I understand that LyCORIS are the new sexy and all that. Honestly, I'm not sure I see the benefits over LoRAs, and maybe someone can explain them. But that's not the point here.
However, from experimenting with LyCORIS, one thing I've noticed is that they seem to be extremely variable. Like, I'll run a celeb-based LyCORIS in a size 5 batch at a whopping 1.8 strength. And I swear, whether or not a generation looks anything like the LyCORIS model seems to be random. For example, this is the Felicia Day model run 10 times.
I'd say it looks like her in 2,4,6,10, and if you really wanted to argue it, number 10. Number 3 MIGHT be a cousin. The others aren't even remotely close.
I'm not used to seeing stuff like that with regular LoRAs.
Is there some fundamental concept I'm not catching here?
1.8 strength should look like dogshiz, are you using the correct like strength application? Network weight shouldn't be above 1 for any of them which likely means you're not even using it. Loras above a 1, should look seriously overcooked, and damaged if not burned or bleached out.
It seems you may not be using the inferred "prompt" keyword or trigger.
lycoris should work the same way in some ways as lora or LoHA- just slightly different.
Check your colab or SD settings, check your logs of when your'e gernerating there may be an error denoting that the mask weight isn't allow or something?
Yeah, well, I guess that's why I'm asking. Because I agree with you (mostly - I've had my own LORAs work fine to about 1.2), but even the original poster for the LyCORIS above has his samples at 1.4: https://civitai.com/models/37088/felicia-day
That's why I'm asking. These LyCORIS are more and more common, and I don't know if there's some fundamental flaw with them or if I'm just using them wrong.
Check out this run - it doesn't even begin to look like recognizably like Felicia Day until 1.4.
And another that doesn't become possibly recognizable until 1.3:
I'd say this one is more in the 1.5 to 1.6 range before you can know for certain who you're looking at: