TripoSplat generates a 3D object from a single image, open source.

TripoSplat uses Density-Sampled Gaussians to generate high-quality open-source 3D assets from a single image, outperforming rivals by over 140 Elo points.

TripoSplat reconstructs a 3D object from a single image. The result takes the form of Gaussian splats, a representation made of a multitude of small particles, already supported by many rendering and game engines, hence its use for creating assets in video games, simulation, or virtual and augmented reality.

The core contribution focuses on density control, meaning the distribution of these particles in space: numerous where the geometry is complex, sparse on simple surfaces. Existing generators poorly managed this distribution adaptively and fixed the number of particles. With Density-Sampled Gaussians (DeG), TripoSplat samples particle positions from a learned density distribution, which it refines based on the rendering error, by densifying poorly reconstructed areas and lightening others.

For the same particle budget, rendering quality improves; for comparable quality, the number of particles decreases. The user can also set this number at the time of generation: light for a background, dense for a central element, or in multiple versions to manage levels of detail. A human preference study places TripoSplat more than 140 Elo points ahead of the best open-source competitor (1137 versus 996).

The weights and inference code are published under an MIT license, in a deliberately lightweight codebase (approximately 2,000 lines, primarily native PyTorch). The tool also benefits from immediate support in ComfyUI, becoming its first officially integrated Gaussian 3D generator.