app/services/detector/transforms.py

"""
Image transforms for the deepfake detector.
"""

from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize


def get_eval_transforms(processor, model_type="vit"):
    """
    Create evaluation transforms based on processor settings.

    Args:
        processor: The image processor from the model
        model_type: Type of model ("vit" or "siglip")

    Returns:
        Composed transforms for image preprocessing
    """
    size = processor.size["height"]
    image_mean = processor.image_mean
    image_std = processor.image_std
    normalize = Normalize(mean=image_mean, std=image_std)

    return Compose(
        [
            Resize(size if model_type == "siglip" else 256),
            CenterCrop(size),
            ToTensor(),
            normalize,
        ]
    )