LocalMesh has opened a beta that turns a single photograph into a gaussian splat and then a textured mesh entirely on a local NVIDIA GPU.
The splat comes out of TripoSplat, Tripo's MIT licensed latent flow model, which pairs a DINOv3 encoder with a FLUX.2 VAE and runs 30 sampling steps to produce 262,144 gaussians. ComfyUI added TripoSplat natively on release day, and TACTYC shipped Onesplat, a Houdini SOP that generates a splat from one image on a local GPU. In the developers' wording, one photo means the back of the object is invented, not measured.
Instead of running Poisson reconstruction over gaussian centers, LocalMesh places 60 virtual cameras on a golden angle sphere, renders a depth map from each through gsplat, carves out the empty space every ray travels through, and fuses the 60 depth maps into a 768 cubed TSDF volume. A Poisson solver has to close every surface and can never report that it sees through a gap. The approach follows the gs2mesh principle, reimplemented from scratch.
The full conversion, covering splat rendering, TSDF fusion, decimation and atlas bake, stays under 7.3 GB of VRAM. A single image reaches a splat in about 60 seconds on an RTX 5090 and about two minutes on a 4060 laptop. LocalMesh ships a gsplat wheel with precompiled kernels for RTX 20 through 50 series, so installation needs no CUDA toolkit, no Visual Studio, no Python and no admin rights.
Meshes export as OBJ and GLB and splats as PLY, at triangle targets from 30k to 300k and quad retopology from 5k to 30k. Baked lighting is fitted as a spherical harmonics model against the surface normals and divided out, and ambient occlusion ships as a separate map.
The announcement does not state is which gsplat build the bundled wheel carries. LocalMesh runs on Windows 10 and 11 only, needs an NVIDIA card with at least 8 GB of video memory, and has no Authenticode certificate, so SmartScreen warns on first launch. Six models download once on first open, the largest of them TripoSplat at 4.5 GB.
LocalMesh is in beta with a seven day trial, then a one time $39 licence covering one machine at a time. It is available here.
