Trendyol DinoV2 Image Similarity Model

This repository contains a fine-tuned DinoV2 model for image similarity and retrieval tasks, specifically trained on e-commerce product images.

Model Details

• Model Type: Image Similarity/Retrieval
• Architecture: DinoV2 ViT-B/14 with ArcFace loss
• Embedding Dimension: 256
• Input Size: 224x224
• Framework: PyTorch
• Format: SafeTensors

Usage

Quick Start

import torch
from PIL import Image
from transformers import AutoModel, AutoImageProcessor

device = 'cuda' 

# Load model and processor from Hugging Face Hub
processor = AutoImageProcessor.from_pretrained("Trendyol/trendyol-dino-v2-ecommerce-256d", trust_remote_code=True)
model = AutoModel.from_pretrained("Trendyol/trendyol-dino-v2-ecommerce-256d", trust_remote_code=True)
model.to(device)

# Load and process an image
image = Image.open('your_image.jpg').convert('RGB')
inputs = processor(images=image, return_tensors="pt")

# Move inputs to CUDA
inputs = {k: v.to(device) for k, v in inputs.items()}


# Get embeddings
with torch.no_grad():
    outputs = model(**inputs)
    embeddings = outputs.last_hidden_state  # Shape: [1, 256]

print("Generated dimensional embedding shape:", embeddings.shape[1])

Preprocessing Pipeline

The model uses a specific preprocessing pipeline that's crucial for good performance:

1. DownScale (Lanczos): Resize to max dimension of 332px
2. JPEG Compression: Apply quality=90 compression
3. Scale Image: Scale to max dimension of 332px
4. Pad to Square: Pad with color value 255
5. Resize: Resize to 224x224
6. ToTensor: Convert to PyTorch tensor
7. Normalize: ImageNet normalization (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

Installation

Install the required dependencies:

pip install transformers torch torchvision safetensors pillow numpy opencv-python

Model Architecture

The model consists of:

• Backbone: DinoV2 ViT-B/14 (frozen during training)
• Projection Head: Linear layer mapping to 256 dimensions
• Normalization: L2 normalization for similarity computation

Training Details

• Loss Function: ArcFace loss for metric learning
• Training Data: E-commerce product images
• Epoch: 9
• PyTorch Version: 2.8.0

Intended Use

This model is designed for:

• Product image similarity search
• Visual product recommendations
• Duplicate product detection
• Content-based image retrieval in e-commerce

Limitations

• Optimized specifically for product/e-commerce images
• May not generalize well to other image domains
• Requires specific preprocessing pipeline for optimal performance
• Requires transformers library for feature extractor functionality

License

This model is released by Trendyol as a source-available, non-open-source model. See the LICENSE file for full details.

You are allowed to:

• View, download, and evaluate the model weights.
• Use the model for non-commercial research and internal testing.
• Use the model or its derivatives for commercial purposes, provided that:
  • You cite Trendyol as the original model creator.
  • You notify Trendyol in advance via scr.datascience@trendyol.com or other designated contact.

You are not allowed to:

• Redistribute or host the model or its derivatives on third-party platforms without prior written consent from Trendyol.
• Use the model in applications violating ethical standards, including but not limited to surveillance, misinformation, or harm to individuals or groups.

By downloading or using this model, you agree to the terms above.

© 2025 Trendyol Group. All rights reserved.

Citation

@misc{trendyol-dinov2-ecommerce,
  title={Trendyol DinoV2 E-commerce Image Similarity Model},
  author={Trendyol Data Science Team},
  year={2025},
  url={https://huggingface.co/Trendyol/trendyol-dino-v2-ecommerce-256d}
}