Describe the solution you'd like
Support for PIL library image input (path) instead of Base64 encoding.
For example, when using models with transformers library, I provide images this way img = Image.open(path).convert("RGB") which isn't an additional 33% increase of memory overhead like it is with Base64.
Describe alternatives you've considered
I am currently using llama.cpp CLI version to process images in batch using GGUF models but I'd rather prefer this functionality here as well.