text-to-image: replace nested dict by height and width properties in the input schema

[hanouticelina]

Flattening height and width parameters for text-to-image, making the API simpler for users and making provider-specific transformations (dict/enum) easier to handle for us.

yes, It's a breaking change but I expect the usage of target_size to be really minimal so far.

[Wauplin]

Thanks!

Just to be sure, it means we need to do the height/width => target_size conversion for "hf-inference" provider, right? Maybe can be done in the same PR for the hf.js client (here).

[hanouticelina]

Just to be sure, it means we need to do the height/width => target_size conversion for "hf-inference" provider, right? Maybe can be done in the same PR for the hf.js client (here).

Not sure about that, if the pipeline we're using for hf-inference is this one : https://huggingface.co/docs/diffusers/main/en/api/pipelines/stable_diffusion/text2img, we don't need to do the conversion.

It seems that it's indeed this one that is used for text-to-image : huggingface/api-inference-community/docker_images/diffusers/app/pipelines/text_to_image.py

[Wauplin]

Even better! Then, there shouldn't have been a target_size in the first place?

[hanouticelina]

Even better! Then, there shouldn't have been a target_size in the first place?

I guess so 😄

[hanouticelina]

failing CI is unrelated, it seems like there is a security check failure when trying to install node.js packages