---
license: apache-2.0
base_model:
- facebook/dinov3-vitl16-pretrain-lvd1689m
pipeline_tag: image-feature-extraction
tags:
- product_recognition
- image_embedding
---

# Cat2Real-DINOv3-384
Cat2Real-DINOv3-384 is a product image embedding model fine-tuned from DINOv3. It maps the real-world product images and catalog packshots to the same embedding space for similarity search.

# How to Get Started with the Model
The example below demonstrates how to obtain an image embedding with the [AutoModel] class.

```python
import torch
import cv2
import numpy as np
from torchvision import transforms
from transformers import AutoImageProcessor, AutoModel

def padding(img):
    """
    :param img: take the image as input, np.uint8 format, [0-255] range
    :return: img: square image with white pixel padded along the shorter side.
    """
    h, w, _ = img.shape
    if h > w:
        new_img = 255 * np.ones((h, h, 3)).astype(np.uint8)
        start_w = int((h-w)/2)
        new_img[:, start_w:start_w+w, :] = img
        return new_img

    elif h < w:
        new_img = 255 * np.ones((w, w, 3)).astype(np.uint8)
        start_h = int((w - h) / 2)
        new_img[start_h:start_h + h, :, :] = img
        return new_img
    else:
        return img

image_path = "your local image path"
dim = 384
image = cv2.imread(image_path)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = padding(image)
image = cv2.resize(image, (dim, dim))
transform = transforms.ToTensor()
image = transform(image)
image = image.unsqueeze(0)
model_name = "zhanganyi88/Cat2Real-DINOv3-384"
processor = AutoImageProcessor.from_pretrained(model_name)
model = AutoModel.from_pretrained(
    model_name, 
    device_map="auto", 
)
inputs = processor(images=image, return_tensors="pt").to(model.device)
with torch.inference_mode():
    outputs = model(**inputs)
embedding = outputs.pooler_output
print("embedding shape:", embedding.shape)
```