What changes?

• Created new file which is the wrapper script that inferences with BiRefNet model: ./BiRefNetModule/wrapper.py
• Added run_birefnet and get_birefnet_usage_options functions to the clip_manager.py script to connect backend to GUI.
• Added kornia package to requirements.txt and toml files.
• .gitignore modified to ignore BiRefNetModule checkpoint folders.
• Model downloads automatically handled within the wrapper script. Can be found at lines 67-82 in wrapper.py.

Code Documnetation

def run_birefnet(
    clips,
    usage="General",
    dilate_radius=0
)

run_birefnet: Starts inference with loaded clips.
usage: Model selection string which is selected from usage_to_weights_file (placed in wrapper script) dictionary. Dropdown selection can be directly passed to this.
dilate_radius: Mask expansion and contraction amount. I think range between -50 to 50 is ideal. Implemented as post-processor at line 177 inside wrapper.py file

def get_birefnet_usage_options(): # This converts `usage_to_weights_file` dictionary to keys list. Instead of including wrapper script to the frontend directly, using this makes it a little bit cleaner.
    return list(usage_to_weights_file.keys())

This is the model dictionary

Key values can be used to create dopdown values for GUI.

usage_to_weights_file = {
    "General": "BiRefNet",
    "General-dynamic": "BiRefNet_dynamic",
    "General-HR": "BiRefNet_HR",
    "General-Lite": "BiRefNet_lite",
    "General-Lite-2K": "BiRefNet_lite-2K",
    "General-reso_512": "BiRefNet_512x512",
    "Matting": "BiRefNet-matting",
    "Matting-dynamic": "BiRefNet_dynamic-matting",
    "Matting-HR": "BiRefNet_HR-Matting",
    "Matting-Lite": "BiRefNet_lite-matting",
    "Portrait": "BiRefNet-portrait",
    "DIS": "BiRefNet-DIS5K",
    "HRSOD": "BiRefNet-HRSOD",
    "COD": "BiRefNet-COD",
    "DIS-TR_TEs": "BiRefNet-DIS5K-TR_TEs",
    "General-legacy": "BiRefNet-legacy",
}

BiRefNet Model Zoo & Resource Guide:

• Default model is General
• Selecting a model will automatically adjust resolution for BiRefNet inference, it will not affect CorridorKey resolution!
• Dynamic models doesn't change the resolution. It will process on the original resolution of the image.
• "Note: Models listed as 'High' resolution require ~10GB+ VRAM. If you encounter OOM errors, consider using medium or small models."