung Research that turns a single ordinary photo of a person's face into a 3D avatar. You provide one image, and the tool produces a full three-dimensional face model that can be viewed from any angle, including angles that were not visible in the original photo. No lengthy setup or training on that specific person is required. The way it works is through a technique called 3D Gaussian Splatting, which represents a 3D scene as a large collection of tiny overlapping blobs rather than traditional polygons. The tool starts from a base face template, then uses image-generating AI to fill in views of the face from dozens of different angles. Those generated views are combined with the template to produce the final 3D model. An identity-preservation step ensures the generated views stay consistent with the person in the original photo. The output is a .ply file, which is a standard 3D point cloud format that can be opened in free 3D viewers designed for this type of content. The README includes the viewer names you would need to look up. Alongside the 3D model, the tool also saves the intermediate diffusion images it generated per camera angle, so you can inspect the process. The hardware requirement is steep: the authors recommend a GPU with 24 gigabytes of memory, and processing one image takes roughly ten to fifteen minutes. This puts it firmly in the research and professional category rather than casual personal use. The project was published alongside a research paper and a companion dataset called CelebA-3D, which was itself built using SplatShot. The code is available for academic and research use, and the project acknowledges several open-source libraries it builds on.