We built LocalMesh, one photo in, a Gaussian splat + textured mesh out, 100% on your own GPU. Beta is open, 7 days free.
Hi everyone,
We've been building LocalMesh for the past few weeks, and the beta just opened.
Short version: one photo goes in, and you get a 3D Gaussian splat (~60s on a 5090, ~2min on a 4060 laptop), then a textured mesh** (UV-unwrapped, photo reprojected, AO, normal map) exported as GLB/PLY/OBJ.
Everything runs locally. No upload, no queue, no credits. Turn off your Wi-Fi, it still works.
The mesh part is the bit we're proud of. Poisson reconstruction on gaussian centers drapes a tarp over your object: a Poisson solver has to close every surface, and can never say "I can see through here".
So instead: 60 virtual cameras on a golden-angle sphere, gsplat renders depth for each, every ray carves out the empty space it travels through, and the 60 depth maps fuse into a 768³ TSDF volume. Carving instead of guessing. (gs2mesh principle, written from scratch.)
Other things you might enjoy:
- It runs on 8 GB cards. The entire conversion (splat rendering, TSDF fusion, decimation, atlas bake) stays under 7.3 GB of VRAM, so a 4060 laptop finishes the exact same job as a 5090, just slower.
- No CUDA toolkit, no Visual Studio, no Python, no admin rights on the client. We ship a fat gsplat wheel with precompiled kernels, RTX 20 to 50.
Limits, upfront: one photo means the back of the object is invented, not measured. Windows + NVIDIA 8 GB only. No Authenticode cert yet, so SmartScreen will grumble on first launch.
7 days free, no card to start, then $39 once, no subscription, because there's no cloud to pay for: https://local-mesh.com
Break it and tell us, that's what the beta is for.
Happy to answer anything about the pipeline in the comments.