EfficientDet-Lite0
GPU80-class COCO detector. Best general-purpose choice — fast enough to run alongside other tasks.
Every model is quantized and cached locally after first download. Sizes are on-disk weights; latencies measured on an M1 Air with GPU delegate.
80-class COCO detector. Best general-purpose choice — fast enough to run alongside other tasks.
Higher accuracy at the cost of latency. Use when small / distant objects matter.
33 3D landmarks, single person. Ideal for fitness rep counting and form feedback.
Sharper landmark accuracy for slow, cinematic motion. Higher GPU pressure.
21 landmarks per hand, up to two hands. Drives the gesture recogniser and touchless cursor.
478 landmarks including irises, plus 52 ARKit-style blendshapes for AR filters.
Drop a .tflite or .task file into the studio's Storage panel. It's hashed, cached and available on next launch — no round-trip to the network.
Export from TF or PyTorch with int8 quantization. Aim for ≤ 8 MB total weight size.
Convert to the Tasks format (.task) if you need classifier or detector heads.
Upload in the studio. Benzo Vision keys it by filename + hash and stores it in IndexedDB.
Launch the studio and switch models in one click. Nothing downloads until you pick it.
Launch Studio