Why clip_guided works better than text2im, inconsistent with the paper's claim?

[a43992899]

I have tested the glide model for a few days (I tried many kinds of prompts), and my result is that clip_guided works better than classifier-free text2im.

clip_guided can correctly follow the intention of my prompt, like "a boat on the top of the mountain", or "Pablo Picasso: Into the wind", and text2im failed to do that.

However the paper claims that classifier-free text2im > clip_guided. I wonder why? Is there anything wrong with the released model?