>>105688112
But the "results" can't be replicated at all. I know for a fact that I can take perfect seeds I've gotten on this newer version and compare to a previous version and the previous version will be worse. In fact I have done that already.
>>105682398
But even if you don't want to take that word for it. Actually I can prove to you that what you're comparing means nothing because what I prompt for is very difficult to get in one try. Take for instance this image I genned of a basic thumbs down. Though I got this first try and got lucky on v36, looking at other seeds the result wasn't so good. If I were to simply do a "same settings, same seed" test as you have shown, I could easily fool myself into thinking that the model has improved, whereas it was just a lucky seed.