Krea 2 LoRA.
Realistic skin, anatomy, and clothing. This LoRA handles all ethnicities, ages, and body shapes. It works for SFW and NSFW content.
Weight Guide: 1.0: General purpose. 1.5 to 2.0: NSFW.
a coffee is appreciated: ☕ Support me on Ko-fi.
Description
Krea 2 LoRA.
Realistic skin, anatomy, and clothing. This LoRA handles all ethnicities, ages, and body shapes. It works for SFW and NSFW content.
Weight Guide: 1.0: General purpose. 1.5 to 2.0: NSFW.
FAQ
Comments (27)
The preview images look great, but uploading them without a single prompt is pretty much useless.
Why would you not add any of the prompts used for generating the images you're promoting your lora with? 🤔
I understand but I am always paranoid and I strip metadata before I publish any image online.
That's why I mentioned that I will upload a few references soon. I will add prompts for each images. Thanks.
@RudySen Even if you strip out the metadata, you can add in the prompt when you add the images. tbf.
@RudySen i get where you're coming from 😅
But the worst people could get from your metadata would be the workflow used to create the image afaik.
Hope it didn't come off as rude, just thought it's a shame that the most realistic looking lora on civitai right now doesn't have any example prompts :D
Thanks for adding the prompts! 🫶🏽
@vslinx NW mate, and glad you like it. So when I create these images I am like mad scientist and my WF looks like my lab, also, I generally keep a lot of notes and prompts that I always use to stress test, So sometimes IDK what's in there. That's why I like to clean up and share the WF and prompts later. If that make sense.
Added all the prompts.
Just as a heads-up, it does seem like this LoRA is not compatible with the Int8 format, while most other LoRA(or every other one I have tired) does work.
AttributeError: 'set' object has no attribute 'calculate_weight'
Not sure what would be causing it, but it is a very popular format so might be worth looking into as to why.
(Nevermind, I just tested another LoRA that ran into the same issue)
You won't have the problem with my WF, as it's clean and not using any set node at all.
But if you want to fix yours, just reconnect the set node (calculate_weight) to the get node for that one, and it should fix it.
It's not LoRA issue, it's a connection issue between set node and get node
@RudySen oh really? Well, the way they are getting Int8 to work is having a Node that loads the Int8 model, into an Int8-specific lora loader and then to regular Ksampler. You saying there should be an additional one?
EDIT: hmm, wait. Do you mean those "connection" nodes? There is none of that being used in my WF. The direction connections are INT8 model loader > INT8 lora loader > Ksampler
EDIT2: Ah, I took a look at your WF. This problem isn't related to the get and set nodes, it is something else. Most likely nothing on your side of things. Thanks for at least answering.
@kossan okay, I sometimes get that "set" error because of the set nodes. Anyways, glad you found out what's happening
@RudySen I found out that if I used the INT8's "pre-lora" node, this lora seems to load just fine.
So if anyone else reads this, if you want make the LoRA's that get the error with INT8, use the pre-lora loader, that input into the INT8 Diffusion Model loader.
Also, thanks for the cool lora, RudySen! It's pretty nice!
did anyone manage to make it nsfw? I mean with krea2 in general, not only with this lora. Only one another lora is merely close to waht anima is capable of. Or am I just too corrupted?
Krea2FilterBypass works just fine.
(search for it here on civitai. With respect to the OP of this model, I don't wanna link to another.)
@kossan Yeah no worries, only reason I made this LoRA because that node was taking away details when you increase the values. Personally I did not like the output I was getting with it. Again, personal opinion.
@RudySen I actually agree. I did not like the node either! But some smart people have made LoRAs instead and it is much less intrusive!
https://civitai.red/models/2728234/krea2filterbypass?modelVersionId=3067151
I personally use the Vector 2 at 1.0 Strength.
Vector 3 gives a similar degradation as the bypass node did, so I'd avoid that unless you need to really really push the uncensoring.
Hi!
Thanks for the great workflow and LoRA. I have a question regarding your custom node, which I really like, by the way. However, I’m running into an issue with it. Regardless of whether I use LM Studio or Ollama as the backend (using uncensored thinking models), the generation always cuts off in the middle of the "thinking" process. I never get the final generated prompt.
Thank you. And very simple fix. Click on System prompt in Muse Chat and you will see the option to change 'Max reply tokens' crank that up to your liking, by default it's 2048, because a lot of people might use it with low VRAM setup, but I generally crank it up to 10k/20k. Setting up 10k doesn't mean it will use all 10k per reply, it's just a window I created because sometimes Qwen models get stuck in loop, I had that same issue in LM Studio as well. So I added the per reply context as well. Hope this helps.
💕👍🏻
even if I say "shaved pussy" it usually gives hairy pussy. can you improve that in next version?
wow that's actually completely opposite in my testing, I have to mention 'hairy' or generally I will get shaved. But sure thing, I will fix that.
I am working on the next dataset so I will definitely take care of this.
can we get lora for int 8 models,it looks really good
I can try
So I am training the next with a new dataset, no face at all, so it should work great with existing faces and character LoRA. I am still working on the dataset. So I will continue to improve this one.
But I would love to hear what you guys think I should make next. Should I make a Creative photography LoRA like I did for ZIT?