Custom Extractors

Condition extractors transform a real image into a control signal that guides ControlNet generation.

Interface

All extractors subclass ExtractorABC and implement extract():

from ciagen.extractors.abc_extractor import ExtractorABC
from PIL.Image import Image


class MyExtractor(ExtractorABC):
    name = "MyExtractor"

    def __init__(self, **kwargs):
        pass

    def extract(self, image: Image) -> Image:
        return processed_image

Registration

1. Create ciagen/extractors/my_extractor.py with the class above

2. Add to ciagen/extractors/__init__.py:

from .my_extractor import MyExtractor

AVAILABLE_EXTRACTORS = (..., "my_extractor")

def instantiate_extractor(control_model: str, **kwargs):
    extractors = {
        # ... existing
        "my_extractor": MyExtractor,
    }
    return extractors[control_model](**kwargs)

1. Add a ControlNet config in ciagen/conf/config.yaml:

model:
    cn:
        - my_controlnet:
            extractor: my_extractor
            sd: my-org/stable-diffusion
            cn: my-org/controlnet-my-extractor

Guidelines

• extract() takes a PIL Image and returns a PIL Image
• Output should be the same size as the input
• Accept **kwargs in __init__ for forward compatibility
• Handle edge cases gracefully (e.g., no face detected)