Turn one photo into a 3D-printed object with AI
Take a photo of a toy, a sculpture or a doodle, and a few minutes later you're holding it in plastic. Microsoft's open TRELLIS.2 model guesses the full 3D shape from one image, including the sides it never saw. A repair script then turns its screen-ready mesh into a solid, flat-based, correctly scaled STL your slicer accepts first time.
- A sliced print: settings for PLA figurines in Bambu Studio, PrusaSlicer or Cura, with time and filament
- A kit: the scripts, a sample image and mesh to practise on, and a settings cheat sheet
- uv to run the Python scripts: Put AI in your terminal. Each script installs what it needs
- A free Hugging Face account for the no-GPU route (you'll create a read token in step 3)
- A slicer: Bambu Studio ↗, PrusaSlicer ↗ or UltiMaker Cura ↗, all free
- An FDM printer, for example a Bambu Lab A1 mini (around £169 at Bambu's UK store ↗), or an online print service
- Optional: Blender ↗ for hand fixes, and Rent a cloud GPU to run the model yourself
Know what image-to-3D can and can't do
Pick the right model, and know which objects come out well before you spend an evening on one.
How it worksAn image-to-3D model has seen hundreds of thousands of 3D objects. Given one photo, it predicts the shape that best explains it, including the back, the underside and anything hidden. The visible side comes out close to the photo; the hidden sides are a confident guess.
The models worth knowing:
- [TRELLIS.2](https://github.com/microsoft/TRELLIS.2) from Microsoft: MIT licence, textured output in GLB format, the best default. Running it yourself needs Linux and an NVIDIA GPU with at least 24GB; Microsoft's free Hugging Face Space ↗ runs it for you.
- [TripoSR](https://github.com/VAST-AI-Research/TripoSR) from Stability AI and Tripo: MIT licence, fast and light (about 6GB of VRAM), lower detail. Good for simple shapes on your own NVIDIA card.
- Hunyuan3D 2.1 from Tencent: strong results, but its licence excludes the UK, EU and South Korea, so it's not one to use here.
What prints wellchunky, matte, single objects with a clear silhouette: figurines, toys, ornaments, sculptures, animals, fruit, shoes, simple products. What struggles: glass, chrome and see-through things, thin wires and hair, text, and anything where precise dimensions matter (a replacement bracket needs CAD, not a photo).
Everything you'll use is in one kitDownload the kit (163KB):
photo_to_glb.py,prep_print.py, a practice mesh generator, a sample image and a settings cheat sheet.