axiom-of-choice/OmniParser-v2-icon-detect

OmniParser v2.0 Icon Detector (YOLO11) — MLX

Native MLX port of the icon detection model from microsoft/OmniParser-v2.0.

Model Details

• —Architecture: YOLO11-family (C3k2 backbone, SPPF, C2PSA attention neck, anchor-free Detect head)
• —Classes: 1 (icon)
• —Input: RGB image at native resolution
• —Output: Bounding boxes with confidence scores

Usage

import mlx.core as mx
from mlx_vlm.models.yolo11 import YOLO11, load_weights, non_max_suppression

model = YOLO11(nc=1)
load_weights(model, mx.load("model.safetensors"))

# Run detection
pred = model(mx.array(image)[None])  # (1, 5, N) — 4 box + 1 score
detections = non_max_suppression(pred, conf_thresh=0.05, iou_thresh=0.1)

Source

Converted from microsoft/OmniParser-v2.0/icon_detect/model.pt (Ultralytics YOLO11, nc=1). Parity verified: identical detections vs PyTorch reference (7/7 boxes, confidences to 4 decimals).