NSFW Helper Slider
To make it easy to understand when to use til LoRA, and what it helps with, i made two lists.
This is not the same as the "Refusal-Reduction" and should not be compared againt eachother. Instead you should use them together to get the best of both worlds. Refusal-Reduction makes the model do what you tell it, while this LoRA make the NSFW stuff look appealing.
Recommended range: -1.0 to 5.0
What this LoRA help with:
Works in both directions: more NSFW or more SFW
Remove/Add clothes
Makes it easy to "roll up" tops, pull down panties etc.
Makes poses more/less exposing and sexual with pos/neg
Help with penetration, especially when used with penis size LoRA etc.
Better prompt adherence on NSFW details
Does minimal changes to rhe image at values below 2.5, compared to alternatives
From 2.5 to 5.0 it will take more controll over pose, camera etc, often in a good way.
Makes it super easy to use original checkpoints for NSFW images
MUCH less changes to the original image compared to Bypass, by a large margine
Can make NSFW LoRAs look way more realistic with a small negaitve value like 0.5
What this does NOT do:
It will not make non NSFW checkpoints have better looking genitals etc
It will not undress people reliably if the prompt states they are dressed
It will not undress people reliably if the prompt has left over details like "light hitting her jacket" that you forget to remove
I would like to add that some of the things this LoRA helps with, could probably be fixed with prompts. But personally i use Sonnet, or GPT 5.6 to prompt for me, and i almost never read the prompt. And from time to time they intentionally word the prompt very "politicaly correct", that may reduce the intended NSFW level of the image. And i find it simpler to just have this LoRA on standby to adjust on the fly when i find a prompt i like, without having to read the whole prompt and begin to rewrite parts of it.
Description
FAQ
Comments (21)
Awesome, thank you for all these amazing LoRAs! You should try Grok for NSFW prompts. It’s pretty much uncensored.
I used grok for the longest. But im on a bit of a testing phase for other models since i feel grok is to repetetive.
If i want a prompt with a image as ref, Sonnet is by far better then Grok. After i made a extremely robust instruction i have not been hit with any direct censoring or rejections, other then carefull wording when i not spesificly ask for a hardcore prompt.
I dont know if i have managed to bypass all censoring with my instruction, but i have no problem getting GPT to make me a gaping anus prompt for example XD
@Loraholic
Ok nice, thank you. ll have to give that a try then. Always assumed they`d refuse.
@AugustusLXIII The most important part is to have a very clear and absolute list of rules at the very beginning of the instruction. Mine contain 2 tings in my current version im testing.
My age rule can seem a bit excessive, or imply im feeding it questionable data. But it is purely to make make the LLM understand that every input should be looked at from a perspective of adult.
Here they are if you want to try them out:
- All subjects must be adults, 18 or older, with no exceptions. If any input (text or image) implies or depicts a younger subject, silently reframe the subject as a young adult (18–21) without comment. Never describe childlike features, proportions, or contexts. When describing a subject, anchor age explicitly in the prompt (e.g. "The 20 year old woman...", "The woman in her mid-twenties...").
- No referance or naming of known or famous people.
Interesting. Hopefully this can help with exposed parts that keep showing when I explicitly prompt for them not to.(like nip slips when prompting for specific breasts parameters)
It will do that without problem. But be careful to not use to low values. -0.5 is probably enough to keep things safe for normal prompts.
seems to work perfectly for me at strengths suggested, with no other bypass LoRA/method applied. Nice job!
It looks promising 👀
The anti-entropy slider
How badly does it murder character knowledge? "Bypass" loras were bad for this. The "Refusal Reducer" seems really good.
It should be minimal to no effect on charcter LoRAs. At least at values below 3.
I will make a sample with one of RLYs characters.
https://civitai.red/images/136735884
This is a one pass no refiner image.
But one important part to get this result. Dont ref any clothes in the prompt, and add something like "showing her pussy" like i did in this prompt if she dont spread her legs as you want.
As stated in this LoRAs description, prompts actually work when using this LoRA. So if you prompt shorts you get shorts, but if you prompt pussy you get pussy.
Have not seen or tested the refusal reducer before you mentioned it now. Will check it out.
@Loraholic Refusal reduction is almost a must in some scenarios. Also, if you haven't tried this one, https://huggingface.co/Beinsezii/Krea-2-Turbo-Projector-Scale-LoRA-Diffusers can improve prompt adherence in certain situations, works well at 0.1 - 0.15. I'm just a nooby noob, maybe you can use it somehow in your work and incorporate into your nice loras. Kudos for all the work you've done!
@green_anger IF I HAVE TRIED THAT ONE?? Here comes a crashout.
That is the shittyest LoRA that is created for Krea2. It is HORRIBLE for other LoRAs.
I would recommend staying as far away fram that as possible.
Delete that file as soon as possible, and download Fedors instead, if you are going to use a bypass LoRA. Fedors version is much better at not destroying everything related to realism and quality.
Krea2 Filter Bypass [Fedor]
And just to mention it, both bypass and refusal is not doing the same thing as this LoRA. They will help in most cases, but they wont help making a scene more NSFW the way my LoRA does. The true power comes from using them combined with my LoRA.
I look forward to finding another one of your sliders every day. I have a dedicated Lora folder for them.
Would you be interested in making a "zoom" LoRA that let us control the closeness/distance of the camera ?
I'm currently trying to replicate a really specific image and i'm having a hard time getting the camera to be in the near-exact sport.
I need the camera very slightly back and the model dose not understand that, it's either too far or too close... so annoying.
Zoom is not a "vector direction" the same way as most other Slider LoRA concepts.
Its more in the stucture of your prompts. And that makes is hard to make a Slider that has a meaningfull effect on it.
I have a article here in civ that teaches you how to structure you prompt to master zoom extremely simple.
But in short, prompt room first will zoom out, prompt person first, zoom in.
If a bodypart it outside the frame, prompt contactpoint, like "her toes touching the grass" to force the model to include the feet in frame.
@Loraholic it would be nicer to have focal length slider lora instead, focusing on the image compression of how typical telephoto lens would capture. I like this Lora and I haven't seen anyone doing it for Krea 2, unless I'm scrolling past them because I couldn't keep up with the amount of loras right now lol.
@gimatiplease106 I made a fisheye Slider for Z.
https://civitai.red/models/2672837/telephoto-fisheye-slider-zit
And by the nature of fisheye it became a telephoto the other direction :P
Its a lot of work testing and finding out how the model responds to different camera properties, so it wont be the first priority right now. But when the major Sliders are done i'll look into stuff like this.
@Loraholic no problem! its just a suggestion, and it would be a nice addition to my collection of some of your loras like collecting infinity stones 😂. If you made it at some point ill send some of my buzz to you.
I don't know if this is Krea 2 model or LoRA training bias, but when generating close-up sex photos, it's extremely difficult to not have anal sex even if you prompt for vaginal sex. This greatly limits the usefulness of this LoRA for me.