mickeyvanolst/yolo-coreml

Ultralytics YOLO models as Core ML packages

Ready-to-run Core ML exports of four Ultralytics models, converted for the AML operator family for TouchDesigner on Apple silicon. Nothing in the models was changed: each is the official Ultralytics checkpoint exported with ultralytics 8.4.116 and coremltools 9.0 to a .mlpackage (fp16, NMS left to the caller), with Ultralytics' own metadata (task, class names, input size, licence) intact.

Package	Task	Input	Classes	Source weights
yolo26n-seg.mlpackage	instance segmentation	640×640	80 (COCO)	yolo26n-seg.pt
yolo11n-pose.mlpackage	pose, 17 keypoints	640×640	person	yolo11n-pose.pt
yolo11n-cls.mlpackage	classification	224×224	1000 (ImageNet)	yolo11n-cls.pt
FastSAM-s.mlpackage	segment anything (prompt-free masks)	640×640	object	FastSAM-s.pt

Export command, per model:

yolo export model=<weights>.pt format=coreml imgsz=<size> half=True

Licence

These files are derived from Ultralytics weights and are distributed under the GNU Affero General Public License v3.0, the licence Ultralytics publishes them under; the full text is in LICENSE. In short: you may use, share and modify them, and anything you distribute or serve that is built on them must be released under the AGPL as well. Ultralytics offers an Enterprise License for use outside those terms. Ultralytics, YOLO and FastSAM are the work of Ultralytics and the FastSAM authors; this repository only hosts a conversion.

Use in TouchDesigner

The AML Model Manager downloads these directly. Any Core ML consumer can use them: the input is an RGB image, the outputs are the raw YOLO tensors (detections plus mask prototypes for the segmentation models), so the consumer decodes boxes, applies NMS and assembles masks itself.