Effectively increases prompt adherence in a manner that's a bit different than CFG. Helps when the model isn't following instructions correctly.
- -0.005 weight = -0.5x
- 0.01 weight = +1×
- 0.1 weight = +10×
- 1.0 weight = +100×I used 0.4 to start with no other loras.
NOTE: If you're having issues with combining other loras start low at 0.01 strength and work up if needed.
Source: https://huggingface.co/Beinsezii/Krea-2-Turbo-Projector-Scale-LoRA-Diffusers
Uploaded here for civitai user convenience. This lora uses the built-in model anatomy knowledge without modifying the image significantly or relying on trained images. Very high quality results.
CONFIRMED working on all (nvfp4/mxfp8/fp8/bf16) checkpoints from https://civarchive.com/models/2726029/krea-2-turbo-official-comfy-org-checkpoints-krea2
Sample workflow uploaded: https://civarchive.com/models/2728073/krea-2-simple-workflow-with-civitai-create-post
Description
Initial version from https://huggingface.co/Beinsezii/Krea-2-Turbo-Projector-Scale-LoRA-Diffusers
FAQ
Comments (70)
please add working workflow
Almost done with a super simple workflow with built-in "Post to Civitai" feature. Hoping to upload later today.
black image
Can you share a prompt example that wasn't working?
delete your comfyui start flags if you have like: --sage attention, --fast etc.
@flo11ok874 I'm using sageattention3 with no problems.
also getting a black image.
@hughchungus420 Which workflow are y'all using? My custom one is working great. Not one black image so far. I have had it keep the woman clothed and needed varying strengths but ZERO black images.
try setting your compute_dtype in the diffusion model loader to fp32. that fixed the black image for me.
@wondier yep that fixed it, however doing so has increased my generation time from 48s to >200s
@ravenghost181374 If you want to use it with fast accumulation you can lower the strength to something like 0.1, it should work then. But it might not be as good (though still decent).
@ravenghost181374 if you find a true fix for it let me know! I just ran with the first thing that worked, but yeah the gen time is rough
@wondier I don't seemingly have that choice on AMD just a bunch of FP8 variations
is it really not working with int8?
I didn't try INT8. Was working on nvfp4 w/o issues. Had to increase to 0.6 so far depending on the prompt.
It does, I use Int8 and this LoRA does make a difference.
Int8 versions many times got issues with loras. Thats way I think of creat int8 model with preloded this lora at low str + nsfw v2 lora = this merge should be fastest and not problematic.
confirmed working on all the Comfy-Org checkpoints. nvfp4/mxfp8/fp8/bf16
@daceheg192491 Not for me on AMD, I just downloaded Krea 2 from the thing that tells you about missing models in ComfyUI, which gave me a fp8 model. I'm just getting a black image
@GPUPoorChad Yeah I don't have an AMD card to test on. Try manually pulling the fp8 model I guess that's linked in recommended resources. I would expect it to be the same though. Also make sure you're not jacking the strength up to high.
Works great! Thanks.
lol at 0.2 strength a regular realistic prompt output an anime image for me, at 0.4 strength as suggested by the description it's a black image. posting both my examples in a bit.
And you're using a comfy-org checkpoint? I just uploaded a post showing it's working for all the comfy-org checkpoints I uploaded. If you're using INT8, that's not from Comfy-Org and others have said it wasn't working also.
@daceheg192491 yes, the fp8 model from here and the fp8/bf16 TE from the huggingface repo.
your lora is downloading at 268 bytes, is that the correct file size? lol
@hughchungus420 Yup, you can see the embedded metadata from each checkpoint on my recent post https://civitai.red/posts/29398111 Could be a workflow issue?
@daceheg192491 I'm using the suggested workflow from the repo brother. My images should be up by now with metadata if you wanna take a look, played around a tiny bit switching from ER SDE Simple to ER SDE Beta 57 out of curiousity, beta as always seems to actually produce higher quality images but i don't think that's affecting your lora.
*my bad, it's this guy's
@hughchungus420 Appreciate the update. Please try with the comfy-org checkpoint and report back. I'm trying to get this workflow ready to publish to go alongside as well.
For black outputs remove comfyui file flags like: --sage attention or --fast. For anime well use slider - https://civitai.red/models/2727636/realistic-slider-krea-2-rank1-7mb?modelVersionId=3066073
@daceheg192491 I'll give the comfyorg repo's diffusion model a shot in that case, for future reference i got my model from here on civitai: https://civitai.red/models/2723583/krea2-turbofp8?modelVersionId=3060999
@flo11ok874 sage attention isn't causing it, i'm getting my gigantic anime tiddy demon girls with it just fine.
