How to return PIL.Image from dataset generator?

I have a custom GeneratorBasedBuilder subclass that I’m trying to get running. It seems like I’m almost there, but my trainer is failing with a type error when trying to access a generated sample—to be more specific, I’m working with the script and it’s failing at the preprocess_train(examples) function…

From what I understand, this is because the actual data is a PIL.Image, but my generator’s features dict lists the features as "image": datasets.Value("string"), which is obviously incorrect. However, it doesn’t seem like I can indicate an Image as a type here, so how am I supposed to do this?..

In my _generate_examples function I’m creating two images and a string as the types for my three columns.

        "image": [Image.Image],
        "text": datasets.Value(("string")),
        "conditioning_image": [Image.Image]

In addition to the above, I’ve tried the original (template) datasets.Value("string"), Image (which fails with cannot pickle module), and Image.Image

I’m doing some preprocessing in my _generate_examples function—basically adding a random mask—which I’d rather not do at the caller, if possible…

My _generate_examples function returns:

yield idx, { "image": target_img, "text": prompts, "conditioning_image": cond_img }

in case that helps… “target_img” and “cond_img” are both type PIL.Image

Why is this difficult… ???

Digging around in the ImageFolder class (well, it’s superclass) I found datasets.Image, but it fails with no useful error:

Exception has occurred: DatasetGenerationError
An error occurred while generating the dataset

During handling of the above exception, another exception occurred:

The above exception was the direct cause of the following exception:

  File "/home/james/src/diffusers/examples/text_to_image/", line 547, in main
    dataset = load_dataset(
  File "/home/james/src/diffusers/examples/text_to_image/", line 831, in <module>

okay, apologies for all the latent (or not so latent) rage… it was datasets.Image()… heh…