Models

PowerAim runs YOLOv8 ONNX models for object detection. Models are interchangeable: download a different one and PowerAim picks up the new input size and class list automatically.

This section covers:

• Using Models — switching models, the in-app store, the benchmark
• Training Your Own — labelling data and training a fresh model from scratch
• Contributing Models — submitting a model to the community store

How models work in PowerAim

• Models are .onnx files in bin\models\
• Input shape is read from the ONNX metadata — both fixed (e.g. 640×640) and dynamic (any multiple of 32) are supported
• Multi-class models are handled — the class names list comes from the model’s names custom metadata field (Ultralytics convention {0: "Enemy", 1: "Teammate"})
• Inference runs via ONNX Runtime with two execution providers:
  • DirectML (default build) — works on every DX12 GPU
  • CUDA (_cuda build) — NVIDIA only, ~10–25% faster
• Both providers fall back to CPU if the GPU path fails

Where models come from

• Bundled — bin\models\ ships several Universal models out of the box
• In-app Store — Models tab → Downloadable Models. Merges this fork’s models/ folder with the upstream Babyhamsta/Aimmy repo
• Trained yourself — see Training Your Own
• Manually placed — drop any .onnx into bin\models\ and click refresh

The merge rule for the in-app store: filename match between repos → newer commit wins; tie → fork wins. Users see one combined list.

---

Table of contents

• Using Models
• Training Your Own
• Contributing Models