@daceheg192491 is not a workflow issue this lora seems very sensitive or erratic in gguf quants or fp8 models
@brahianvalles I've literally posted samples using the COMFY-ORG FP8 checkpoint. Other FP8's or GGUF may not work... I asked in your other post for a link to the GGUF so I can test with it.
@daceheg192491 Either way, it seems to break at strengths higher than 0.25. I've tried the other NSFW loras here on civitai at twice the strength with no issues (minus quality degradations lol). By this point i've upgraded to the comfyorg MXFP8 and using your lora in combination with this lora https://civitai.red/models/2725430/krea-2-nsfw-v2?modelVersionId=3065362 to pull off what i need, black images no longer an issue at 0.2 strength.
@hughchungus420 Still curious about the workflow because I've run it but never less than 0.4 and as high as 1.0 with still zero black images. Feels like something else is going on.
There are a lot of reasons that can cause this. For me it was the "--fast" flag when launching. It could also be ConditionZero or NAG.
@daceheg192491 my batch file for some reason had fp16 accumulation flag turned on, i had no idea this whole time despite my never using it in the kj node lol
Works perfectly. I use it at strength 1.0 with https://civitai.red/models/2725430/krea-2-nsfw-v2?modelVersionId=3065362
Can even combine with multiple character loras. No face bleed.
Can you share one of your workflows please?
@nofrickinway77 Try this one also https://civitai.red/models/2728073/krea-2-simple-workflow-with-civitai-create-post
@daceheg192491 Thank you! i'll give that a shot
This lora is very limited with my gguf model it gives a black screen
Well it isn't advertised as supporting anything other than the checkpoints it was tested on which are the Comfy-Org ones linked. Int-8, other fp8, and your gguf aren't in that list. ;) Happy to test if you point me in the right direction though
Works great, not only for NSFW.
How is this different from the NSFW conditioning: https://civitai.red/models/2725346/krea-2-nsfw-conditioning
The conditioning requires a custom node pack and IIRC the values vary depending on the checkpoint dtype? This works across all the Comfy-Org checkpoints and can be used simply in any lora loader. It targets prompt adherence, so works with SFW/NSFW equally AFAIK.
It's actually working great for me. It can override baked in anatomical bias in NSFW loras that maybe you want to change (like a gender bias - specifically genitals).
how are you using it? can you explain to me and share your workflow? thank you.
@necroryona One workflow that I posted is https://civitai.red/models/2728073/krea-2-simple-workflow-with-civitai-create-post
@necroryona just like any other lora. I'm using the workflow embedded in the realism engine lora sample pics. Only change is adding in this lora at strengths (.1-1.0). Oh and slight decrease to Realism Engine strength. That's about it. I've also been doing the same workflow but swapping out RE for Mystixxx.
This lora destroys image quality, makes images look extremely soft. I tried both 0.4 and 1 weights and no weight is good. Turning it off improved quality drastically.
Are you using it with some specific loras? I haven't seen any softness and for me it's the same quality as using no loras just with better prompt following. Also as you can see from the gallery... many people are getting GREAT results with it.
Could we get a version 2? Version 1 is causing skin artifacts and skin blotches.
Which diffusion model version are you using? What strength? You can see others are posting images without any artifacts/blotches. Keep strength reasonable... 0.6 -> 2.0 maybe. Try the FP8 version if that's not what you're using already https://civitai.red/models/2726029/krea-2-turbo-official-comfy-org-checkpoints-krea2?modelVersionId=3064297
@daceheg192491 Could it be step count related? I had mine at 12 and had issues. I had been using a different model that suggested 12 but with this lora it would have weird broken skin issues. 8-10 steps seems perfect. Sampler also seems to matter a lot. Tried Res_2s for a different lora, worked great. switched to this one. baaaaaaddd.
@Renkuya Euler / beta was best in my experience for sampler/scheduler and did make a huge difference. I mostly run mine at 8 steps but have done 12 also w/o breaking things. I've still not had a single black screen across mxfp8/fp8 or any of the Comfy-Org checkpoints
It's not their LoRA, they are just grifting. This is the source from the creator, as they had mentioned in their comments: https://huggingface.co/Beinsezii/Krea-2-Turbo-Projector-Scale-LoRA-Diffusers/blob/main/pytorch_lora_weights.safetensors
The creator also hadn't mentioned prompt adherence regarding NSFW so it is misleading. Refer to the latest comment of the person throwing a tantrum yesterday for more info and tea.
