AI generated art experiences for NSFW content

[fitgirlbestgirl]

In theory, AI and porn could have mind blowing possibilities: no longer praying and hoping that your favorite smut author/artist will issue something which tickles your darkest fancies and no more worrying lest someone finds your shady browsing history. The AI will cater to your darkest desires with no moral boundaries, just hop on its back and ride the vastness of Smutland to wherever you heart desires.

As long as "AI" remains stuck in a "type in an autistic prompt to produce a somewhat passable result" paradigm there is no danger of it actually replacing competent writers/artists. Even assuming it gets a lot better at that, sometimes you don't know what "tickles your darkest fancies" until you actually see it.

 

[Sleepyguylol]

Even assuming it gets a lot better at that, sometimes you don't know what "tickles your darkest fancies" until you actually see it.

This part I absolutely agree with. There are some kinks that initially scoffed at... until i played some VNs and I realized that its not so bad lol

 

[lancelotdulak]

As long as "AI" remains stuck in a "type in an autistic prompt to produce a somewhat passable result" paradigm there is no danger of it actually replacing competent writers/artists. Even assuming it gets a lot better at that, sometimes you don't know what "tickles your darkest fancies" until you actually see it.

Did you look at the art i posted? I have a lot more of it. And every one of those images was better than 98% of what is posted on f95zone

 

[Earliestbird]

AI art can get a little funky, but it's nothing that personal editing can't fix. I already have a 2gb folder of generated art that is quite beautiful.

 

[3D Reaver]

AI art can get a little funky, but it's nothing that personal editing can't fix. I already have a 2gb folder of generated art that is quite beautiful.

View attachment 2351325 View attachment 2351326

I'd rather have a character be objectively worse made but be consistent in appearance. The constant shift in look is distracting and you can't really be too specific with what you want. You will more than likely have to work around what images you get instead of the other way around.

I think AI is great for generating backgrounds or for lets say art pieces if one of the characters is an artist or something like that.

 

[Earliestbird]

I'd rather have a character be objectively worse made but be consistent in appearance. The constant shift in look is distracting and you can't really be too specific with what you want. You will more than likely have to work around what images you get instead of the other way around.

I think AI is great for generating backgrounds or for lets say art pieces if one of the characters is an artist or something like that.

You can get consistent imagery and style - you just have to use a singular model to achieve that. Which is 100% possible in the future. Bur for now people are still making brand new models that surpass the older ones on many levels, so most people keep jumping from one to the next.

A year or two from now, and there will be no issue generating style-consistent images.

 

[dusty stu]

So.. anyway... started programming 40 years ago... because i wanted to do ai... anyway... i have stable diffusion with automatic1111 installed. About 6 checkpoints (thing MASSIVE databases of pictures). Ive generated some pretty sweet erotic stuff.. ill post a few. But saying "it isnt replacing porn anytime soon etc" when youve played with it 2 hours is like saying you cant create photorealistic content with daz after an hour.. or a week. Oh also there are companies actually starting to do it.

Guided creation is harder and requires more work and expertise. You use tools that people just tossing in prompts dont even know exist. And the KEY is the checkpoint. The big checkpoints are built from pictures that specifically exclude nsfw. We'll need people doing a huge amount of work building checkpoints for nsfw .. which is already happening. BTW you folks thinking about investing money into these big corporate ai's... dont. Theyre all hyper obsessed about excluding nsfw and any content they dont like (a lot of which we wouldnt like). Porn built the internet folks. Those companies are going to go byebye (especially as their creations will be based on others work and.. actionable). Each of these took < 20 seconds with a 3080ti..
I dont like the generated sex photos yet.. i need to do more work and need better checkpoints etc

Damn there be a lot of fake tits in the training data XD
Unless you told the ai to make fake tits? In which case, why?

 

[dusty stu]

So.. anyway... started programming 40 years ago... because i wanted to do ai... anyway... i have stable diffusion with automatic1111 installed. About 6 checkpoints (thing MASSIVE databases of pictures). Ive generated some pretty sweet erotic stuff.. ill post a few. But saying "it isnt replacing porn anytime soon etc" when youve played with it 2 hours is like saying you cant create photorealistic content with daz after an hour.. or a week. Oh also there are companies actually starting to do it.

