Turn a picture into a 3D mesh on your own graphics card. No cloud, no account, nothing uploaded. Drop an image in, press Start, save the result.
- Install a model. Open the Models screen, pick one, press Get the model and then Set it up. Local Mesh downloads the weights straight from Hugging Face and builds itself a private Python environment with the right PyTorch for your card. Nothing touches the rest of your computer, and you only do this once per model.
- Drop a picture in. Every image you drop on the Generate screen becomes a job in the queue, named after the file. Drop ten and walk away.
- Press Start. Jobs run one after another on your GPU, and each shape appears in the viewer as it finishes.
- Tidy it up, then save. Remove floaters, fill holes, reduce, smooth.
Every edit is a step you can walk back. Save writes a
.glb(or.obj,.stl,.ply) into your outputs folder.
That is the whole app. The knobs an expert might want, such as steps, guidance and resolution, sit behind a Quality picker (Fast / Balanced / Detailed) on each job, and everything else lives behind an Advanced switch in Settings.
Smallest card first. The Models screen checks each one against the graphics card it finds and says whether it fits.
| Model | VRAM | License | In short |
|---|---|---|---|
| TripoSR | ~4 GB | MIT | A rough shape in seconds. Good for checking a picture works. |
| Hunyuan3D 2 mini turbo | ~5 GB | Tencent Hunyuan Community | The one to start with. Clean shapes, fast, fits an 8 GB card. |
| Hunyuan3D 2 mini | ~5 GB | Tencent Hunyuan Community | The same model without the shortcut: slower, a little finer. |
| Hunyuan3D 2 turbo | ~6 GB | Tencent Hunyuan Community | The full-size model, still quick. |
| Hunyuan3D 2 | ~6 GB | Tencent Hunyuan Community | The sharpest you can get under 8 GB. |
| TripoSG | ~7.5 GB | MIT | Very sharp; tight on an 8 GB card. |
| Hunyuan3D 2.1 | ~10 GB | Tencent Hunyuan 3D 2.1 Community | The sharpest here. Needs a 12 GB card. |
| Step1X-3D | ~10 GB | Apache-2.0 | Sharp, and free for commercial use. Needs a 12 GB card. |
| TRELLIS | ~16 GB | MIT | A different approach that fails on different images. Needs a 16 GB card. |
There is also a Test shape behind the Advanced disclosure on the Models screen: it makes a simple mesh instantly, with nothing to download and no GPU, so you can see the queue and the viewer work before committing to gigabytes.
- An NVIDIA graphics card with 4 GB or more of memory, from a GTX 1080 up to an RTX 5090. Setup reads the card and installs a matching PyTorch. Without one the models run on the CPU, which works for the small ones but takes minutes.
uvandgiton your PATH. The Models screen tells you if either is missing.- To run from source: Node 20 or newer.
npm install
npm run dev # Vite dev server + Electron, hot reload
npm run build # typecheck + bundle to dist/ and dist-electron/
npm start # Electron against the production buildPipelines. A node editor for changing how a mesh is made, step by step: image → background removal → model → clean-up ops in any order → export. It appears in the sidebar after your first mesh, or from Settings → Advanced. A job can run a saved pipeline instead of a model.
Settings → Advanced. Idle unload, device, precision and low-VRAM mode. The defaults pick themselves per machine: fp32 on Pascal cards (no bf16, slow fp16), low-VRAM mode on when the card has less than twice what the model needs.
Logs. Ctrl+J opens three channels (general, errors, generation) in a
dock at the bottom of the window, backed by plain files in ~/.local-mesh/logs.
Everything lives in ~/.local-mesh: the Python environment, model weights,
each job's input copy and unsaved revisions, saved outputs, pipelines and logs.
The Electron main process spawns one long-lived worker.py from the
environment and talks to it over JSON lines
(protocol). The worker keeps a model resident
between jobs, reports memory, cancels cleanly and unloads itself when idle.
Backends are small Python modules, one per model, behind a common
load / generate / unload interface, and they ship inside the app in
resources/python.
src/core/ contracts: IPC, model registry, pipeline graph, worker protocol
src/main/ Electron main: env setup, Hugging Face downloads, worker + queue
src/ui/ React renderer (plain CSS, no UI library)
resources/python worker, backends, requirements, manifest
Adding a model is an entry in src/core/models.ts, a backend in
resources/python/backends/ with its line in backends/registry.py, an entry
in resources/python/manifest.json, and a bump of resources/python/VERSION.
The Models screen, the job settings and the downloader pick it up from there.
MIT for the app. Each model keeps its own license; see the table above.