Problem in understanding change log of 0.8.5

[Anustup900]

We are using old versions of Kohya in which we get stable results for generating our fashion images in terms of near accurate prints/ shape/ orientation. But while testing with this version things are poor. So trying to understand what is the change or bug that you mentioned is actually is ? :

• What was the bug ? what was the concept to change the down scale frequency factor to 0 ?
• If i play around the values of the frequency factor should I will have any luck ? or there is some other catch ?

[Anustup900]

Ccing @kohya-ss if you can please give small explanation

[kohya-ss]

Sorry for the late reply.

This bug was caused when porting SDXL from SGM (the original implementation of Stability AI) to this repository. With this fix, the timestep embedding values will be correct. I am not certain how this fix will affect the results, but users have reported that the noise has been reduced. However, the impact seems to be minor overall.

If there is a difference between 0.8.4 and 0.8.5, it is due to this problem, but if you have compared with other versions, I think the problem you got is probably not due to this fix, but may be due to other factors.

If you get good results by changing the downscale_freq_shift, then that is not a problem. It is quite possible that you will get better results with your data set.