BEN2 - Background Erase Network

BEN2 (Background Erase Network 2) is a state-of-the-art background removal model integrated into OpenTryOn. It provides high-quality background removal for fashion and product images, making it ideal for e-commerce, virtual try-on preprocessing, and product photography.

Overview

BEN2 uses a transformer-based architecture with window attention mechanisms for precise foreground-background separation. The model automatically downloads weights from Hugging Face on first use.

Key Features

• High Precision: State-of-the-art accuracy for background removal
• Automatic Weight Download: Weights automatically downloaded from Hugging Face
• Foreground Refinement: Optional refinement for higher quality edges
• Batch Processing: Process multiple images efficiently
• Flexible Input: Accepts file paths, URLs, BytesIO, or PIL Images
• GPU Acceleration: CUDA support for faster processing

Requirements

• CUDA 11.8+ (recommended for GPU acceleration)
• PyTorch 2.1+
• Model weights are automatically cached after first download

Installation

BEN2 is included with OpenTryOn. The model weights are automatically downloaded on first use:

pip install opentryon

Quick Start

Single Image

from tryon.api.ben2 import BEN2BackgroundRemoverAdapter
from PIL import Image

# Initialize adapter (weights download automatically on first use)
adapter = BEN2BackgroundRemoverAdapter()

# Remove background from a single image
result = adapter.remove_background("model.jpg")

# Save the result (PNG with transparency)
result[0].save("model_no_bg.png")

Batch Processing

from tryon.api.ben2 import BEN2BackgroundRemoverAdapter

adapter = BEN2BackgroundRemoverAdapter()

# Process multiple images
images = ["model1.jpg", "model2.jpg", "model3.jpg"]
results = adapter.remove_background_batch(images)

# Save all results
for i, img in enumerate(results):
    img.save(f"output_{i+1}.png")

Command Line Usage

BEN2 can be used directly from the command line:

# Single image background removal
python bg_remove.py --mode single --image model.jpg --output_dir outputs

# With foreground refinement for higher quality
python bg_remove.py --mode single --image model.jpg --refine

# Batch processing multiple images
python bg_remove.py --mode batch --images model_1.jpg model_2.jpg model_3.jpg

# Batch with refinement
python bg_remove.py --mode batch --images *.jpg --refine --output_dir results

CLI Arguments

Argument	Type	Default	Description
--mode	string	single	Processing mode: single or batch
--image	string	-	Path/URL for single image mode
--images	list	-	Multiple paths/URLs for batch mode
--output_dir	string	ben2_outputs	Output directory for results
--refine	flag	False	Enable foreground refinement

Python API

BEN2BackgroundRemoverAdapter

The main adapter class for background removal.

Constructor

BEN2BackgroundRemoverAdapter(
    weights_path: str = None,  # Custom weights path (optional)
    device: str = None         # Device: "cuda" or "cpu" (auto-detected)
)

Methods

remove_background

Remove background from a single image.

def remove_background(
    self,
    image: Union[str, io.BytesIO, Image.Image],
    refine: bool = False
) -> List[Image.Image]

Parameters:

• image: Input image (file path, URL, BytesIO, or PIL Image)
• refine: Enable foreground refinement for higher quality edges

Returns:

• List containing a single PIL Image with transparent background

Example:

adapter = BEN2BackgroundRemoverAdapter()

# From file path
result = adapter.remove_background("model.jpg")

# From URL
result = adapter.remove_background("https://example.com/model.jpg")

# From PIL Image
from PIL import Image
img = Image.open("model.jpg")
result = adapter.remove_background(img)

# With refinement
result = adapter.remove_background("model.jpg", refine=True)

# Save result
result[0].save("output.png")

remove_background_batch

Remove background from multiple images.

def remove_background_batch(
    self,
    images: List[Union[str, io.BytesIO, Image.Image]],
    refine: bool = False
) -> List[Image.Image]

Parameters:

• images: List of input images (file paths, URLs, BytesIO, or PIL Images)
• refine: Enable foreground refinement for all images

Returns:

• List of PIL Images with transparent backgrounds

Example:

adapter = BEN2BackgroundRemoverAdapter()

# Process multiple images
images = [
    "model1.jpg",
    "model2.jpg",
    "https://example.com/model3.jpg"
]

results = adapter.remove_background_batch(images, refine=True)

# Save all results
for i, img in enumerate(results):
    img.save(f"output_{i+1}.png")

print(f"Processed {len(results)} images")

load_image

Load and normalize image from various input formats.

def load_image(
    self,
    input_data: Union[str, io.BytesIO, Image.Image]
) -> Image.Image

Supported Inputs:

