README.md

---
license: cc-by-nc-4.0
library_name: coreml
pipeline_tag: image-segmentation
tags:
  - background-removal
  - image-segmentation
  - coreml
  - apple-neural-engine
  - birefnet
  - rmbg
language:
  - en
base_model: briaai/RMBG-2.0
---

# RMBG-2 CoreML - Background Removal Model

CoreML conversion of [BRIA AI's RMBG-2.0](https://huggingface.co/briaai/RMBG-2.0) background removal model, optimized for Apple Neural Engine (ANE).

## Model Description

This is a native ML Program format CoreML model converted from RMBG-2.0 (BiRefNet architecture) for high-quality background removal on Apple devices.

### Key Features

- **Native CoreML ML Program format** - Full ANE compatibility
- **INT8 weight quantization** - Reduced model size (233MB vs 461MB)
- **No custom layers** - Uses only native CoreML operations
- **1024x1024 input resolution** - High-quality segmentation
- **Fast inference** - ~3-5 seconds on Apple Silicon

## Usage

### With RMBG2Swift Package

```swift
import RMBG2Swift

// Simple one-liner
let rmbg = try await RMBG2()
let result = try await rmbg.removeBackground(from: image)

// Access the result
let outputImage = result.image  // Image with transparent background
let mask = result.mask          // Grayscale segmentation mask
```

**Swift Package:** [github.com/VincentGourbin/RMBG2Swift](https://github.com/VincentGourbin/RMBG2Swift)

### Manual Integration

```swift
import CoreML

// Load the model
let config = MLModelConfiguration()
config.computeUnits = .all  // Enable ANE

let model = try await MLModel.load(contentsOf: modelURL, configuration: config)

// Prepare input (1024x1024, NCHW format with ImageNet normalization)
let input = MLDictionaryFeatureProvider(dictionary: ["input": inputArray])

// Run inference
let output = try model.prediction(from: input)

// Get mask from output_3 (full resolution)
let mask = output.featureValue(for: "output_3")?.multiArrayValue
```

## Model Details

| Property | Value |
|----------|-------|
| Architecture | BiRefNet |
| Input Size | 1024 x 1024 |
| Input Format | RGB, NCHW [1, 3, 1024, 1024] |
| Normalization | ImageNet (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) |
| Output | 4 scales (output_0 to output_3), use output_3 for full resolution |
| Format | ML Program (.mlpackage) |
| Weight Quantization | INT8 (symmetric linear) |
| Model Size | ~233 MB |
| Minimum OS | macOS 13+ / iOS 16+ |
| Compute Units | All (CPU, GPU, ANE) |

## Performance

Tested on Apple Silicon:

| Device | Compute Units | Inference Time |
|--------|--------------|---------------|
| M1 Pro | .all (ANE) | ~5s |
| M1 Pro | .cpuAndGPU | ~3s |

> Note: ANE performance varies by device. The INT8 quantization improves memory efficiency and model size while maintaining quality.

## Files

Two model versions are available:

| File | Size | Description |
|------|------|-------------|
| `RMBG-2-native.mlpackage/` | 461 MB | Original FP32 model |
| `RMBG-2-native-int8.mlpackage/` | 233 MB | INT8 quantized model (recommended) |

**Recommended:** Use the INT8 version for better memory efficiency and smaller download size with equivalent quality.

## License

**Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0)**

- **Free for non-commercial use**: Research, personal projects, education
- **Commercial use requires a separate license** from BRIA AI

For commercial licensing, contact: [bria.ai/contact-us](https://bria.ai/contact-us)

## Attribution

This CoreML conversion is based on the RMBG-2.0 model by BRIA AI:

- **Original Model**: [briaai/RMBG-2.0](https://huggingface.co/briaai/RMBG-2.0)
- **Architecture**: BiRefNet
- **Creator**: [BRIA AI](https://bria.ai)

## Related Links

- [RMBG2Swift - Swift Package](https://github.com/VincentGourbin/RMBG2Swift)
- [Original RMBG-2.0 Model](https://huggingface.co/briaai/RMBG-2.0)
- [BiRefNet Paper](https://github.com/ZhengPeng7/BiRefNet)