Anthro Hands Detailer | Segmentation Model (YOLO26s-seg)

A single-class hand segmentation model for ADetailer, trained specifically on AI-generated anthro art - pawpads, claws and all the poses stock human-hand detectors choke on. Drop it into your A1111 / Forge / SD.Next or ComfyUI install, point ADetailer at it, and let it find and repaint the hands in your anthro generations.

This one is a instance segmentation model: ADetailer gets an actual per-hand mask instead of a rectangle stretched over the bounding box, so inpainting touches the hand (and just enough of the wrist) instead of a chunk of the background.

Quick facts

• Architecture: YOLO26s-seg (end-to-end, NMS-free)

• Classes: hand

• Native resolution: 1024 (letterboxed, gray padding)

• Training data: ~5,500 AI-generated anthro art images, ~9,000 hand masks

• Size / speed: ~10M params, ~22 MB

Validation results (v1.0)

Measured on a held-out validation set of 566 images / 993 hand instances (best checkpoint):

• Mask: Precision 0.950, Recall 0.937, mAP@50 0.968, mAP@50–95 0.811

• Box: Precision 0.948, Recall 0.934, mAP@50 0.967, mAP@50–95 0.821

In practice this means: it rarely invents hands that aren't there (high precision), and it finds hands down to fairly small sizes.

Installation

Automatic1111 / Forge / SD.Next

1. Download the .pt file.

2. Copy it to stable-diffusion-webui/models/adetailer/

3. Restart the WebUI (full restart if it was already open).

4. Select the model in the ADetailer tab.

> Tip: in Settings → ADetailer you can raise the number of ADetailer passes (up to 10) if you want to run a hand pass together with face/eyes detailers.

ComfyUI (Impact Pack)

1. Copy the .pt file to ComfyUI/models/ultralytics/segm/

2. Select it as the SEGM model in the UltralyticsDetectorProvider node.

Requirements

The model uses the YOLO26 architecture and needs ultralytics >= 8.4 (January 2026). If ADetailer throws an error on load, upgrade ultralytics inside your WebUI's venv:

venv\Scripts\python.exe -m pip install -U ultralytics

(portable A1111: run from the WebUI root; Forge/SD.Next — same idea, upgrade in the environment the WebUI runs from).

Recommended ADetailer settings

Starting point that worked well for me:

• Confidence threshold: 0.25–0.40 (drop to ~0.15–0.20 for small or partially hidden paws)

• Inpaint denoising strength: 0.30–0.40 (up to ~0.5 if the hand is badly melted)

• Use the ADetailer TAB to check detections before committing a full render.

Notes

The training set is anthro art, so anthro hands are the strong suit; human hands are outside the training distribution and may be detected inconsistently.

If it saves you a reroll, a review and a posted result are the best thanks. Feedback is welcome and directly feeds the next version.