Now at 0.3 - 0.5 ms / frame for Face Detect, Crop, 468 Keypoint Detection, Hand Detect, Crop, Hand Keypoint Detection, including video processing full screen drawing and 1 euro filter smoothing / data processing and more. Not bad on an old M1 :)
Theres something so satisfying about optimization: Using some compute techniques and optimized models, and a custom MPS metal graph, Fabric will have sub 1ms hand pose tracking (on M1 Max) - meaning vision is 'free' at 120Hz with room to spare:
Why is CoreML / Vision so slow..?