Tools / Picture to 3D Object
Picture to 3D Object
Turn one picture of a thing - a chest, a burger, a bottle - into a small textured 3D model a game can drop straight into a level.
Working - admin onlyComing soon for players
16.7 saverage per object, picture in to game file out (warm)
~1.7 MBper object at 20,000 triangles with one 1024 texture
8objects below, made from pictures and icons
$0per object: no per-use fees
How it works
The steps
- Start from one picture of one object on a plain background - a photo, or an icon Flux 3 drew. The background is keyed out; thin sticker borders are trimmed, the picture is enlarged and padded to a square.
- TRELLIS builds the 3D shape and colours in about 7 seconds.
- Loose shards are removed, the mesh is reduced to 20,000 triangles, unwrapped, and one 1024 texture is baked - with the picture itself projected on the front.
- Out comes a GLB of about 1-2 MB, facing forward, standing on the ground, ready for three.js.
Examples
Turn them around
Two bench objects and six grab items from a painted-rooms test game, each from one picture. They turn slowly; drag to look at the back.
The 3D view loads when you scroll here
Drag to turn it, pinch or scroll to zoom.
Icons in, items out
These Flux 3 icons were the only input for the six items in the viewer.
Bench, 31 Oct 2026
Three models compared
| Model | Avg time | Licence | Verdict |
|---|---|---|---|
| TRELLIS + 20k game version (recommended) | 16.7 s | MIT code + MIT weights (microsoft/TRELLIS-image-large); DINOv2 encoder Apache-2.0 - commercial OK | Best props of the bench: crisp, straight, standing level (chest sits flat), believable back, right colours, 0.7-1.8 MB GLB at 20k tris. |
| TRELLIS (raw mesh) | 7.4 s | MIT code + MIT weights (microsoft/TRELLIS-image-large); DINOv2 encoder Apache-2.0 - commercial OK | Same shapes as above at full density: 0.2-1.3M triangles and 4-26 MB per GLB, too heavy for a browser game as is. |
| TripoSR + 20k game version | 4.5 s | MIT code + MIT weights (stabilityai/TripoSR) - commercial OK | Fastest and lightest: about 4-6 s per object, 3.7 GB VRAM, 1.3-1.6 MB GLB; looks the same as the full mesh at prop size. |
| TripoSR (raw mesh + its texture bake) | 42.5 s | MIT code + MIT weights (stabilityai/TripoSR) - commercial OK | Model itself takes 0.2-0.4 s, but its xatlas texture bake on the raw 110-280k-triangle mesh takes 20-60 s, and the GLBs are 4-9 MB. |
| Hunyuan3D-2mini turbo (shape only, no texture) | 16.3 s | Tencent Hunyuan Community Licence: NOT allowed in the EU, UK or South Korea; over 1M monthly users needs a deal; outputs may not train other models | Cleanest geometry of the bench: smooth, crisp edges, standing level, plausible back; model only 1.6-2 s (turbo, 5 steps), plus 10-19 s CPU clean-up; 6.3 GB VRAM; 0.7 MB at 40k tris. |
Models and licences
What it runs on
| Step | Model or tool | Licence |
|---|---|---|
| 3D shape and colour | TRELLIS-image-large (Microsoft) - our pick | MIT (code and weights); DINOv2 Apache-2.0 |
| Fast fallback | TripoSR (Stability AI) | MIT |
| Tested, not used | Hunyuan3D-2mini | Tencent community licence - not allowed in the EU, UK or South Korea |
| Cut-out | rembg (U2-Net) or our chroma key | MIT / Apache-2.0 |
| Bake | pymeshlab, xatlas, moderngl | GPL-3.0, MIT, MIT |
Rigid objects are TRELLIS's sweet spot - no skeleton, small texture. A drawn bust cut off at the edge came out as a flat card, so drawn characters go through the Photo to 3D Character tool instead. The back of an object can come out paler than the front.
Benchmarks and live pages