Contributing Models

PowerAim’s in-app Downloadable Models tab pulls from two GitHub repos and merges them transparently:

1. This repo — fgilde/AI-Ming (the PowerAim fork)
2. Upstream — Babyhamsta/Aimmy

If the same filename exists in both, the newer commit wins; on a tie, this repo wins. Users see one combined list with no idea which side a model came from.

Naming convention

If you’d like to be credited, name your model:

[Game Name] by [Your Handle].onnx

If you’d rather stay anonymous:

[Game Name].onnx

How to upload

You can submit to either repo — both feed the same in-app downloader.

Option A — Upload to PowerAim’s fork

1. Fork https://github.com/fgilde/AI-Ming
2. Drop your .onnx into the models/ folder of your fork
3. Open a Pull Request against main
4. Once merged, your model appears in everyone’s in-app Downloadable Models tab on next refresh

Option B — Upload to the upstream Babyhamsta/Aimmy repo

1. Fork https://github.com/Babyhamsta/Aimmy
2. Add your .onnx to models/
3. Open a PR against master
4. The same model will be visible in PowerAim once the upstream PR is merged

Quality bar

We don’t reject models — anyone can submit. But for your model to be useful:

• Train on varied data (different maps, weapons, lighting, view angles)
• Export with dynamic input axis (dynamic=True)
• Include a names metadata field for multi-class models
• Test in-game first — open PowerAim, enable Show Detected Player, and verify confidences are reasonable

See Training Your Own for the full pipeline.

Removing a model

If you want a model removed, open another PR removing the file, or open an Issue against the repo where it lives.

Tips for high-quality models

• YOLOv8 architecture exported to ONNX with NHWC → NCHW input order
• Input image size from 192 to 1280 (multiple of 32) — PowerAim picks this up automatically from the ONNX metadata
• Multi-class supported — names custom metadata field with Ultralytics convention (e.g. {0: "Enemy", 1: "Teammate"})
• Thousands of varied images — PowerAim’s detection quality is bottlenecked by training data, not by inference
• Test in-game with Show Detected Player enabled — verify class confidences are reasonable

---

Thank you for contributing! PowerAim and the wider community benefit massively from shared models.