I've been messing around for DAZ some time now and I wanted to start testing with the Stable Diffusion AI image generators and image2image, but it censors even cleavage. Anyone know how others are managing to make sexy content on these? I came across a link to bypass the NSFW stuff, but it was really complex code and stuff and I couldn't get my head around it. If anyone can help me set it up simply or know of communities for this (NSFW) stuff it would be appreciated.
AI image generators (Stable Diffusion)
- Thread starter dudedudester4vrvr4
- Start date
It depends if you want to run Stable Diffusion Locally or use an Online Service.
Localy on your PC:
I am to lazy to explain all the installation steps, but its very easy to install. Just search for a tutorial to install stable diffusion.
Something like this :
After installing Stable Infusion you can disable the nsfw filter:
(I think the optimized version doesnt even have a nsfw filter)
Online Service:
You need to use collab (or a service) that has the nsfw filter turned off.
This should probably work i guess?:
Never worked with collabs so i cant explain how to use it but i am sure there are some easy tutorials.
I think reddit Communites are getting banned that share nsfw art. I think there was a quite a large Discord with NSFW stuff so may search for that.
If you are intrested to see some creations with stable diffusion:
Fuck i know see the date of the post XD
Localy on your PC:
I am to lazy to explain all the installation steps, but its very easy to install. Just search for a tutorial to install stable diffusion.
Something like this :
You must be registered to see the links
After installing Stable Infusion you can disable the nsfw filter:
You must be registered to see the links
(I think the optimized version doesnt even have a nsfw filter)
Online Service:
You need to use collab (or a service) that has the nsfw filter turned off.
This should probably work i guess?:
You must be registered to see the links
Never worked with collabs so i cant explain how to use it but i am sure there are some easy tutorials.
I think reddit Communites are getting banned that share nsfw art. I think there was a quite a large Discord with NSFW stuff so may search for that.
If you are intrested to see some creations with stable diffusion:
You must be registered to see the links
Fuck i know see the date of the post XD
- Dec 2, 2016
- 150
- 373
NovelAI has a stable diffusion model that works for nsfw, but its limited to anime art style or general weirdness with the furry variant (as in it makes odd humans sometimes rather then furred individuals)
It also cost a subscription too.
It also cost a subscription too.
I've been playing around with NovelAi for a few days on a subscription, it's pretty impressive. It can absolutely generate high fidelity anime-style and anime-adjacent style art with a little fiddling and some patience, comparable or better than most anime art in games being posted here by indie creators.
I feel like we're headed for an inevitable tidal wave of NovelAi-powered VN games on this site.
The upside is that it'll empower a lot of talented writers and game developers who are otherwise lacking in artistic skills, but we might end up with a lot of VERY samey-looking games soon.
Feel free to post any requests, I'm still playing around with the NovelAI generator and learning its quirks.
I feel like we're headed for an inevitable tidal wave of NovelAi-powered VN games on this site.
The upside is that it'll empower a lot of talented writers and game developers who are otherwise lacking in artistic skills, but we might end up with a lot of VERY samey-looking games soon.
Feel free to post any requests, I'm still playing around with the NovelAI generator and learning its quirks.
- Dec 2, 2016
- 150
- 373
If you haven't discovered yet, NovelAI stable diffusion can also emulate a wide multitude of artistic styles and is not just limited to Anime, which is just the default.
Here's a bunch of samples for a lot of the art styles it can currently emulate.
You must be registered to see the links
One issue with stable diffusion is that it's not very good at knowing how boobs look when they're not in a bra/clothes. They usually stay squished together like they're in an invisible bra even when they're not.
Also draws nipples in the wrong locations often.
Also draws nipples in the wrong locations often.
- Dec 2, 2016
- 150
- 373
Thats just anime boob logic, all boobs are impossibly perky.
As for nipples, use the edit feature and just fix it.
basically just smear the normal skin tone over the weird nipple placement and then smear a new nipple elsewhere.
Yes.
Either specify an actor that the Ai knows about by name,
Or use Stable diffusion with Dreambooth, and give it a few pictures of the character you want to keep consistent.
Last edited:
- Dec 2, 2016
- 150
- 373
If you mean in the text version, yes.
Simply make a lorebook for the character if you want the AI to remember the character and use it occasionally, or when prompted.
For consistent and constant use, say for a main character, then simply specify as such in either the memory or author notes.
1st person would be something like, [ I am Peter. ] and 2nd person would be, [ You are Peter. ]
Spaces included and in those brackets.
- Sep 12, 2021
- 510
- 746
Har har har, I'm a genius: I'll make quick, unfinished renders and then use Stable Diffusion's style transfer to make them look good! This technique will overcome the consistency issue with AI-generated art!
AI: turn this crappy render into a baroque masterpiece painted by Diego Velasquez himself!
Stay close to my image: use 0.3 denoise.
Come on AI, this is shit!
Try again at 0.35
0.4 then
0.45?
One last time at 0.5
You win this one too, computer, but one day I will find a consistent way to create decent art...
