Image to 3D Model.

Exact 20-Step.
On Your Mac.

V2 combines browser WebGPU conditioning and spatial decoding with the untouched official fp32 flow sampler on local Apple MPS.
The final 20-step latent and camera match the official fixture bit-for-bit, while the inaccurate and slower WebGPU DiT is not downloaded.
Your image and flow tensors stay on your device.

Try it out Why this matters
SYSTEM CHECK

Preparing your browser

STEP 1 / IMAGE

Choose a starting image

ON DEVICE

Use a transparent PNG or WebP, or prepare an opaque photo locally before generation. The processor removes the background, frames the subject, and creates the official 1024px TripoSplat input.

No image selected yet.
OPTIONAL / LOCAL PROCESSING

Prepare a photo for TripoSplat

Opaque photos start background removal and the official subject crop automatically in this browser. Your pixels stay on this device; use the button to process again.

Choose an image to enable local processing.

Continue to browser engine
STEP 3 / RESULT

Your spatial scene

Waiting

A completed scene will appear here without leaving this page.

STEP 2 / EXACT MAC ENGINE

Preparing the model

CACHE CHECK

V2 keeps image conditioning and spatial decoding in WebGPU, then runs all 40 flow calls through the bit-exact official fp32 sampler on your Mac. The inaccurate 1.64 GB WebGPU DiT artifact is skipped.

REQUIRED / MAC V2

Connect the local MPS sampler

NOT CONNECTED

Start the v2 Mac launcher, then use its authenticated link or paste the printed token here. Tensor traffic stays on loopback; the service refuses non-official source and weight revisions.

Waiting for the authenticated local Mac service.

Checking model and browser storage…
Use a different model server
STATUS

Checking browser compatibility…

ENGINEERING PIPELINE

What this browser is doing

Technical mode shows exact graph stages, sampler steps, CFG invocations, and decode boundaries as they happen.

  1. 01
    Source imageInspect alpha and prepare local pixels.
    READY
  2. 02
    Model packageVerify manifest, integrity, and browser cache.
    WAITING
  3. 03
    ConditioningPreprocess → DINOv3 → Flux VAE.
    WAITING
  4. 04
    Exact Mac flow20 official fp32 steps · 40 MPS DiT calls.
    WAITING
  5. 05
    Spatial decodeOctree occupancy → Gaussian decoder.
    WAITING
  6. 06
    Scene handoffEncode PLY / .splat and swap the viewer.
    WAITING

Your image stays on this device. TripoSplat artifacts are requested from the official Hugging Face package unless you choose another server. When you select local image processing, its browser-local background-removal model is fetched and cached separately; transparent PNG or WebP files do not need that step.

WHY THIS MATTERS

A monumental attempt to put CUDA-era 3D generation in reach of more people.

TripoSplat started as a CUDA-based 3D generation pipeline. With Codex GPT 5.6 Sol Ultra and GPT 5.6 Sol Max in Kiro, this project is porting that work to WebGPU—replacing the usual requirements for NVIDIA hardware, CUDA, terminal commands, environment setup, and local dependency installation with a browser-native workflow.

The goal is not a cosmetic port. It is an attempt to make advanced AI 3D generation radically more accessible: for GPUPoor MacBook users, students, independent developers, and anyone without an expensive GPU. When a browser passes the system check, they can generate a 3D asset locally on hardware they already own—no terminal, no installation, and no cloud GPU bill for inference.

This is still an engineering preview, with compatibility and quality under active qualification. But the direction is clear: keep powerful spatial-generation tools inspectable, local, and available to more builders.

Read the source, benchmarks, and open issues on GitHub