Guided creation is harder and requires more work and expertise. You use tools that people just tossing in prompts dont even know exist. And the KEY is the checkpoint. The big checkpoints are built from pictures that specifically exclude nsfw. We'll need people doing a huge amount of work building checkpoints for nsfw .. which is already happening. BTW you folks thinking about investing money into these big corporate ai's... dont. Theyre all hyper obsessed about excluding nsfw and any content they dont like (a lot of which we wouldnt like). Porn built the internet folks. Those companies are going to go byebye (especially as their creations will be based on others work and.. actionable). Each of these took < 20 seconds with a 3080ti..
I dont like the generated sex photos yet.. i need to do more work and need better checkpoints etc

Actually if you could give a few tips on how to prompt the AI better that would be awesome.
I'm using:
automatic1111 with uberRealisticPornMerge_urpmv12

Spoiler: example of my inputs

You don't have permission to view the spoiler content. Log in or register now.

Spoiler: what i made last night

You don't have permission to view the spoiler content. Log in or register now.

Particularly I want to avoid the Eldrich abominations of extra limbs and people sharing limbs.
@fitgirlbestgirl
You'll probably get a kick out of some of the dumb shit my attempts have generated so far XD

 

[Zuburban]

Getting good-looking realistic porn with pure prompting, without img2img or inpainting, is still hard, not impossible though. Now hentai - that's a different story, already a lot of great models\embeddings\LORA's to get great looking hentai pictures with simply prompting. As an example - I'v done this today in a span of half an hour, and it only took so long cause I'm still rocking ol'reliable GTX950 with 2GB VRAM. And that's doing bare minimum by prompting, not using inpainting to fix minor issues or outpainting to add more stuff.

Spoiler: Pictures

You don't have permission to view the spoiler content. Log in or register now.

Actually if you could give a few tips on how to prompt the AI better that would be awesome.
I'm using:
automatic1111 with uberRealisticPornMerge_urpmv12

Too many steps, prompt size and CFG scale a bit too big for URPM, try going less specific first, also would recommend 512x512 resolution for it. Once you get something good - grab the seed and try adding removed stuff gradually on top of it, or inpaint it. Highrez fix should be helpful in battling abominations that appear when you generate on higher resolutions.

 

[dusty stu]

Once you get something good - grab the seed and try adding removed stuff gradually on top of it, or inpaint it.

Can you explain that?
I've new to this and havent messed around with seeds, img2img, or inpainting yet.What is outpainting?

It's super Facinating that higher res would make the abominations. :thinking:

I'll eventually switch to a Hentai model, but I'm enjoying making human abominations for now. If I'm using automatic1111, is there a good model to go with that for hentai?

 

[Zuburban]

Can you explain that?
I've new to this and havent messed around with seeds, img2img, or inpainting yet.What is outpainting?

It's super Facinating that higher res would make the abominations. :thinking:

I'll eventually switch to a Hentai model, but I'm enjoying making human abominations for now. If I'm using automatic1111, is there a good model to go with that for hentai?

Each picture you generate has a seed, it's random one unless specified. When you generate with the same seed, prompt and other settings you will produce same image as before. So once you generate a picture with composition you like - you can use it's seed and start adding things to prompt to enhance it. Use PNG info to straight up copy all the settings from previous generations.

Spoiler: Visual cue

You don't have permission to view the spoiler content. Log in or register now.

Inpaint is "smart" photoshop basically, you can use it to change\add\remove stuff from certain parts of images. It takes like a minute to fix faces and hands with it for example. Too much to write about it, so just google a guide, plenty of them around at this point.

Outpaint expands picture - you can for example generate a full body portrait from a face photo in same style. Google "a1111 openoutpaint" for a neat extension that makes it very easy, or use "script" tab on bottom of img2img.

As for hentai model\embeddings - search Master_Corneo on Civitai site, his stuff is easiest to start from, in my opinion.

 

[dusty stu]

Each picture you generate has a seed, it's random one unless specified. When you generate with the same seed, prompt and other settings you will produce same image as before. So once you generate a picture with composition you like - you can use it's seed and start adding things to prompt to enhance it. Use PNG info to straight up copy all the settings from previous generations.

Spoiler: Visual cue

You don't have permission to view the spoiler content. Log in or register now.

Inpaint is "smart" photoshop basically, you can use it to change\add\remove stuff from certain parts of images. It takes like a minute to fix faces and hands with it for example. Too much to write about it, so just google a guide, plenty of them around at this point.

Outpaint expands picture - you can for example generate a full body portrait from a face photo in same style. Google "a1111 openoutpaint" for a neat extension that makes it very easy, or use "script" tab on bottom of img2img.

As for hentai model\embeddings - search Master_Corneo on Civitai site, his stuff is easiest to start from, in my opinion.

Ok, got any tips for improving these bad boys?

Spoiler: pls help

You don't have permission to view the spoiler content. Log in or register now.

My settings:

 

[Zuburban]

Ok, got any tips for improving these bad boys?

Spoiler: pls help

You don't have permission to view the spoiler content. Log in or register now.