• File paths (local files)
• URLs (http/https)
• BytesIO / file-like objects
• PIL Image objects

Use Cases

Virtual Try-On Preprocessing

Remove backgrounds from garment images before virtual try-on:

from tryon.api.ben2 import BEN2BackgroundRemoverAdapter
from tryon.api import KlingAIVTONAdapter

# Step 1: Remove garment background
bg_adapter = BEN2BackgroundRemoverAdapter()
garment_clean = bg_adapter.remove_background("garment.jpg", refine=True)
garment_clean[0].save("garment_clean.png")

# Step 2: Use cleaned garment for virtual try-on
vton_adapter = KlingAIVTONAdapter()
result = vton_adapter.generate_and_decode(
    source_image="model.jpg",
    reference_image="garment_clean.png"
)

E-Commerce Product Photos

Create professional product photos with transparent backgrounds:

from tryon.api.ben2 import BEN2BackgroundRemoverAdapter
from pathlib import Path

adapter = BEN2BackgroundRemoverAdapter()

# Process all product images in a folder
input_dir = Path("raw_products")
output_dir = Path("processed_products")
output_dir.mkdir(exist_ok=True)

images = list(input_dir.glob("*.jpg"))
results = adapter.remove_background_batch([str(img) for img in images], refine=True)

for img_path, result in zip(images, results):
    output_path = output_dir / f"{img_path.stem}.png"
    result.save(output_path)
    print(f"Processed: {output_path}")

Model Swap Preprocessing

Prepare images for model swapping:

from tryon.api.ben2 import BEN2BackgroundRemoverAdapter

adapter = BEN2BackgroundRemoverAdapter()

# Remove background for cleaner model swap
person_clean = adapter.remove_background("person.jpg", refine=True)
person_clean[0].save("person_clean.png")

Performance Tips

GPU Acceleration

BEN2 automatically uses CUDA if available:

import torch

# Check if CUDA is available
print(f"CUDA available: {torch.cuda.is_available()}")

# BEN2 will automatically use GPU if available
adapter = BEN2BackgroundRemoverAdapter()  # Auto-detects device

# Or explicitly set device
adapter = BEN2BackgroundRemoverAdapter(device="cuda")  # Force GPU
adapter = BEN2BackgroundRemoverAdapter(device="cpu")   # Force CPU

Batch Processing

For multiple images, batch processing is more efficient:

# Less efficient: processing one by one
for img in images:
    result = adapter.remove_background(img)
    # process result...

# More efficient: batch processing
results = adapter.remove_background_batch(images)
for result in results:
    # process results...

Refinement Trade-offs

The refine parameter improves edge quality but increases processing time:

# Fast processing (no refinement)
result = adapter.remove_background("image.jpg", refine=False)

# Higher quality (with refinement) - takes longer
result = adapter.remove_background("image.jpg", refine=True)

Model Architecture

BEN2 uses a Swin Transformer-based architecture with:

• Window Attention: Efficient self-attention in local windows
• Shifted Windows: Cross-window connections for global context
• Multi-scale Processing: Captures both fine details and global structure
• Foreground Refinement: Optional post-processing for cleaner edges

The model weights are downloaded from Hugging Face: PramaLLC/BEN2.

Troubleshooting

Common Issues

CUDA Out of Memory

For large images on limited GPU memory:

# Use CPU instead
adapter = BEN2BackgroundRemoverAdapter(device="cpu")

# Or resize images before processing
from PIL import Image

img = Image.open("large_image.jpg")
img.thumbnail((2048, 2048))  # Resize to max 2048px
result = adapter.remove_background(img)

Weights Download Issues

If automatic download fails:

# Manually download weights
import huggingface_hub

huggingface_hub.hf_hub_download(
    repo_id="PramaLLC/BEN2",
    filename="BEN2_Base.pth",
    local_dir="tryon/api/ben2"
)

# Then use custom path
adapter = BEN2BackgroundRemoverAdapter(
    weights_path="tryon/api/ben2/BEN2_Base.pth"
)

Image Format Issues

BEN2 automatically converts images to RGB:

# RGBA images are converted to RGB automatically
result = adapter.remove_background("image_with_alpha.png")

# Grayscale images are also converted
result = adapter.remove_background("grayscale.jpg")

API Reference Summary

Method	Description
remove_background(image, refine=False)	Remove background from single image
remove_background_batch(images, refine=False)	Remove background from multiple images
load_image(input_data)	Load image from path, URL, or BytesIO

Related Documentation

• Preprocessing Overview
• Garment Segmentation
• Virtual Try-On API
• Kling AI

References

• BEN2 on Hugging Face
• Swin Transformer Paper