AI: turn this crappy render into a baroque masterpiece painted by Diego Velasquez himself!
Stay close to my image: use 0.3 denoise.
Come on AI, this is shit!Try again at 0.35
0.4 then
0.45?
One last time at 0.5
You win this one too, computer, but one day I will find a consistent way to create decent art...
- Dec 2, 2016
- 150
- 373
Turn the strength down to .4-5 and the noise down to .005 or less to preserve the original image and then proceed to generate and edit as usual.
- Sep 12, 2021
- 510
- 746
Mhmh... At those settings (Guidance 0.5, Prompt Strenght 0.1 in Cmdr2 GUI) it doesn't transfer the style.
"Velasquez":
"Caravaggio":
- Dec 2, 2016
- 150
- 373
Ah, thought this was in NovelAI. Wasnt even aware of Cmdr2 GUI's existence, but a quick look shows me its of a seemingly weaker model, and a rather lacking UI.
It took me ten attempts to make this with NovelAI's model, using the initial image you posted as the base.
For NovelAI Strength controls how much the uploaded image is altered, so its the equivalent of the Guidance setting.
Noise is the fancy setting for adding detail to an image, but its real fucking finicky so you have to tweak it a lot.
Prompt strength in NovelAI is Scale, which i set to 7 for this image.
Prompt Used: "A blond haired young man ripping the clothes off a pretty girl, {by Diego Velasquez}, nsfw, detailed"
Strength: 0.4
Noise: 0.05
Steps: 50 (default)
Scale: 7
Edit: I tried about 200ish more attempts at this and I can confidently say, that stablediffusion was not trained on enough images of this format to make anything more then passable images. You'd have to also edit these images substantially to aid the AI in creating this prompt.
Last edited:
- Sep 12, 2021
- 510
- 746
Cmdr2 GUi is, well, just a GUI for SD (the good: one-click install, support for legacy hardware out of the box and a detailed styles GUI, plus the guy is always very nice and helpful. The bad: it's a rather weak and very barebones GUI and img2img in particular sucks great time - I installed Auto's WebUI at work and it's a lot more professional).
The models are whatever you put in: I'm mostly using a blend of f111 with pretty much everything else, that I found on the Unstable Diffusion Discord.
f111_0.4-NAI_0.25-gg1342_0.125-r34_e4_0.125
I never tried pure NAI: from what I gathered, as most NSFW models around, it's focused on Anime/Manga look, and I'm simply not interested in that style.
While I plaud your result as a great achievement on the road of lewdness, it's well... Anime.
Currently, I'm experimenting wildly to understand what causes the style transfer to trigger. Sometimes the AI gives me a batch in renaissance-style (not baroque, but hey, one has to be realistic about these things) with an initial image and a modern-looking batch with prompt-only (exactly the same prompt), sometimes it does the opposite.
I think bright, contrasted images force the modern look and that Prompt Strength and Guidance Scale don't have much of a word in it.
I also suspect long prompts are more likely to ignore the style.
Since I'm on CPU-only, due to my GT730, in this format an image takes several minutes (my beloved 512x256 format doesn't work for NSFW), so I can't really go full scientific in this research.
Also, GFPGAN doesn't run on my home computer: that one would've helped with small faces like those.
Raising steps beyond 20 didn't make any difference for me.
You have one more parameter than I have:
Strength -> Prompt Strenght
Scale -> Guidance Scale
Noise -> -
I'm going to ask some questions in the GUI discord...
I fully and heartly agree with you: humankind needs more NSFW training. Especially for group sex: when I try groups I mostly get sequels of Carpenter's Thing.
P.S.
I can't believe I didn't save the prompt.
It was something like (young blond man) ripping the clothes of a (pretty) girl, with (perky tits), nipples, ((beautiful face)) in a warehouse
P.P.S.
I just finished downloading a f222 blend.
This is its take on the task (Scale 4, Strenght 0.4, 20 steps):
style: detailed and intricate, pastel drawn art, by Caravaggio
- Dec 2, 2016
- 150
- 373
This is NovelAI or NAI for short, the hacked and stolen source of it, prior to any current updates.
It is anime by default, but it can emulate a wide variety of artists without any hint of anime if you set the scale right, I found 1-3 to work best for Velasquez during my 200 or so generations.
Tried to upload one, but F95 was being weird and wouldn't let me.
Also, NAI has a furry model, that despite it's name is actually just a non-anime nsfw model that has no problems making normal humans.
It also just happens to be better at making furry content then the anime one, which mostly makes monstrosities or kemonos.
And some advice.
Give up on getting every piece of the anatomy to be pleasant looking.
Instead, cut and paste the parts that worked and form a frankenstein image and then run that through the process to have the AI clean it up for a nice seamless merge.
As I did with this picture, which is 5 different generations I spliced together to form.
Futa warning.
You don't have permission to view the spoiler content.
Log in or register now.
I used a different head, face, right arm, penis and balls, and cum shot to finally land on this image.
Because the odds of all of these things working well in a single generation are pretty fucking slim, especially hands and faces.
