killjoyelite/anime-eye-yolov8

Anime Eye Detector (YOLOv8)

A YOLOv8 object detection model fine-tuned to detect eyes in anime-style character art, intended for use with ComfyUI + Impact Pack for automated eye detailing/inpainting workflows (similar to how face_yolov8n.pt and hand_yolov8n.pt are used).

Model details (YOLOv8n and YOLOv8s)

• Base model: yolov8n.pt / yolov8s.pt (Ultralytics)
• Task: Object detection, single class (eye)
• Training data: 212 self-generated anime-style images (AI-generated, primarily female characters), manually labeled with bounding boxes around each visible eye
• Training config: 100 epochs, image size 640, batch size 8

Performance (on validation split)

Metric	Nano (yolov8n)	Small (yolov8s)
mAP50	0.995	0.994
mAP50-95	0.681	0.705
Precision	0.998	0.971
Recall	1.000	1.000

Known limitations

• Trained predominantly on female anime characters — detection on male character eyes is less reliable and may miss detections.
• Struggles with very large, cartoony/chibi-style eyes that deviate significantly from standard anime proportions or sometimes closeup of faces (WIP).
• Trained entirely on a single generation style/checkpoint's output — may generalize less well to very different art styles (e.g. heavily stylized, painterly, or non-anime art) than to mainstream anime/semi-realistic anime styles.
• Small dataset (212 images) — while validation metrics are strong, real-world robustness across the full diversity of anime art is inherently more limited than a larger, more varied dataset would provide.

If you find specific failure cases, feel free to open a discussion — this is a good candidate for community-driven dataset expansion over time.

Examples

Detection preview — the model correctly finds eyes across different poses/styles:

Eye color change/Eye fixing — using the detected eye region with Detailer (SEGS) to redraw eye color/detail from a prompt, while keeping the rest of the image untouched:

Usage (ComfyUI)

1. Download the size you want (eye_yolov8n.pt or eye_yolov8s.pt) from the Files tab.
2. Rename it if you want and place it in:

   ComfyUI/models/ultralytics/bbox/
   

3. Restart ComfyUI.
4. In your workflow:

   Load Image → UltralyticsDetectorProvider (select the eye detector model) → BboxDetectorSEGS → Detailer (SEGS)
   

5. Recommended Detailer (SEGS) starting settings for eye detailing:
   • guide_size: 512
   • denoise: 0.5–0.7 (lower = closer to the original eye, higher = more prompt-driven reinterpretation)
   • feather: 5–10

License

Released under the MIT License. Training images were self-generated by the author; users should independently verify licensing terms of any base checkpoint used to generate their own training/inference images if that matters for their use case.