All my generations are now hypernetworked

[leonnn1]

I don't know what happened but from one generation to another all my images became a product of the hypernetworks and I don't know how to fix it anyone is having the same issue?

This is the result that I should have

And this is what I am getting now

T_T

[leonnn1]

I tried everything but then when I was about to reinstall everything I found out that if I delete the files inside of the hypernetworks/pycash it start running back properly, then if I change anything the error starts again, I don't know what it is

[cgessai]

I've kind of tried to mess around with this issue a bit (I'm just some random dude, not a developer) and have you tried feeding your prompt into a place like dezgo.com and see whether the image it spits out is closer to your TOP or BOTTOM image? I think there's something screwed up with AUTOMATIC1111 because I'm not a cutting edge power user and when I put what looked like your prompt in to my install, I also received what looked like your 'unwanted' image. When I fed it into Dezgo.com, it came out looking like your 'wanted' image but the site won't accept seeds so only the basics of the prompt are rendered.

So I'm getting the vibe that you and I have something similarly wrong, while dezgo doesn't.

[leonnn1]

I realized the problem, at least to me, the cap was removed on the recent versions of automatic1111 but I jsut did a git pull today, so when I did it it extendend the cap above what I used before and my prompt before probably was exceding the cap, the problem is that this cap seems to be there for a reason, and the part of the prompt that extended the cap was a anime artist and because of this it was taking too much weight over the concept, when I removed that part after the cap it stoped, the problem is that I lost a bit on the process.
I don't know why it's doing this but if the first part of the extended cap does that, it seems a bit pointless to me to use it cause it will change all the weights you put on the first tokens

[cgessai]

Whoa, nice catch! I didn't think of that as a possibility. Even tried running the same prompt with all 12 "Stop At last layers of CLIP model" setting just in case that might be it. I just truncated your prompt after "artgerm" and it's dialed into the style of your 'wanted' image.

[EliEron]

I don't know why it's doing this but if the first part of the extended cap does that, it seems a bit pointless to me to use it cause it will change all the weights you put on the first tokens

They way the extended cap essentially works is that it divides your prompt into groups of 75 tokens and then adds their weight together. So token 1 and token 76 is equivalent weight wise, token 2 and token 77 is equivalent and so on. The simplest way of thinking about it is that you are starting over again weight wise once you get to token 76. And this happens again once you get to token 151 and so on.

[leonnn1]

this is what I thought but still remains the same to me, doesn't look much useful, it complicates a process in my oppinion, it would be better to structurize the prompt with 75 tokens, I ended up understanding what you explain after some hours of frustration trying to figure the issue out haha, at least it was a good learning ahha

[Jilocasin]

I don't think there's supposed to be a pycache folder in there, should it? I don't have one at least.

[leonnn1]

I don't know exactly I have it for some reason ahha