MINC-Materials-23 ONNX

ONNX export and INT8 dynamic quantized version of prithivMLmods/Minc-Materials-23 (fine-tuned SigLIP 2 google/siglip2-base-patch16-224 on the MINC - Materials in Context dataset).

Designed for fast, lightweight PBR material classification and surface detection in 3D/VFX applications (Blender Addons, Unity, Unreal Engine, WebGL).

📦 Model Files

File	Precision	Size	Description
`model_quantized.onnx`	INT8 Dynamic	~84.7 MB	Recommended. Ultra-fast MatMul/Gemm INT8 quantization for lightweight CPU/GPU deployment.
`model.onnx`	FP32	~354 MB	Full-precision standard ONNX model.

🏷️ 23 Material Classes

The model predicts probabilities across 23 visual surface material categories:

brick, carpet, ceramic, fabric, foliage, food, glass, hair, leather, metal, mirror, other, painted, paper, plastic, polishedstone, skin, sky, stone, tile, wallpaper, water, wood

💻 Python ONNX Runtime Usage

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

# 1. Load quantized ONNX model session
session = ort.InferenceSession("model_quantized.onnx", providers=['CPUExecutionProvider'])

# 2. Preprocess input RGB image (224x224)
img = Image.open("sample.jpg").convert("RGB").resize((224, 224), Image.Resampling.BILINEAR)
img_np = np.array(img, dtype=np.float32) / 255.0

# 3. SigLIP Normalization: (x - 0.5) / 0.5
norm_img = (img_np - 0.5) / 0.5
tensor_in = np.transpose(norm_img, (2, 0, 1))[np.newaxis, ...].astype(np.float32)

# 4. Run ONNX Inference
logits = session.run(None, {"pixel_values": tensor_in})[0][0]

# 5. Softmax over logits
exp_logits = np.exp(logits - np.max(logits))
probs = exp_logits / np.sum(exp_logits)

classes = [
    'brick', 'carpet', 'ceramic', 'fabric', 'foliage', 'food', 'glass', 'hair',
    'leather', 'metal', 'mirror', 'other', 'painted', 'paper', 'plastic',
    'polishedstone', 'skin', 'sky', 'stone', 'tile', 'wallpaper', 'water', 'wood'
]

top_idx = np.argmax(probs)
print(f"Top Material Class: {classes[top_idx]} ({probs[top_idx]*100:.2f}%)")

📜 License & Attribution

License: Apache License 2.0
Original Model: prithivMLmods/Minc-Materials-23
Base Architecture: google/siglip2-base-patch16-224 (Google Research)
Dataset: MINC (Materials in Context) Database (Bell et al., Cornell University)
ONNX Conversion & Quantization: Manh Huynh (manhhuynhsd)

Downloads last month: -; Downloads are not tracked for this model. How to track