A local 3D workshop. Type a prompt or hand over an image; a 3D model gets made on your own machine — viewed, styled, and exported without a single byte leaving localhost.
"garden gnome with red hat" on the Ultra tier — Hunyuan3D-2mini sculpts 100,000 triangles, TripoSR bakes the 1024² texture, ~30 seconds warm on an RTX 5060 Laptop — entirely offline. The shelf below holds earlier makes.
Two engines share one bench:
- AI · realistic — a fully local pipeline: your prompt is translated if needed
(opus-mt, CPU), painted into a reference image (SD-Turbo, 8 steps for a cleaner,
vividly-lit subject), cut out (rembg/isnet), sculpted into a real mesh (TripoSR)
on your GPU, then smoothed and baked into a UV-textured GLB. The baked color is
lifted out of TripoSR's naturally washed-out triplane range so textures read
lifelike rather than muddy. A prompt like
çeşmebecomes an actual tiered fountain — crisp 1024² texture included — in well under half a minute. Four detail tiers: Draft / Standard / Fine trade marching-cubes resolution for speed (5–16s), and Ultra hands the sculpting to a dedicated shape-generation model, Hunyuan3D-2mini (0.6B flow-matching DiT) at a 448 octree resolution, for visibly sharper geometry — 100k triangles in ~35s warm. TripoSR still paints it: its triplane color field is baked onto the Hunyuan mesh, so Ultra keeps the same textured-GLB output. Image uploads take PNG/JPG/WEBP and iPhone HEIC photos; the subject is cut out of the background automatically and the cutout is shown next to the result. - Instant — a deterministic procedural generator. It parses Turkish/English keywords (house, tree, car, robot, rocket… plus color/size/material adjectives) and builds parametric models with seeded variations. No GPU, no wait, works on any machine.
Everything downstream is engine-agnostic: orbit viewer with studio lighting, material editing, stylization (voxelize / low-poly / toon), transforms, an IndexedDB library shelf, and GLB / OBJ / STL export (STL is 3D-print ready).
Prompts can be Turkish or English — the UI is English, the understanding is both.
- Both servers bind to
127.0.0.1only. - Model weights are downloaded once at setup; after that the whole thing runs offline.
- Images, prompts, and generated meshes never touch the network. The top bar has a live sentinel that flags any request leaving localhost — it should never trigger.
backend/setup.bat clones the TripoSR and Hunyuan3D-2 source repos and downloads
model weights, then imports/runs that third-party code and loads the weights on
your machine. Some weights are pickle-based (.ckpt), which can execute code when
loaded. This is normal for a local ML tool, but it means setup trusts those
upstream sources — a compromised repo or weight file would run with your
privileges. Only run setup if you trust the upstream projects and your network.
| Instant engine | AI engine | |
|---|---|---|
| OS | any modern browser | Windows 10/11 (tested); Linux should work with minor script changes |
| Node.js | 18+ | 18+ |
| Python | — | 3.12 (installed automatically via uv) |
| GPU | not needed | NVIDIA, ~6.5GB free VRAM (8GB card recommended); Blackwell (RTX 50xx) needs the CUDA 12.8 wheels the setup installs. CPU fallback works but is slow (~1–2 min/model) |
| RAM | any | 16GB+ for Draft–Fine; 24GB recommended for Ultra (idle models park in system RAM) |
| Disk | ~200MB | ~12GB (PyTorch + model weights); +8GB if you use the Ultra tier (Hunyuan3D-2mini weights, downloaded on first Ultra make) |
Measured on an RTX 5060 Laptop (8GB): a Draft/Standard/Fine make is a 5–16s burst depending on the detail tier (Fine bakes the 1024² texture), ~45W / 64°C peak, ~6.5GB VRAM peak, idle between requests. An Ultra make is ~35s warm (~65-70s for the first request after the backend starts, while models load into RAM; longer still on the very first Ultra make ever, which also downloads 8GB of weights): diffusion sampling is ~11s of that, the rest is the 448-resolution surface decode, texture baking, and model shuffling — SD-Turbo, TripoSR, and Hunyuan all take turns on the GPU so the whole thing stays inside 8GB instead of spilling into shared GPU memory. A 512 octree resolution was tried and rejected: it looks sharper but its system RAM footprint (not VRAM) pushed the backend process past 23GB and into pagefile swap, tripling make times on a 24GB machine. Laptop-friendly. One Windows gotcha handled in code: a minimized backend console can be put in EcoQoS ("efficiency mode"), which made generation 4x slower — the backend opts itself out at startup.
npm install # web UI dependencies
backend\setup.bat # one-time: Python env, PyTorch cu128, TripoSR, model weights (~5GB download)Skipping backend\setup.bat is fine — the app runs in Instant mode without it.
start.bat # starts backend + web UI, opens http://127.0.0.1:5173TripoSR normally requires torchmcubes, which needs a C++/CUDA build. Kilnform
ships a small shim (backend/torchmcubes.py) that
provides the same interface via skimage.measure.marching_cubes, so nothing has
to compile. Mesh winding is normalized and the result is rotated z-up → y-up
before export. Texture baking (backend/texbake.py) is
adapted from TripoSR's bake code with two fixes: one reused GL context instead
of a fresh (leaked) one per make, and triplane color queries on the GPU.
The Ultra tier uses only the shape half of Hunyuan3D-2 (hy3dgen.shapegen),
which also needs no compiled extensions — the parts that do (its texture
painter) are exactly the parts Kilnform replaces with its own bake. Hunyuan's
mesh comes out in a different coordinate frame than TripoSR's; a fixed axis
permutation plus a per-make bounding-box fit (against a quick low-res TripoSR
mesh of the same image) lines the two up so the color bake lands correctly
(backend/hunyuan.py).
Kilnform's own code is MIT-licensed — see LICENSE. That covers this repository's own source only; the third-party pieces it stands on keep their own licenses (weights are downloaded by each user at setup, never redistributed here):
- three.js — MIT
- TripoSR (code & weights) — MIT
- SD-Turbo weights — Stability AI
non-commercial research license. Weights are downloaded by each user at setup,
not redistributed here. If you need commercial use, swap the model id in
backend/pipeline.pyfor a permissively licensed one. - Hunyuan3D-2 (Ultra tier, code & weights) — Tencent Hunyuan 3D 2.0 Community License. Commercial use is allowed up to 100M monthly active users, but use is not licensed in the EU, the UK, or South Korea — if you are there, stick to the Draft/Standard/Fine tiers, which don't touch Hunyuan. Weights are downloaded by each user on the first Ultra make, never redistributed here.
- Helsinki-NLP opus-mt-tr-en — CC-BY 4.0 · rembg/u2net — Apache-2.0
- UI fonts Fraunces and
Schibsted Grotesk — SIL OFL 1.1,
served locally from
public/fonts(no runtime font CDN requests)