---
license: apache-2.0
---

# WithoutBG Snap Models

**Free, high-quality background removal models powered by AI**

[![Apache 2.0 License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
[![ONNX Runtime](https://img.shields.io/badge/Runtime-ONNX-orange.svg)](https://onnxruntime.ai/)

This repository contains the **Snap tier** models for the [withoutbg](https://github.com/withoutbg/withoutbg) library - a complete set of ONNX models for local, free background removal processing.

## 📋 Model Overview

The Snap tier implements a sophisticated 3-stage pipeline for background removal:

| Model | File | Purpose | Input | Output | License |
|-------|------|---------|-------|---------|---------|
| **Depth Estimation** | `depth_anything_v2_vits_slim.onnx` | Stage 1: Depth map generation | RGB (518×518) | Inverse depth map | Apache 2.0 |
| **Matting** | `snap_matting_0.1.0.onnx` | Stage 2: Initial background separation | RGBD (256×256) | Alpha channel (A1) | Apache 2.0 |
| **Refiner** | `snap_refiner_0.1.0.onnx` | Stage 3: High-resolution refinement | RGB+D+A (original size) | Refined alpha (A2) | Apache 2.0 |

## 🚀 Quick Start

### Using the withoutbg library (Recommended)

```bash
pip install withoutbg
```

```python
from withoutbg import remove_background

# Automatically downloads and uses these models
result = remove_background("image.jpg")
result.save("output.png")
```


## 🔄 Processing Pipeline

### Stage 1: Depth Estimation
- **Model**: Depth Anything V2 ViT-S (Apache 2.0 licensed)
- **Input**: RGB image (518×518 pixels, ImageNet normalized)
- **Output**: Inverse depth map (0-255 range)
- **Purpose**: Provides spatial understanding for better background separation

### Stage 2: Matting
- **Input**: RGBD (RGB + depth concatenated as 4-channel input, 256×256)
- **Output**: Initial alpha channel (A1)
- **Purpose**: Performs initial foreground/background segmentation

### Stage 3: Refining
- **Input**: RGB + depth + A1 (5-channel input at original resolution)
- **Output**: Refined alpha channel (A2) with high detail
- **Purpose**: Enhances edge quality and removes artifacts


## 🔧 Technical Specifications

### Model Details
- **Framework**: ONNX (compatible with ONNX Runtime)
- **Providers**: CPU, CUDA (automatically detected)
- **Precision**: FP32
- **Total Size**: ~140 MB (all three models)

### Input Requirements
- **Format**: RGB images (any resolution)
- **Preprocessing**: Automatic resizing and normalization
- **Output**: RGBA images with transparent background

## 🏗️ Integration Examples

### Batch Processing
```python
from withoutbg import remove_background_batch

results = remove_background_batch([
    "image1.jpg", 
    "image2.jpg", 
    "image3.jpg"
], output_dir="results/")
```

### Custom Model Paths
```python
from withoutbg.models import SnapModel

model = SnapModel(
    depth_model_path="custom/depth_model.onnx",
    matting_model_path="custom/matting_model.onnx", 
    refiner_model_path="custom/refiner_model.onnx"
)

result = model.remove_background("image.jpg")
```

## 📄 Licensing

### Open Source Components
- **`depth_anything_v2_vits_slim.onnx`**: Apache 2.0 License
  - Based on [Depth-Anything V2](https://github.com/DepthAnything/Depth-Anything-V2)

### Snap Tier Components  
- **`snap_matting_0.1.0.onnx`**: Apache 2.0 License
- **`snap_refiner_0.1.0.onnx`**: Apache 2.0 License
- Free for commercial and non-commercial use
- Open source models for the Snap tier

## 🔗 Related Links

- **Main Library**: [withoutbg/withoutbg](https://github.com/withoutbg/withoutbg)
- **Documentation**: [withoutbg.com/docs](https://withoutbg.com/documentation)
- **Demo**: [Hugging Face Space](https://huggingface.co/spaces/withoutbg/demo)
- **Commercial Licensing**: [withoutbg.com/focus](https://withoutbg.com/focus)

## 🎯 Use Cases

- **Development & Prototyping**: Free local processing
- **E-commerce**: Product photo background removal  
- **Social Media**: Profile picture editing
- **Content Creation**: Video thumbnails and graphics

## 🤝 Contributing

We welcome improvements to the open source components:

1. **Depth Anything V2 optimizations**: Submit PRs to improve inference speed
2. **Preprocessing enhancements**: Better image handling and normalization
3. **Documentation**: Examples, tutorials, and integration guides
4. **Bug reports**: Issues with model loading or inference

## 📧 Support

- **Technical Issues**: [GitHub Issues](https://github.com/withoutbg/withoutbg/issues)
- **Community**: [GitHub Discussions](https://github.com/withoutbg/withoutbg/discussions)

---

**Try the models**: Install `pip install withoutbg` and start removing backgrounds instantly!