---
license: apache-2.0
language:
- en
base_model:
- facebook/dinov3-vitl16-pretrain-lvd1689m
pipeline_tag: image-feature-extraction
---




# GR-Lite: Fashion Image Retrieval Model

GR-Lite is a lightweight fashion image retrieval model fine-tuned from [DINOv3-ViT-L/16](https://huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m). It extracts 1024-dimensional embeddings optimized for fashion product search and retrieval tasks.

GR-Lite achieves state-of-the-art (SOTA) performance on LookBench and other fashion retrieval benchmarks.See the [paper](https://arxiv.org/abs/2601.14706) for detailed performance metrics and comparisons.


## Resources

- 🌐 **Project Site**: [LookBench-Web](https://serendipityoneinc.github.io/look-bench-page/)
- 📄 **Paper**: [LookBench: A Comprehensive Benchmark for Fashion Image Retrieval](https://arxiv.org/abs/2601.14706)
- 🗃️ **Benchmark Dataset**: [LookBench on Hugging Face](https://huggingface.co/datasets/srpone/look-bench)
- 💻 **Code & Examples**: [look-bench Code](https://github.com/SerendipityOneInc/look-bench)

## Usage

### Installation

```bash
pip install torch huggingface_hub
```

For full benchmarking capabilities:
```bash
pip install look-bench
```

### Loading the Model

```python
import torch
from huggingface_hub import hf_hub_download
from PIL import Image

# Download the model checkpoint
model_path = hf_hub_download(
    repo_id="srpone/gr-lite",
    filename="gr_lite.pt"
)

# Load the model
device = "cuda" if torch.cuda.is_available() else "cpu"
model = torch.load(model_path, map_location=device)
model.eval()

print(f"Model loaded successfully on {device}")
```

### Feature Extraction

```python
# Load an image
image = Image.open("path/to/your/image.jpg").convert("RGB")

# Extract features using the model's search method
with torch.no_grad():
    _, embeddings = model.search(image_paths=[image], feature_dim=1024)

# Convert to numpy if needed
if isinstance(embeddings, torch.Tensor):
    embeddings = embeddings.cpu().numpy()

print(f"Feature shape: {embeddings.shape}")  # (1, 1024)
```


### Using with LookBench Dataset

```python
from datasets import load_dataset

# Load LookBench dataset
dataset = load_dataset("srpone/look-bench", "real_studio_flat")

# Get query and gallery images
query_image = dataset['query'][0]['image']
gallery_image = dataset['gallery'][0]['image']

# Extract features
with torch.no_grad():
    _, query_feat = model.search(image_paths=[query_image], feature_dim=256)
    _, gallery_feat = model.search(image_paths=[gallery_image], feature_dim=256)

# Compute similarity
import numpy as np
query_norm = query_feat / np.linalg.norm(query_feat)
gallery_norm = gallery_feat / np.linalg.norm(gallery_feat)
similarity = np.dot(query_norm, gallery_norm.T)
print(f"Similarity: {similarity[0][0]:.4f}")
```

## Benchmark Performance

GR-Lite is evaluated on the **LookBench** benchmark, which includes:

- **Real Studio Flat**: Flat-lay product photos (Easy difficulty)
- **AI-Gen Studio**: AI-generated lifestyle images (Medium difficulty)
- **Real Streetlook**: Street fashion photos (Hard difficulty)
- **AI-Gen Streetlook**: AI-generated street outfits (Hard difficulty)

For detailed performance metrics, please refer to:
- Paper: https://arxiv.org/abs/2601.14706
- Benchmark: https://huggingface.co/datasets/srpone/look-bench

## Evaluation

Use the `look-bench` package to evaluate on LookBench:

```python
from look_bench import evaluate_model

# Evaluate on all configs
results = evaluate_model(
    model=model,
    model_name="gr-lite",
    dataset_configs=["real_studio_flat", "aigen_studio", "real_streetlook", "aigen_streetlook"]
)

print(results)
```

## Model Card Authors

Gensmo AI Team

## Citation

If you use this model in your research, please cite:

```bibtex
@article{gao2026lookbench,
      title={LookBench: A Live and Holistic Open Benchmark for Fashion Image Retrieval}, 
      author={Chao Gao and Siqiao Xue and Yimin Peng and Jiwen Fu and Tingyi Gu and Shanshan Li and Fan Zhou},
      year={2026},
      url={https://arxiv.org/abs/2601.14706}, 
      journal= {arXiv preprint arXiv:2601.14706},
}
```