There won't be headings to this, just a bunch of "Well they told me to do it this way, and they don't me not to do it this way" - but with several trains that I pulled past normal settings everyone's been sharing...
Things have been looking very WEIRDLY -- more "STABLE" the more I've been breaking into "old habits" for SD 1.5 style training.
The opinionated guide I WROTE no less than a month or so ago? it's opinionated alright, and it was entirely -- well..
PUT IT THIS WAY:
Anything that anyone says you should or shouldn't do isn't a hard and fast rule except:
Don't do anything illegal, that's like -- a give in, like don't train it on ... MMMMM yea not going there LMAO. was gonna pull a meme about drugs and remembered that's kinda against TOS LOL!
Anyways.
Let's take a look at a couple pictures of what would be considered "STANDARD" SDXL training:
Don't ask me what my settings were, i've been using MOSTLY standard settings via CivitAI trainer laetly.
And then for reference, POLTERGEIST XL which i've almost reFUSED to bring to CivitAI becuase it just won't play ball. - this was trained via bmaltais back in like August.
FOR REFERENCE: THE TOP TWO WERE UNDER 1000 STEPS EACH, AND HAD FAIRLY. DECENT DATA SET SIZES, the likelyhood is the top was maybe AdamW8BIT ifi t's the 2nd miqote retrain.
NOW- Let's take THANCRED as your "We used roughly the basic idea of what you do bar maybe repeats on a Holostrawberry colab setting for SD 1.5" for SDXL:
2000ish steps, I can't remember repeats, and yes its' AdamW8bit - which we were all told at the first "DOES NOT WORK FOR SDXL" (and yes it does lol)
Note the fact that it's not just ACCURATE mostly in like the outfits and such, but evne tho of course it refuses to do a gunblade all one piece LOL - it's FAIRLY ACCURATE even on 1-1 or even 0.5
Also that thing they tell you about NEEDING it all to be ONE SIZE?
or even AT LEAST 1024 on all sides OR MORE?
I don't reccomend ALWAYS skirting around this, there are times SDXL will pick at the SMALLEST OF THE WORST OF YOUR DATA - but you can fudge it with a pre-made SD 1.5 dataset if you know what you're doing. (Do this sparingly, and don't blame me if it doesn't work)
One other thing:
WHILE YOU CAN TRAIN ON UNDER 50 IMAGES, SDXL gets really weird with 3D and semi realistic data if there's NOT enough of it - i've noticed it tries to break loras.
EXAMPLES:
AND THIS ISNT EVEN THE WORST i've seen, and my mother who adores horror films is like "PLEASE CHILD SAVE THE PICS SO I CAN CACKLE AT THEM" .... it's like a big middle finger to anyone who hates AI going 'ITS ALL PERFECT WE ARE ALL OUT OF JOBS' "No we're not it can't even do hands let alone faces' XD
Let's just say this about how you learn about SDXL from now on:
It's only rules are stated in your laws, your morals and your ethics.
Beyond that, if you want a lighter train and it doesn't give you the same troubles as I have? GO FOR IT!
If you like them stronger, do the strong!
I"m just commenting on my frustrations is all with this small tiny study.