ZipSplat

ZipSplat is a research tool from ETH Zurich and Microsoft for reconstructing 3D scenes from ordinary photographs. It uses a technique called 3D Gaussian Splatting, where a scene is represented as a large collection of small, blurry blobs (called Gaussians) that together describe the shape, color, and opacity of everything in view. When rendered from different angles, these blobs produce photorealistic images of the scene.

What makes ZipSplat different from earlier approaches is speed and compactness. Most systems that do 3D reconstruction require either expensive per-scene optimization that takes minutes, or they produce enormous numbers of Gaussians because they assign one to every pixel in the input images. ZipSplat instead compresses the input images into a compact set of internal tokens, each of which decodes into a small group of Gaussians. This happens in a single forward pass, meaning reconstruction completes in under a second. A single compression setting lets you trade between using fewer Gaussians (smaller, faster) and more (higher quality).

You give the model one or more photographs of a scene, and it outputs a 3D representation you can render from any angle, export as a standard file, or view interactively in a browser. The interactive viewer includes a slider for adjusting the compression level in real time so you can see the quality tradeoff visually. The model accepts still images, image folders, or a video clip as input and works even from a single photo.

The project includes code for running inference on your own scenes, evaluating against standard benchmarks, and retraining the model from scratch. Pretrained weights are downloaded automatically from Hugging Face on first use. A CUDA GPU is required because 3D Gaussian rendering depends on GPU-accelerated rasterization.

This is an academic research release accompanying a published paper. The code is written in Python and the repository is from the Computer Vision and Geometry group at ETH Zurich.