mickeyvanolst/triposr-coreml

TripoSR — CoreML (MLProgram fp16) for the AML TouchDesigner family

CoreML conversion of stabilityai/TripoSR (Tochilkin et al., Stability AI + Tripo AI, MIT) for the Geo Gen operator in AML — Apple ML operators for TouchDesigner. Single image → 3D mesh on the Apple Neural Engine in ~2 s on an M-series Air.

package	in	out	runs on
ImageToTriplane.mlpackage (800 MB)	image [1,3,512,512] RGB 0–1, subject on 0.5 gray, foreground filling 85%	triplane [1,3,40,64,64]	ANE / GPU / CPU
NeRFQuery.mlpackage (100 KB)	triplane, xyz [262144,3] in [−0.87, 0.87]	density [262144,1], color [262144,3]	CPU (fastest)

Marching cubes over a 3D grid of density queries runs on the host (TripoSR default threshold 25). Mesh frame: TripoSR is z-up with the subject facing +x.

Converted with coremltools 9 from the original PyTorch weights (transformers 4.45.2 / torch 2.7.0); conversion script and notes: labs/geogen/convert_coreml.py in the AML repository. Requires macOS 14+.

License

MIT — the original TripoSR license applies (© Stability AI, Tripo AI). The conversion is provided under the same terms.