Training Your Own Model

You can train a YOLOv8 model on your own labelled data and export it to ONNX for PowerAim. PowerAim picks up dynamic input shapes and multi-class metadata automatically.

This page is the short, opinionated version. For a full walkthrough, see the video tutorial linked at the bottom.

The pipeline

Capture frames  →  Auto-label / hand-label  →  Train YOLOv8  →  Export to ONNX  →  Drop into bin/models/

1. Capture frames

PowerAim ships built-in capture for training data:

Settings → Collect Data While Playing — saves capture frames to disk while you play
Settings → Auto-Label Data — additionally writes a YOLO .txt label file from the current model’s detections (the “auto-label” trick)

For a fresh class with no existing model, just capture frames manually. The Replay Buffer is also a quick way to grab clean frames.

In-app training-data capture

Together these two toggles bootstrap a dataset you can curate and retrain on — no separate recorder needed.

Collect Data While Playing saves capture frames as JPEGs to bin\images\ while the AI loop runs. Frames are throttled to roughly a couple per second (the save is rate-limited to at most one frame every 500 ms), so a play session yields a steady trickle of varied images rather than a flood of near-duplicates.
Auto-Label Data (in addition to Collect Data While Playing) writes a matching YOLO-format label file to bin\labels\ for each saved frame, named to match its image. The label is derived from the primary (nearest-to-centre) detection of the model that is currently loaded.

Each label file holds one line in the standard YOLO format — <class-id> <x-center> <y-center> <width> <height> — with the box coordinates normalized to the frame size.

Auto-labels are only as good as the model that produced them. Treat them as a head start: load the captured bin\images\ and bin\labels\ into a labelling tool, fix the misses, and add classes the current model can’t see yet before you retrain.

2. Auto-label with MakeSense.ai

The repo includes a TFJS export under Universalv3_web_model/ specifically for auto-labelling new data via MakeSense.ai:

Open makesense.ai
Load your captured images
Pick Object Detection as the task
Choose AI / YOLOv5 (TensorFlow.js) under the AI assistant
Upload the Universalv3_web_model/ files
Run the model — most of your frames get auto-labelled
Hand-correct the obvious misses

Export labels in YOLO format.

3. Train YOLOv8

Use Ultralytics’ YOLOv8 trainer. Minimal example (Python):

from ultralytics import YOLO

# Start from a pretrained checkpoint
model = YOLO("yolov8n.pt")  # or yolov8s.pt / m / l / x depending on GPU

model.train(
    data="dataset.yaml",   # standard YOLO dataset config
    epochs=100,
    imgsz=640,
    batch=16,
    name="poweraim-model",
)

A typical dataset.yaml:

path: ./dataset
train: images/train
val: images/val

names:
  0: Enemy
  1: Teammate

PowerAim reads the names field from ONNX metadata, so multi-class models work out of the box.

4. Export to ONNX

model.export(format="onnx", dynamic=True, simplify=True)

Key flags:

dynamic=True — exports with dynamic axis. PowerAim respects this and lets users pick image size at runtime.
simplify=True — strips redundant ops. Smaller file, faster load.
opset=12 or higher — matches what ONNX Runtime expects.

The exported file lands as runs/detect/poweraim-model/weights/best.onnx.

5. Drop into PowerAim

Copy best.onnx into <PowerAim install>\bin\models\ (rename to something descriptive — MyGame_v1.onnx).

Restart PowerAim or click refresh on the Models tab — your new model appears in the local list. Click to load.

Tips for high-quality models

Use a YOLOv8 architecture. PowerAim’s inference pipeline is built around YOLOv8’s anchor-free output. Other YOLO versions may work but are not officially supported.
Input size 192 to 1280, multiple of 32. PowerAim auto-detects this from metadata.
Multi-class support. List class names in the names custom metadata field.
Thousands of varied images. Different maps, weapons, lighting, view angles. Detection quality is bottlenecked by training data, not by inference.
Test in-game with Show Detected Player enabled. Verify class confidences are reasonable (>= 0.5 for clear targets).

Video walkthrough

A short walkthrough video showing the MakeSense.ai labelling → YOLOv5/8 training → PowerAim drop-in workflow.