TripoSplat in the browser
Show HN: TripoSplat Running in the Browser
Last verified:
Visit TripoSplat in the browser
What is TripoSplat in the browser?
TripoSplat in the Browser is a web-based tool that converts a single 2D image into high-quality 3D Gaussian splats. Built from the open-source TripoSplat model developed by TripoAI/VAST AI Research, this browser implementation allows users to generate 3D assets entirely online without installing software or having specialized GPU hardware. The tool uses a novel learned density-control approach called Density-Sampled Gaussians (DeG) to create variable-number 3D Gaussians from one input image.
Key features include the ability to load an image and generate a .splat file, save the generated Gaussian splat for download, and load existing .splat files for viewing. The interface includes advanced options like quantization mix settings, sampling controls, preprocessing options, kernel selection, flash attention support, and GPU IO binding. The entire process runs client-side in the browser using WebGPU technology.
This tool is designed for asset creators, AR/VR developers, game developers, simulation environment builders, 3D artists, and anyone working with 3D Gaussian rendering. Since 3D Gaussian rendering is now supported by mainstream rendering and game engines, TripoSplat serves as a powerful pipeline tool for integrating single-image-to-3D conversion into production workflows without requiring multi-angle photo shoots.
TripoSplat in the browser pricing
Pricing model: Freemium
TripoSplat is completely free and open-source under the MIT license. The model weights and inference code are publicly available on Hugging Face (VAST-AI/TripoSplat) and GitHub (VAST-AI-Research/TripoSplat). The browser demo at splat.sqrm.co appears to be a community implementation with no official pricing tier. Hugging Face Spaces hosting is free with CPU Basic or ZeroGPU options, though GPU hardware for self-hosted Spaces starts at $0.40/hour for Nvidia T4. Users can run the model locally with no cost if they have a compatible NVIDIA GPU (minimum RTX 3060 12GB recommended).
TripoSplat in the browser pros
- Converts single 2D image to 3D Gaussians without multi-angle photo shoot
- Completely free and open-source under MIT license
- Runs entirely in browser with no software installation required
- No server uploads - all processing happens client-side
- Variable number of 3D Gaussians adapts to scene complexity
- Learned density control allocates more Gaussians to complex areas
- Inference-time budget control for custom Gaussian counts
- Higher visual quality than existing open-source 3D Gaussian generators
- 140+ Elo points ahead of next best model in user studies
- Lightweight codebase with only ~2,000 lines of code
- Bypasses heavy external frameworks like transformers or diffusers
- Relies almost entirely on native PyTorch
- Day-0 ComfyUI support for 3D Gaussian generation
- Supports JPG, JPEG, PNG, WEBP, GIF, AVIF input formats
- Can extract multiple LOD versions from same asset instantly
- GPU-accelerated via WebGPU for fast processing
- Save and download .splat files for use in other applications
- Load existing .splat files to view in browser
- Advanced settings including quantization mix and flash attention
- Ideal for stylized assets and artistic content
TripoSplat in the browser cons
- Requires WebGPU-compatible browser for GPU acceleration
- Browser-based processing may be slower than local GPU
- Single image input limits geometric accuracy compared to multi-view capture
- No official paid hosting - users must run their own instance
- Hugging Face Space demo may have queue times or resource limits
- Genetic 3D generation may produce artifacts on complex geometry
- Requires GPU with sufficient VRAM for larger Gaussian counts
- No built-in 3D editor - only generation and viewing
- .splat files may not work in all 3D software without converters
- Learning density control is novel but less tested than traditional split/prune
Frequently asked questions about TripoSplat in the browser
What is TripoSplat?
TripoSplat is an open-source model from TripoAI/VAST AI Research that converts a single 2D image into high-quality and variable-number 3D Gaussians. It uses a novel learned density-control approach called Density-Sampled Gaussians (DeG) that samples 3D Gaussian centers from a hierarchical octree-structured density function, enabling adaptive density control during generative modeling.
How does TripoSplat differ from traditional 3D Gaussian Splatting?
Traditional 3DGS requires hundreds of photos from different viewpoints and uses splitting/pruning for density control during optimization. TripoSplat is generative - it works from a single image using flow-matching models. Instead of predetermined pixel-aligned Gaussians, it samples from a learned density distribution, allocating more Gaussians to underfitted regions (splitting) and fewer to well-fitted regions (pruning) through policy-gradient training.
What input image formats does TripoSplat support?
TripoSplat accepts JPG, JPEG, PNG, WEBP, GIF, and AVIF image formats. Users can drag and drop image files from their computer, paste images from web pages or clipboard (Ctrl/Cmd+V), or provide a URL to an image.
What is Density-Sampled Gaussians (DeG)?
DeG is TripoSplat's novel representation where 3D Gaussian centers are sampled from a hierarchical octree-structured density function rather than predicted explicitly. This makes the non-differentiable sampling process end-to-end trainable using a policy-gradient approach from reinforcement learning, where density at each location is adjusted based on how much Gaussians at that location contribute to fixing rendering error.
Can I control how many Gaussians are generated?
Yes, TripoSplat offers inference-time budget control. Since Gaussian locations are independently sampled from the learned density distribution, you can specify your desired sample count at inference time. Use low counts for background elements to minimize storage and maximize frame rates, or scale up for hero assets with intricate details. You can also extract multiple LOD versions instantly.
What hardware do I need to run TripoSplat locally?
For local execution, you need Ubuntu or Windows with WSL2, and an NVIDIA RTX 3060 12GB GPU (or equivalent). The codebase is lightweight at ~2,000 lines and relies almost entirely on native PyTorch, bypassing heavy frameworks like transformers or diffusers that cause dependency issues.
Is TripoSplat free to use?
Yes, TripoSplat is fully open-sourced under the MIT license. Both the weights and inference code are publicly available on Hugging Face and GitHub. You can run it locally with no cost if you have compatible hardware, or use free Hugging Face Spaces with CPU Basic or ZeroGPU options.
What can I use TripoSplat-generated Gaussians for?
TripoSplat serves as a powerful pipeline tool for asset creation, AR/VR development, game development, simulation environments, and beyond. Since 3D Gaussian rendering is supported by mainstream rendering and game engines, the generated .splat files can be viewed and used in modern 3D pipelines for protoyping, background assets, hero assets with LoD systems, and stylized content.
How does TripoSplat's quality compare to other models?
In a comprehensive user study with 94 input images and 399 pairwise human preference choices from 32 participants, TripoSplat achieved an Elo rating of 1137, sitting over 140 points ahead of the next best model (TRELLIS at 975). It delivers significantly higher visual quality with the same number of Gaussians, or matches TRELLIS quality while using drastically fewer particles.
Does TripoSplat work in ComfyUI?
Yes, TripoSplat has Day-0 support in ComfyUI version v0.23.0, making it the first officially supported 3D Gaussian generator in the ComfyUI ecosystem. Users can update ComfyUI, open the Template Library, search for TripoSplat, choose the workflow template, download models, upload an image, and run the workflow.