@Cyberdelia yeah I'm having much better prompt adherence with 14.1 than with 15

It's pony. You can get away without the score tags now-days because of updates to model understanding. Though, using them still produces better results most of the time.

in the included model description: "it's also possible to remove the normal pony tags"

Full-body scans may reduce the weight of facial details, which may need to be controlled using LoRa sliders similar to facial detail magnification or enhancement.

v15 is by no means bad or anything (it's still very good), but so far I think I actually prefer v14 more. I'll have to test further to conclude, but v15 seems less versatile once you start adding lots of LoRAs or changing lots of things to manipulate image details. Might just be bad luck with small sample sizes, but I'm seeing increased anatomy defects (more than usual).

If I'm entirely honest, I kind of question the numbered iterative releases of Pony/SDXL checkpoints... Newer versions always imply updated & improved, right?

But it kind of feels as if Pony/SDXL has already very much plateaued now; & we're not seeing (or even capable of seeing much in the way of improvements at this point), but instead are just getting alt versions & different styles of checkpoints at this stage.

This isn't really directed at CyberRealistic Pony either, but pretty much all Pony/SDXL checkpoints. I'd be curious to hear Cyberdelia's take on this?

PS - I hope you have a very Merry Christmas btw (or whatever relevant festive celebration) if we don't see each other before the 25th ^^

Thanks for the comparison images. V14 has by far the best eyes/pupils. I'm coming from v12. Although every image situation is different, I don't feel ready for v15 at the moment.

Videos are made with another model, Wan 2.2 mainly. Pony is just for Image generations.

had same issue, changed scheduler to karras worked for me

@assowski501 thank you! it is working better for me too, had to adjust some things

@Cyberdelia yess, when I don't have to rely on the civit generator, I will definitely use version 11, its wonderful! I am getting the hang of version 15 now, I had to make some adjustments. I will always be thankful for your work :)

for sure, did you get one done yet?

how come?

@jaikapis120 same question

Find an image that's already NSFW, then check the prompt on the right hand side as a guide of what works and what doesn't. It's all about what you do or don't tell it to show when you're generating.

prompting "make pr0n" does a great job =P

or seriously you can try actual working prompts... such as:
"woman having sex with a man"
and the thing just does it. then you refine it. read comfyui's "prompting". is all about prompting, there you dress them, undress them, hair, eyes, and everything about it.

although keep in mind creating actual sex scenes to desired taste can take a lot of work and hours. models are trained with general poses, but they cannot read minds. again: prompting is everything.

have fun ;)

welcome to civitai. go on "create" -> "generate", choose a basic model (all the cyberrealistic pony" are good). use prompts (what you want to see) like: 1girl, solo, blonde, on beach, topless, exposed breasts, red bikini bottom, medium breasts, looking at viewer, waving, smile. use some negative prompts (what not should be part of the pic). sendpara is right: you should/can start with a good pic you have found, try to change the prompts / start with a remix. btw: slider are great tools too. there are: skin color slider, breast size slider, and many more. im sure you will learn the important stuff soon. if you have question: just ask.

The first rule of NSFW gen is you don't talk about NSFW gen.

use negative prompt for undesired things

I use zero negative (blank), everything I do is on the positive. For the nipples not showing, make sure you don't have ANY prompting regarding breasts or general attitude (aroused, horny, cold, etc). Avoid anything that could trigger erect nipples. Perhaps even using "Nipples are not visible" or "No nipples" (at the positive) would do the trick. I've learned you can pretty much do *anything* without needing any negatives, at least on ksampler (ComfyUI).

Using correct scheduler also does matter. sgm_uniform does an amazing job avoiding extremely useless details, my fav scheduler.

@Cyberdelia  why you don't train at high resolution?

i use 1456x2000 but i have a lora with 25000 steps trained at 2000x2600, going above 1500x2000 does to much errors, but if you did trained the checkpoint at higer resolution, over time we could increase it a lot 1456x2000 give already a very big improve

@Cyberdelia , ah but if you train at 1024, means your input is 1024 pixel also right? so if you want to make an output at 2000 or higher where does the details come from? , i don't use upscaler, specially with a lora, it will change your character face, it s already hard to make the exact same face, need a lot of try, if you use un upscaler then it s even worse, i want the same face not someone what looks similar

1472x704 (my current fav), 1600x640, 1344x768, 1216x832... and reversed resolutions. Those work fantastic and render fast with mostly all SDXL models. I do get much slower speeds while rendering square images (RTX 3090).

@carlosurportable419 how do you train loras?

@navyguy757

what do you mean with what program or the settings ?

i use one trainer