What are the parameters in CLIPTextEncodeSDXL?

[vedantroy]

The CLIPTextEncodeSDXL has a lot of parameters. After reading the SDXL paper, I understand that

• There are 2 text inputs, because there are 2 text encoders
• crop_w/crop_h specify whether the image should be diffused as being cropped starting at those coordinates
  • I would expect these to be called crop top left / crop top right if this was the case. Not sure what's going on here.
• (I'm guessing) target_w/target_h specify the rendered aspect ratio of diffused image

However, I'm not sure what the original width/height parameters do.

[vedantroy]

Answers:

• target_w/target_h correspond to $c_{ar} = (h_{tgt},w_{tgt})$ (conditioning on the bucket size used when training)
• crop_w / crop_h let you simulate the amount of cropping during inference (didn't manage to get this to work)
• width/height correspond to c_size, which is used to set the apparent resolution

[vedantroy]

(All of this can be found in the paper)