This LoRA should be deleted. It gives the Krea-2 a bad reputation because its giving people all kind of sideeffects. It messes with other LoRAs and wrecks quality.
The Krea2FilterBypass is much more predictable and made correctly.
Just look at the comparison i made image that show how this LoRA add random cencoring on the pussy at seemingly random values, her breasts are never exposed, and the lora has full effect, and wrecks the prompt completely from 0.1. While the correct lora has a gradual buildup to 10.0, witch makes it MUCH more friendly to use, and actualy makes the experience of using Krea-2 good.
From what I understand, this is not a de-censor node, but is meant to help guide it to follow the prompts better. In my test, I've combined it with Krea2FilterBypass for improved results. But I keep it quite low, at 0.01. As the original Creator explained, 0.01 = 1 strength.
I don't believe it should ever be used at 0.4 like this poster recommended.
Works with SFW/NSFW. I use this one at 2.0+ without issues. Maybe your lora is conflicting somehow? I also don't see it touching quality in the pic you shared. Gallery speaks for itself.
@daceheg192491 First of all, its not even YOUR LoRA. So i dont understand why you are posting it in the first place. Create your own models instead of just uploading other peoples work.
When you are not the creator of the LoRA you cant event explain how it works, or how it should be used correctly.
I have been in contact with mulitiple people in my DMs that have problems with my LoRAs, and EVERY SINGLE ONE of them has used this LoRA. And when they changed to the Bypass LoRA all problems went away.
This LoRA WREKS the model. It sure creates some good images, BUT those images would have been MUCH BETTER on the correct LoRA.
JUST DONT UPLOAD A LORA YOU DONT OWN AND DONT KNOW HOW IT WAS MADE OR HOW TO USE IT.
@Loraholic You sound super salty bro. I referenced the source in the post. Everything we're all doing is derivative including your amazing slider loras. You're benefiting from posting on the galleries just like everyone else. Options don't hurt anyone. Maybe make your loras work in a more compatible way? <3 Feel free to hit me up in DMs or on discord if you want a real conversation about it.
@daceheg192491 Im upset because i put in huge amount of work, only for this stupid LoRA to wreck the potential, because you have a recommended 0.4 start value in the description. If people only used 0.01 it would possibly be ok.
@Loraholic 0.4 was what I used with NO other loras to get some basic nsfw working. At the time... your Loras did not exist :) The landscape has changed a ton and now there are a LOT of loras which is great for the community. More than happy to update the description and say "If you're having issues with combining other loras start low at 0.01" But that's not what you asked for... you came in demanding I delete it and questioning my motivation for posting a publicly available lora that people are getting benefit from. No potential has been "wrecked." At most you've fielded some comments/messages from folks having issues. It's a new model, this is expected.
Some people (including me) don't like to use Lora with models. It feels like a hack to the model instead of using the full potential.
This one small Lora only plays with the tensors, for people who really needs to bypass the refusals.
The model will not do more than originally trained tho.
@civitailwhkh995 If used at 0.01 to 0.1 its ok, ish.
But subjects go from looking normal to looking like supermodels. I guess you want to use only the base model to get the "natural" Krea2 image, and sorry to say, but this LoRA affect the original "look" way more compared to alternatives.
Let's be honest. This or any similar lora I have tested ruins coherency for more difficult prompts. There are no good solutions yet.
This is deceptive content and should be reported as such. You could make serious legal backlash on civitai through defamation on the initial creator. The NSFW part is misleading since the huggingface creator never stated that it was NSFW prompt adherence. Also, you're grifting both off of Raison D'etre (Beinsezii), the real creator on HuggingFace and the Krea2FilterBypass by replicating it's name.
To answer most people's questions since this person clearly has no technical experience: Putting a FP32 on a gguf is pushing it. Having it on an FP8 is considerable, but only when used minimally (like 0.1).
Imagine dropping a turbo in a Chevy Cavalier. Can you do it? Sure. But it is a hit or miss and will likely blow the engine if you make one small tune to it (like adding another LoRA). It seems like a redundancy putting an entire FP32 LoRA on a model just for a slight weight adjustment that the Krea2 Developers might make. The Krea2FilterBypass is what you're likely looking for, even if the creator embarrassed themself in the comments.
Using it at 0.2 alongside DarkBeastKrea2 checkpoint is fantastic. Unlock NSFW very well, and lets it compete with Chroma for king of the local hardcore gens. No other LORA required.
Difference for me with that, is that other jailbreaks kill things like TV/Movie reference styles, where this does not.
I get issues with totally black screen though if I push to 0.4 or above.
which checkpoint version are you using?