V3 Is HERE!
All images have WF embedded, I recommend you to use euler/beta 12 steps on turbo. It gives better skin and textures.
Description
Improved stability by a large margin
FAQ
Comments (31)
There are some great Loras for Allround-NSFW stuff on the site but what I like about yours most is the fact that it seems to keep the subjects the way they are prompted. Most other loras bleed their hyper beautiful unnatural plastic queens into the generation. Yours not so much. Thank you!
EXACTLY bro, i was having a tough time getting porn of a "normal" looking woman with other loras, this one respects the base model MUCH more, first try with it and got what i wanted.
Works great! Prompt adherence is far better. Keep on updating the LoRA.
Dude i always loved your LORA for ZiT but the skin details are really bad for krea compared to other NSFW loras. I mean 0 disrespect. Any tips?
All images have WF embedded, I recommend you to use euler/beta 12 steps on turbo. It gives better skin and textures.
This skin looks bad to you? https://civitai.red/images/136308312
You need to stack multiple realism loras together. Just using this lora alone will not push it out of the plastic skin look. I stack like 5 realism loras together. The base model is incredibly biased towards the digital art synthetic look and you need to push it very hard to break it out of that style.
@fox23vang226 yeah you're right, I mix and match a few to get the desired result, It's a great Lora just on its own its not achieving believable results.
@alcaitiff Its doing a great job at prompt adherence but just isn't looking realistic in mostly daylight scenes on its own. Thanks for all your work though!
Really nice thanks!
Does this only work with turbo versions or can it work with Raw or else Raw with the turbo lora at a lower strength? Just wondering if it can be used with CFG > 1 and for that I assume we need a Raw model. Also I hear the Raw models have more variety of results than the turbo. Thanks!
It must work, the training was entirely did on the raw model.
what are the differences between the versions?
V1 uses a smaller dataset and has less concepts. V2 has a dataset 2 times bigger and know a lot more concepts. V3 is a stabilization improvement for less body horror and better prompt following. I tested V3 with 8k generations and will post several examples yet, it really do with easy what others struggle.
how do you get a woman to insert something that isnt a dick into her pussy such as a finger, toy, or other object. no matter what language i use they just seem to touch it to the outside. thx
You describe the final state of the object (that is, don’t say inserting/pushing). Instead, describe it as already inside or only half visible outside, etc.
thanks. any other tips?
@TheyTookErJerbs That principle also applies to pieces of clothing too if you’d like to remove them (half way undressed, for example).
Prompt adheres to few s*x positions but not all. How about all the kamasutra positions? Is it possible? if so would love to see that in next version
Wow. For this I'll need to fine tune the model. I don't think a small lora will do.
Best NSFW lora to use with Character Loras and it's not even close, this actually improves the likeness of your character. I dont know how it does it but keep it up.
What strength do you use usually?
@TheIntuit depends, sfw or lewd .5 need 1 for more
.5 is the sweet spot and just inpaint the penetrations with a Klein lora. I feel that at 1.0 strength it will overexpose the photo and slightly change the face. will also modify the fingers and toes.
It does cum on skin really well, but not ejaculating/projectile cum. Am I just prompting it wrong?
works wonders, loving it. Though skin realism is something it lacks a bit. Using at strength 0.90, skin has not texture. For this, I use RawGirlV2Spicy_epoch_10 at varying strength for better skin. I am using Euler/Beta and 12 steps.
getting strong jpg artifacts around silhouettes of people in high contrast lighting situations. fyi
For some reason, its generating comic/anime style images some times, even if me mention as photograph. Have to expliccitly mention in the prompt as
```Avoid comic, anime, digital art style images. Keep it photo realistic.```
Krea does this from time to time. Use my realism slider, it solves the problem. But really detailed prompts about the media like my examples will not generate the wrong media.
if your too descriptive of the subject like you used in z-image it will exaggerate the body proportions and assume its a cartoon. example: is@bel is a 18-year-old female, exceptionally short height, extremely petite, very skinny, ultra-thin build, tiny frame, delicate bone structure, narrow shoulders, thin arms, thin legs, very small waist, lightweight physique, miniature proportions, petite stature, lean figure, youthful appearance. big eye, full lips, long dark wavy hair. this will give me a good subject in Z-image but a cartoon in Krea2.
Does this include bypass or text refusal remover in it?
The training auto removed the filter, you don't need anything else.