The documentation is not available anymore as the PR was closed or merged.

Is there a rationale why this was changed from 7.5 to 5.0? IMO, the documentation was correct and the actual default was wrong. The default used to be 7.5 in previous versions, and SD literature always refers to "between 7 and 10" as the optimal value.

In the paper, it says they used 5 (see the last paragraph in section 2.2) :)

Thanks for the info! However, the scale of 5.0 appears only once in the paper, in the specific context of conditioning on image size. There are four instances of example image outputs in the paper, on page 5, 9, 16 and 17, and in all of these cases, the scale used is 8.0. (EDITED TO ADD: In https://github.com/Stability-AI/generative-models/blob/main/sgm/inference/api.py, the default value is 6.0. Having rendered thousands of images at various cfg_scales, my impression is that a value between 6.0 and 8.0 is preferable to 5.0 across a wide variety of subjects.)

I'm still convinced that you are erroneously adopting a wrong default value here, and that your original documentation was correct.

Oh interesting, thanks for providing the additional info! Pinging @patrickvonplaten or @sayakpaul who can help provide more insight into why we chose 5.0

The original documentation came from a bad copy-paste. Let's ask stability what they think is best here :-)

@timudk @rromb - any tips on what should be the cfg default for SDXL?