Forkable Factory · Concept → CAD bridge

Photo to printable.

Drop a photo. Get a printable STL. Compare open-source image-to-3D models on the one dimension nobody else benchmarks: does it actually print? Run them here, or run them on your own GPU and bring the mesh back for the same score.

Methodology: v1.0·How it's scored·See in the FF map →

Honest about limits

Single-image reconstruction guesses the back of the object. The mesh you get is the model's best fit to one viewpoint. For full accuracy, real CAD or multi-view scanning is still the right tool — multi-view support coming next.

Option 1

Score a sample mesh

instant · no api key

Try a sample · no api key needed

3 pre-baked meshes

Option 2

Generate in the cloud

needs a replicate key

Drop a photo

or click to choose. JPG / PNG / WebP. ≤8MB.

Plain background works best. We'll remove the background automatically.

Runs on Replicate, so it needs an API key. Set REPLICATE_API_TOKEN in your environment, or paste a personal token via the “Bring your own API keys” drawer below. No key? Option 3 costs nothing.

Option 3 · your own GPU

Run it locally with Modly

freeno api keynothing uploaded

Modly is an open-source desktop app that runs these same image-to-3D models on your own machine. It isn't something this page can call for you — it binds a local API to http://127.0.0.1:8765 on your computer, and your images never leave it. So: generate there, then drop the mesh below and it goes through the exact same printability rubric as the hosted models.

the models it installs — and how they map to the cloud lane

TripoSG
Same model this page runs in the cloud lane — so a local run and a hosted run are directly comparable.
Hunyuan3D 2 Mini
Distilled sibling of the Hunyuan3D-2 running here. Smaller and faster; the cloud lane runs the full model.
Trellis2 GGUF
Quantised Trellis. The cloud lane runs unquantised TRELLIS, so expect the local mesh to be a little coarser.

How to run it — 4 stepsshow

1
Install Modly and add a model
Grab a build from the repo, then add TripoSG (or Hunyuan3D 2 Mini / Trellis2 GGUF) from the in-app Models page. Models install from GitHub — nothing is bundled.
2
Check the app is up
```
python tools/modly-cli/agent.py health
```
The desktop app must be running. The CLI talks to it over the local API, and fails with a structured API_UNAVAILABLE if it cannot reach it.
3
Generate a mesh from your photo
```
python tools/modly-cli/agent.py generate \
  --image ./input.png \
  --output ./export.glb \
  --progress
```
Swap the extension on --output for .obj, .stl or .ply if you prefer. All four load in the scorer below.
4
Drop the result in below
Same printability rubric, same viewer, same scorecard as the hosted models — no API key, nothing uploaded.

Commands are quoted from Modly's own CLI reference. Windows · Linux · Apple Silicon macOS · MIT · 6.3k ★ — checked 2026-08-17.

which model made it? — labels the scorecard, nothing else

Drop the mesh Modly gave you

or click to choose. GLB / OBJ / STL / PLY.

Read in this tab — never uploaded anywhere.

Part of the Forkable Factory project. Methodology is public and versioned.

Disagree with a score? Open an issue at github.com/sinkrest — every score is reproducible from the input mesh.