List inputs in lda encode/decode, and clip text_encoder.

[piercus]

Naming of lda.decode_latents(latents) feels strange cause lda.decode(latents) is also decoding latents.

I suggest using lda.decode_image/lda.decode_images and lda.encode_image/lda.encode_images

To not break examples, the current PR is keeping an alias lda.decode_latents(latents) -> lda.decode_image(latents)

[piercus]

@deltheil @limiteinductive

In current code latents and latent are both used for a single Tensor.
I want to suggest to switch every single tensor variable to singular latent and use latents only when there is a sequence.

The naming for helpers would be image_to_latent and images_to_latent

WDYT ?

[deltheil]

In current code latents and latent are both used for a single Tensor.
I want to suggest to switch every single tensor variable to singular latent and use latents only when there is a sequence.

Not sure using plural will mean the same depending on the reader (e.g. latents -> latent variables vs. sequence of latent vectors). Perhaps we could make our life simpler and just adopt "latents" systematically (seems pretty widespread already: "CLIP latents", etc).

@limiteinductive any strong preference on your end?

[limiteinductive]

@limiteinductive, is there any strong preference on your end?

It's almost a convention that "latent" is always plural. Since a Tensor is a batch, you can have a single Tensor is still have multiple "latents."

[piercus]

Thanks for your help @deltheil @limiteinductive
Please review this new version.

[deltheil]

Thanks! Only minor feedback, please take a look

[deltheil]

@piercus do you mind rebasing this branch onto main and solve the conflicts? (ideally, you should also squash into a single commit - but I can do it myself at the end if that's simpler).

[piercus]

Squash and merge done