glowlytics-skin-models

Glowlytics Skin Analysis Models

On-device ML models for real-time skin health analysis. Part of the Glowlytics platform.

Models

1. Skin Signals (skin_signals.onnx) — 0.6 MB

Unified multi-head EfficientNet-B0 that predicts 4 skin health signals from a single face image.

Training: Knowledge distillation from Claude Sonnet 4 teacher labels on 4,717 images (UTKFace + FFHQ). 30 epochs on A10G GPU.

Input: RGB image, 224x224, ImageNet-normalized
Output: 4 float values in [0, 1] (multiply by 100 for 0-100 score)

2. Acne Lesion Detector (acne_detector.onnx) — 43 MB

YOLOv8s object detection model for identifying acne lesions (comedones, papules, pustules, nodules).

Training: 150 epochs + 50 epoch fine-tune (frozen backbone) on 1,843 images with 18,717 bounding box annotations. Trained at 1280px resolution on A10G GPU.

Input: RGB image, 640x640
Output: Bounding boxes with class labels and confidence scores

Files

Usage

import onnxruntime as ort
import numpy as np
from PIL import Image
from torchvision import transforms

# Skin signals
transform = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
])

img = Image.open("face.jpg").convert("RGB")
input_tensor = transform(img).unsqueeze(0).numpy()

sess = ort.InferenceSession("skin_signals.onnx")
scores = sess.run(None, {"image": input_tensor})[0][0]

signals = ["structure", "hydration", "sunDamage", "elasticity"]
for name, score in zip(signals, scores):
    print(f"{name}: {score * 100:.1f}/100")

Architecture

• Skin Signals: EfficientNet-B0 backbone (pretrained ImageNet) with shared hidden layers (1280 -> 512 -> 256) and 4 independent signal heads (256 -> 64 -> 1 with sigmoid). Total: 4.86M parameters.
• Acne Detector: YOLOv8s with cosine LR, AdamW optimizer, heavy augmentation. Phase 1: full training. Phase 2: frozen backbone fine-tuning.

License

MIT