And its a nightmare to get the AI to draw a decent penis, unless you use the furry model.
- Sep 12, 2021
- 510
- 746
OK, today at work I had some time alone with Auto's WebUI and did some experiments (standard SD 1.4 model). Even enhancing the initial image to separate the characters from the background, the result is me giving up. Again.
My conclusions are:
- There are two axes: how good the image is and how consistent it is with the original source (the latter is very important if one doesn't want a one-off image but a consistent style).
- It's a zero-sum game. Worse, the two axes meet at a point where the image both looks shitty and the characters are no longer recognizable. I admit that the image picked was a very difficult one, but that's good for a crash test.
- with CMG (Guidance?) Scale set to 12 the art styles start differentiating around at 0.3 denoising and become interesting at 0.45. At that level, the image is completely unrecognizable.
- both Velasquez and Caravaggio should have spent more time painting princes-in-rags ripping dresses off the nubile bodies of vagrant girls. That's very rude of them, but the model had clearly no idea of what I was asking in the prompt.
- with the time and effort required to get some promises of a, maybe, result in the future, I could've just made better renders and called it a day, making this a net failure.
But it works well only when you have a simple composition, like a character portrait over a simple background. Unfortunately for me, I got single-char portraits out of my system shortly after Daz's Victoria4 got out but, there are situations where the technique works even with more complex scenes.
img2img from a movie still (one is from a different batch with less denoise, but same still).
became
This is a quick, dirty but effective way to get some 2D VN scenes done, but the problem is always having the characters stay consistent.
Using more modern art styles than baroque is, intuitively, going to help sanity.
Pyro's blowjob model works very well too. As long as all you want is a closeup of a blowjob.
My conclusions are:
- There are two axes: how good the image is and how consistent it is with the original source (the latter is very important if one doesn't want a one-off image but a consistent style).
- It's a zero-sum game. Worse, the two axes meet at a point where the image both looks shitty and the characters are no longer recognizable. I admit that the image picked was a very difficult one, but that's good for a crash test.
- with CMG (Guidance?) Scale set to 12 the art styles start differentiating around at 0.3 denoising and become interesting at 0.45. At that level, the image is completely unrecognizable.
- both Velasquez and Caravaggio should have spent more time painting princes-in-rags ripping dresses off the nubile bodies of vagrant girls. That's very rude of them, but the model had clearly no idea of what I was asking in the prompt.
- with the time and effort required to get some promises of a, maybe, result in the future, I could've just made better renders and called it a day, making this a net failure.
As a veteran photoshopper, that one is my default modus operandi. Together with "clonestamp that shit away."
But it works well only when you have a simple composition, like a character portrait over a simple background. Unfortunately for me, I got single-char portraits out of my system shortly after Daz's Victoria4 got out but, there are situations where the technique works even with more complex scenes.
img2img from a movie still (one is from a different batch with less denoise, but same still).
became
This is a quick, dirty but effective way to get some 2D VN scenes done, but the problem is always having the characters stay consistent.
Using more modern art styles than baroque is, intuitively, going to help sanity.
(Yup. See above.)
Pyro's blowjob model works very well too. As long as all you want is a closeup of a blowjob.
- Dec 2, 2016
- 150
- 373
Yeah, as it currently stands Stable DIffusion and all the other AI generated art programs are simply not that great at making anything that isnt a single character or creature.
Backgrounds on the other hand can be accomplished quite well if you manage to generate a decent one and then guide the AI into hyper focusing on the character, essentially ignoring the background entirely.
Resulting in a detailed environment with a detailed character.
It's why imo, this works best as a tool for someone who actually has the ability to draw, as this is a great supplementary tool for them.
And all of us non-artistically gifted folks can simply use it for quick portraits or landscapes.
edit: Also i downloaded the Cmdr2 GUi and the NAI model to test it, and its definitely a weaker model.
But I am unclear as to why that is.
Just to be clear the GUI doesnt mean its weaker. You can find very specific models to use and get very well done renders. I have seen stable diffusion users draw stick figures and turn them into full fledged anime poses. It takes work but I its possible. I came to the forums looking to see if anyone was training a model on DAZ models. Havent tried training a model myself but I know a few people that might give it go if they can get a hold of enough raw images to process.
cmd2 SDUI is great for beginners, and has a lot of nice easy to use features, it is still very powerful. I use it to create bulk batches of images.
automatic1111 webui is amazing, so many features and gadgets that its overwhelming at first, second, and third times. But its amazing, I use it to clean up images, enlarge them, or merge models together and see what the mix gets me. Im slowly learning all the bells and whistles, it will probably be my top pick by the end of the week.
cmd2 SDUI is great for beginners, and has a lot of nice easy to use features, it is still very powerful. I use it to create bulk batches of images.
automatic1111 webui is amazing, so many features and gadgets that its overwhelming at first, second, and third times. But its amazing, I use it to clean up images, enlarge them, or merge models together and see what the mix gets me. Im slowly learning all the bells and whistles, it will probably be my top pick by the end of the week.