huggingface/transformers · error · ValueError

Input image type not supported: {type(image)}

Error message

Input image type not supported: {type(image)}

What it means

to_pil_image accepts PIL images, torch tensors, and numpy arrays only. After converting tensors to numpy, any other type (list, str, path, bytes) hits this ValueError. It is a boundary check before channel-format juggling and uint8 conversion.

Source

Thrown at src/transformers/image_transforms.py:188

            and `False` otherwise.
        image_mode (`str`, *optional*):
            The mode to use for the PIL image. If unset, will use the default mode for the input image type.
        input_data_format (`ChannelDimension`, *optional*):
            The channel dimension format of the input image. If unset, will use the inferred format from the input.

    Returns:
        `PIL.Image.Image`: The converted image.
    """
    requires_backends(to_pil_image, ["vision"])

    if isinstance(image, PIL.Image.Image):
        return image

    # Convert all tensors to numpy arrays before converting to PIL image
    if is_torch_tensor(image):
        image = image.numpy()
    elif not isinstance(image, np.ndarray):
        raise ValueError(f"Input image type not supported: {type(image)}")

    # If the channel has been moved to first dim, we put it back at the end.
    image = to_channel_dimension_format(image, ChannelDimension.LAST, input_data_format)

    # If there is a single channel, we squeeze it, as otherwise PIL can't handle it.
    image = np.squeeze(image, axis=-1) if image.shape[-1] == 1 else image

    # PIL.Image can only store uint8 values so we rescale the image to be between 0 and 255 if needed.
    do_rescale = _rescale_for_pil_conversion(image) if do_rescale is None else do_rescale

    if do_rescale:
        image = rescale(image, 255)

    image = image.astype(np.uint8)
    return PIL.Image.fromarray(image, mode=image_mode)


def get_size_with_aspect_ratio(image_size, size, max_size=None) -> tuple[int, int]:

View on GitHub (pinned to a597f97485)

Solutions

1. Load files with PIL first: to_pil_image(PIL.Image.open(path)).
2. Unwrap lists: to_pil_image(images[0]).
3. Convert other array types to numpy before calling.

Example fix

# before
pil = to_pil_image('photo.png')  # raises ValueError

# after
from PIL import Image
pil = to_pil_image(Image.open('photo.png'))

import numpy as np
from PIL import Image
if isinstance(image, (str, bytes)):
    image = Image.open(image)
if isinstance(image, list):
    image = image[0]
if not isinstance(image, (np.ndarray, Image.Image)) and not is_torch_tensor(image):
    raise TypeError(f"unsupported image {type(image)}")

def is_supported_pil_input(x) -> bool:
    import numpy as np
    from PIL import Image
    from transformers.utils import is_torch_tensor
    return isinstance(x, (np.ndarray, Image.Image)) or is_torch_tensor(x)

When it happens

Trigger: to_pil_image([np_image]) (image wrapped in a list), to_pil_image('image.png'), to_pil_image(raw_bytes), or passing a tf.Tensor in an environment where it is not a torch tensor.

Common situations: Batch loops that accidentally forward a whole list of images, file-path confusion (thinking the function loads files), or non-torch tensor types leaking in.

Related errors

• Input image must be of type np.ndarray, got {type(image)}
• The image to be converted to a PIL image contains values out
• The image to be converted to a PIL image contains values out
• image must be a numpy array
• Invalid input type. Must be a single audio or a list of audi

AI-assisted analysis of huggingface/transformers@a597f97485 (2026-08-14). Data as JSON: /api/errors/5694f5b28d2b6304. Report an issue: GitHub.