Background Removal Tool

A Rust-based background removal tool using ONNX models (BiRefNet) with GPU support.

Features

• Download images from URLs
• Automatic ONNX model download and caching
• Background removal using BiRefNet-general-lite model
• Side-by-side comparison display (original vs processed)
• Checkered background visualization for transparency
• CLI with flexible options

Building

With Nix (Recommended)

# Enter the development environment
nix develop

# Build the project
cargo build --release

Without Nix

If you don't use Nix, you'll need to ensure you have the following system dependencies:

• Rust toolchain
• OpenSSL development libraries (libssl-dev on Ubuntu/Debian)
• pkg-config
• X11 libraries (for window display)
• Vulkan loader

Usage

Basic usage:

# Download and process an image
cargo run -- https://example.com/image.jpg

# Use custom model
cargo run -- https://example.com/image.jpg --model-path ./custom_model.onnx

# Offline mode (use only cached model)
cargo run -- https://example.com/image.jpg --offline

How It Works

1. Download: Fetches the image from the provided URL
2. Model Loading: Auto-downloads BiRefNet model (cached in ~/.cache/remove_background/models/)
3. Preprocessing: Resizes image to 1024x1024, normalizes pixels, converts to CHW format
4. Inference: Runs ONNX model to generate foreground mask
5. Postprocessing: Applies mask to original image, creates RGBA output
6. Display: Shows side-by-side comparison with checkered background

Project Structure

src/
├── main.rs           - CLI interface and orchestration
├── model.rs          - Model download and ONNX session management
├── preprocessing.rs  - Image preprocessing (resize, normalize, CHW conversion)
└── postprocessing.rs - Mask application and side-by-side display creation

Dependencies

• ort: ONNX Runtime Rust bindings
• ndarray: N-dimensional array operations
• image: Image loading and manipulation
• reqwest: HTTP client for downloads
• show-image: Cross-platform window display
• clap: Command-line argument parsing
• sha2: Hash verification

Future Enhancements

• CUDA execution provider support (GPU acceleration)
• Batch processing support
• Additional model options (U2-Net, ISNet, etc.)
• Save output to file
• Local file input support
• Progress bars for downloads
• WebAssembly support

Controls

• ESC: Close the window and exit

Model Information

Default model: BiRefNet-general-lite

• Size: ~50MB
• Input: 1024x1024 RGB
• Output: 1024x1024 mask
• Source: https://huggingface.co/ZhengPeng7/BiRefNet

License

See LICENSE file for details.