Dynamic Height / Width

[eolszewski]

Fantastic library - really appreciate all the work you all have done to provide such an amazing tool.

I have been poking around a bit with image interpolation and was curious if there was a path to using this model as a means of generating images of various sizes (instead of just 256x256).

I thought that I would be able to just hard code a few parameters (e.g. decoder_latents, super_res_latents), but, when I do this, I get something along the lines of:

Internal server error with unclip_images: Unexpected latents shape, got torch.Size([12, 3, 512, 512]), expected (12, 3, 256, 256)

This is due to the fact that what you pass in is expect to match the these are expected to match the UNet2DModels passed into super_res_first and super_res_last. This leads me to believe that I must be misunderstanding something for the fact that it's unclear why these parameters would even be included if they are just going to be checked against the related models, anyhow.

Any insight here is greatly appreciated.