reading .png not working

[meissna]

Great work and thank you for putting the model and code open source!
If you read .png images into PIL, they might have either 3 or 4 channels, depending if they are transparent.
If a 4 channel .png-Image is used as input_images in:
results = pipeline(
prompt=instruction,
input_images=input_images,

I get an error.
To fix it, add .convert("RGB") after reading in the PIL-Image. e.g. in the preprocess-function.

def preprocess(input_image_path: List[str] = []) -> Tuple[str, str, List[Image.Image]]:
"""Preprocess the input images."""
# Process input images
input_images = []

if input_image_path:
    if isinstance(input_image_path, str):
        input_image_path = [input_image_path]
        
    if len(input_image_path) == 1 and os.path.isdir(input_image_path[0]):
        input_images = [Image.open(os.path.join(input_image_path[0], f)).convert("RGB") 
                      for f in os.listdir(input_image_path[0])]
    else:
        input_images = [Image.open(path).convert("RGB") for path in input_image_path]

    input_images = [ImageOps.exif_transpose(img) for img in input_images]

return input_images