Spaces:
Running
title: TripoSplat WebGPU
short_description: Image-to-3D Gaussian splats with local WebGPU or MutualGPU.
emoji: 🫧
colorFrom: blue
colorTo: purple
sdk: static
pinned: false
models:
- Yosun/TripoSplat-WebGPU
TripoSplat WebGPU
Generate a 3D Gaussian splat from one image using either qualified browser-local WebGPU or community-contributed MutualGPU compute. Both modes return to the same interactive preview and export binary PLY plus browser .splat files. Qualified devices can also opt in as MutualGPU providers. Hugging Face hosts this static app and the model artifacts; it does not provide inference compute.
Before you run
- Local mode: use a qualified WebGPU browser and allow roughly 6.47 GB for the first verified model download and browser cache. The recorded qualification environment is Chrome 150 on an Apple M3 Max with 128 GB unified memory; Edge and 16 GB devices are not yet qualified.
- MutualGPU mode: the public
https://mutualgpu.com/exchange is selected by default, so you can choose an image and proceed immediately. It does not require local WebGPU or model storage, but it uploads the selected image to the exchange and assigned community provider. - Contribute mode: a device that passes local preflight can enter an operator-issued provider key, advertise an honest machine profile, and process one assignment at a time. The key remains only in page memory.
- Choose PNG, WebP, JPEG, or AVIF. Local opaque-image preparation downloads and caches a separate browser model on first use. MutualGPU submissions are normalized to a supported format and a maximum 1024-pixel edge.
- The standard run requests 20 sampling steps / 40 CFG DiT calls. This remains an engineering execution path, not a numerical-parity or production-readiness claim.
For local execution, the app downloads versioned model artifacts directly from Yosun/TripoSplat-WebGPU, verifies their declared size and SHA-256, and caches them in browser storage when permitted. Local mode keeps source-image inference on the device. MutualGPU generation sends the image to the configured exchange and assigned provider; provider mode receives untrusted requestor images. Deployment analytics, CDN logs, browser extensions, and image URL hosts remain separate privacy considerations.
Engineering preview: the four-step prepared-image path has a structural/export/viewer pass but misses a stricter diagnostic. Whole-scene numerical/render parity, repeated-run memory behavior, Edge, and 16 GB devices are not yet qualified.
Source, status, benchmarks, and issue tracker · Vercel mirror
Deploy this card and the prebuilt app
This is the production Space card for Yosun/TripoSplatWebGPU. The staging/test deployment lives at Yosun/TripoSplat-WebGPU-Demo and uses huggingface-space/README.demo.md.
Deploy from the repository root with the tracked script, which builds, runs pnpm verify:web-runner, and uploads the prebuilt bundle plus this card:
corepack enable
pnpm install --frozen-lockfile
DEPLOY_ENV=production node scripts/deploy.mjs --hf-only # production → Yosun/TripoSplatWebGPU
DEPLOY_ENV=staging node scripts/deploy.mjs --hf-only --test # staging → Yosun/TripoSplat-WebGPU-Demo
Both origins share the same model manifest namespace and CDN base, so redeploying either Space does not re-download files already cached in a visitor's browser. Only the first visit to each distinct Space origin downloads the model, because browser OPFS/Cache storage is origin-isolated.
The root dist/index.html is already built; this card intentionally uses only sdk: static and does not request Hugging Face compute or a custom build command. Hugging Face may still require plan credits to activate updated Static Space hosting.