My settings:
View attachment 2358250

Gotta use hires fix or stick to basic resolutions (512x512, 512x768, 768x512). Try different samplers, euler a can be very wild, also keep prompts simple until you get the hang of it, you can always add more stuff later.

You using embedding for pov missionary scene and then prompt for mating press, not gonna work all that well. Mating press in general is one of the harder positions to get, altough I think I saw a LORA for it on Civitai, you can search there.

Edit - Did a quick try with simple prompt of what I think you wanted, start like this and then add more things into prompt\inpaint what you need.


Also - I believe you need to change CLIP Skip setting to 2 when you using anime models, at least I do. It's in settings -Stable diffusion tab, at the bottom.

 

[dusty stu]

Gotta use hires fix or stick to basic resolutions (512x512, 512x768, 768x512). Try different samplers, euler a can be very wild, also keep prompts simple until you get the hang of it, you can always add more stuff later.

You using embedding for pov missionary scene and then prompt for mating press, not gonna work all that well. Mating press in general is one of the harder positions to get, altough I think I saw a LORA for it on Civitai, you can search there.

Edit - Did a quick try with simple prompt of what I think you wanted, start like this and then add more things into prompt\inpaint what you need.
View attachment 2358762

Also - I believe you need to change CLIP Skip setting to 2 when you using anime models, at least I do. It's in settings -Stable diffusion tab, at the bottom.

So what would be the differance of doing 6 batches of 2 -versus- 2 batches of 6 ? Would one be faster? Would the one with more batches have more variety?

 

[dusty stu]

Gotta use hires fix or stick to basic resolutions (512x512, 512x768, 768x512). Try different samplers, euler a can be very wild, also keep prompts simple until you get the hang of it, you can always add more stuff later.

You using embedding for pov missionary scene and then prompt for mating press, not gonna work all that well. Mating press in general is one of the harder positions to get, altough I think I saw a LORA for it on Civitai, you can search there.

Edit - Did a quick try with simple prompt of what I think you wanted, start like this and then add more things into prompt\inpaint what you need.
View attachment 2358762

Also - I believe you need to change CLIP Skip setting to 2 when you using anime models, at least I do. It's in settings -Stable diffusion tab, at the bottom.

Also my pics are now tending to come out 'grainy' 'murkey' or 'washed out'. Any ideas how to improve?

Grainy:

Spoiler

You don't have permission to view the spoiler content. Log in or register now.

My settings (abyssorange x2, ran in through img2img a few rounds):

Spoiler

You don't have permission to view the spoiler content. Log in or register now.

 

[baloneysammich]

So what would be the differance of doing 6 batches of 2 -versus- 2 batches of 6 ? Would one be faster? Would the one with more batches have more variety?

I believe 2 batches of 6 would be faster. But it also increases VRAM usage for the process as a whole. IIRC, for me with only 6G VRAM, generation would often crash if I went above batch size of 4 when generating 512x768 images.

Whether you do 6 batches of 2 or 2 batches of 6, the only difference is based on the seed used for each image (based on your Seed setting of -1). IOW, batch size doesn't affect variety.

Also my pics are now tending to come out 'grainy' 'murkey' or 'washed out'. Any ideas how to improve?

I recall a situation where someone solved a similar problem by manually specifying the model's .vae file. But it was a different model and IIRC txt2img, so it may be completely unrelated.

 

[dusty stu]

I believe 2 batches of 6 would be faster. But it also increases VRAM usage for the process as a whole. IIRC, for me with only 6G VRAM, generation would often crash if I went above batch size of 4 when generating 512x768 images.

Whether you do 6 batches of 2 or 2 batches of 6, the only difference is based on the seed used for each image (based on your Seed setting of -1). IOW, batch size doesn't affect variety.



I recall a situation where someone solved a similar problem by manually specifying the model's .vae file. But it was a different model and IIRC txt2img, so it may be completely unrelated.

Generally speaking, what are the effects of these prompts:

Prompt order:
Prompts that are near the front are weighted more heavily?

1. standing, eating, ... etc... blue hair
2. blue hair, standing, eating ... etc...

Would 2 have more blue hair than 1?


| operator:
(It essentially mixes the two things together to make a hybrid?)

1. running, (cat|dog), smiling
2. running, (cat|dog:1.5), smiling
3. running, |cat|dog|, smiling
4. running, (red hair|blue hair), man

Translations for the above | :

1. running, cat-dog hybrid, smiling
2. running, cat-dog hybrid favoring but favor dog by 150% when mixing them together, smiling
3. running, cat-dog hybrid, smiling
4. running, purple hair, man

Does that look correct?



