Prompt truncation detection and return of tokenized prompts

[shirayu]

Is your feature request related to a problem? Please describe.

There are examples of long prompts on the web, but many people are unaware that long strings are truncated.
However, there is no way to know wheter StableDifusionPipelineOutputprompt truncated long prompt or not.

Long prompts are truncated here implicitly.

diffusers/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

Lines 175 to 181 in f6fb328

	text_input = self.tokenizer(
	prompt,
	padding="max_length",
	max_length=self.tokenizer.model_max_length,
	truncation=True,
	return_tensors="pt",
	)

Describe the solution you'd like

Returns a bool value indicating whether it has been truncated or not.
It is more convenient to return a tokenized prompt as well.

diffusers/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py

Lines 276 to 279 in f6fb328

	if not return_dict:
	return (image, has_nsfw_concept)
	return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

Describe alternatives you've considered

Additional context
In my app, after generating images, it tokenizes prompts separately to check whether they has been truncated or not.
However, this is a waste to tokenize them again.

https://github.com/shirayu/purepale/blob/ab89a20e1a0e1728d4cc98a3d8381b131c7c6319/purepale/serve.py#L167-L172

[patrickvonplaten]

Thanks for opening this issue - that's a good comment indeed!

Cc @anton-l @patil-suraj here