Weighting:

(prompt):

1. ..., (blue hair:1.5), ... -> weight blue hair 150%
2. ..., (blue hair: 1.5), ... -> weight blue hair 150%
3. ..., blue hair:1.5, ... -> weight blue hair 150%
4. ..., (blue hair:0.5), ... -> weight blue hair 50%

(negative prompt):

1. ..., (blue hair: 2.0), ... -> decrease blue hair by 200%, I really don't want blue hair

Spaces:

1. ..., blue hair, ...
2. ..., blue_hair, ...
3. ..., blue-hair, ...

These are essentially the same prompt?



Do all those examples look right? If not, how would they be correctly typed?

 

[baloneysammich]

I haven't been following things closely for a couple months, so take all this with a grain of salt:

Prompt order

There's no exact science. Start simple and and tweak it bit by bit. Or borrow a prompt from a creator or series of images you like and tweak that.

You must be registered to see the links

recommended grouping similar prompt terms together in a particular order, for example one user's "formula" was:

the picture's quality
the picture's subject
their physical appearance
their emotion
their clothing
their pose
the picture's setting

But that's pretty old advice by now, and even months ago I saw some great stuff made with a relatively haphazard ordering of terms.

| operator

Well, there's

You must be registered to see the links

native function of the operator, but I never used it and don't really see any value in it.

OTOH, with a script or extension (such as

You must be registered to see the links

) the operator can modify the prompt on the fly. For example, with Dynamic Prompts and a prompt like running, {cat|dog}, smiling, you'd get some images made with prompt running, cat, smiling and some made with running, dog, smiling. See the extension Readme for more details.

Weighting

I didn't do much with weighting so I won't try to explain it. See

You must be registered to see the links

. I do seem to recall there was some uncertainty about how weights worked in negative prompts, but don't quote me on that.

Spaces

My best advice is to keep things simple at first, then experiment. So create some images using "blue hair" in the prompt, then use the same seeds to re-create the images with "blue hair" changed to "blue-hair" or "blue_hair" and see what happens. I believe it will also depend on the model, based on how the images used to produce it were tagged.

 

[dusty stu]

I haven't been following things closely for a couple months, so take all this with a grain of salt:

Prompt order

There's no exact science. Start simple and and tweak it bit by bit. Or borrow a prompt from a creator or series of images you like and tweak that.

You must be registered to see the links

recommended grouping similar prompt terms together in a particular order, for example one user's "formula" was:


But that's pretty old advice by now, and even months ago I saw some great stuff made with a relatively haphazard ordering of terms.

| operator

Well, there's

You must be registered to see the links

native function of the operator, but I never used it and don't really see any value in it.

OTOH, with a script or extension (such as

You must be registered to see the links

) the operator can modify the prompt on the fly. For example, with Dynamic Prompts and a prompt like running, {cat|dog}, smiling, you'd get some images made with prompt running, cat, smiling and some made with running, dog, smiling. See the extension Readme for more details.

Weighting

I didn't do much with weighting so I won't try to explain it. See

You must be registered to see the links

. I do seem to recall there was some uncertainty about how weights worked in negative prompts, but don't quote me on that.

Spaces

My best advice is to keep things simple at first, then experiment. So create some images using "blue hair" in the prompt, then use the same seeds to re-create the images with "blue hair" changed to "blue-hair" or "blue_hair" and see what happens. I believe it will also depend on the model, based on how the images used to produce it were tagged.

Generally speaking, how do sampling steps work, and how do they affect the overall quality of the outputs?

I obviously know that more steps = more time.
I noticed that steps >55 can produce a lot of nonsense / abominations.
People recommend 35-50 steps a lot; but why, though?

 

[baloneysammich]

Generally speaking, how do sampling steps work, and how do they affect the overall quality of the outputs?

If you don't know, images start out as "noise" basically. Aside from that initial noise, in each step the AI uses the output of the last step and "does its thing" to try to create a modified output matching its interpretation of your prompt and other settings.

I made a couple grids (using X/Y plot script) to give an idea of the progression. This is really inefficient because each image is made from scratch, but I don't know how to direct the tool to save the product of intermediate steps. Anyway, prompt, seed, etc are the same for all images, the only thing that changes is the number of steps:

Spoiler

You don't have permission to view the spoiler content. Log in or register now.

You can kind of follow the progression as the number of steps increases.

From what I understand, the ideal number or range of steps varies based on the Sampling Method.

As to why things get weird with "too many" steps, I can't really give a technical explanation. My understanding is that once it reaches a certain point, further "tweaks" start to compound and spin out of control. I think of it like carving something from a block of stone or wood. At some point you're only gonna make things worse by removing more material. Whereas here, each step reduces noise/randomness/entropy.