I've been using dreamstudio.ai for landscape/creature/character art for a few months, but I can never get something specific (I get nightmare fuel anytime I try to get something as detailed as some of you guys get, no matter what subject matter). There's been a few more AI art posts and I wanted to ask their generators about the workflows they use to make it; what programs, models, and inputs do you use? There's been some discussion in the description of specific works and in the "What do we do with AI art" thread but I was hoping to centralize that somewhat.

The best way to learn, for a lot of things AI gen included, is to copy someone who knows what they are doing then piece together why exactly what they do works.

I personally use NovelAi which costs some money, but I have heard that some people download Stable diffusion then add the 'waifu' model.

If someone is starting trying to generate hypno content I would go with a simple prompt, copy and paste:

{{hypnotized girl}}, {{{glowing pink eyes}}}, {{{hypnotized expression}}}

With quality tags on, undesired content set to none, and the model set to anime(full) the with the following negitive tags:

lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, jpeg artifacts, broken body, extra limbs, amputee, conjoined limbs, Siamese twins, incomplete, pubic hair, watermark, body horror, horror, greyscale

With that you should be able to get something good every 1 in 5 gens.

Here are some good inputs I've found.

if you want more than one girl, use {multiple girls}, the problem with this is that novel AI has a hard time with limps the more people there are.

{fantasy} makes it so that you get different girls instead of clones.

{NSFW} is usually the only thing you need to type to get nudes.

{full body} should be obvious.

The less tags there are in the data base the more {} you need to use in order to see an effect.

The way I think about a prompt is that I divide it into four parts:

Main target
Background
effects
style

If you have at least one decent tag for all of those you will usually get something nice looking when it turns out good.

Lastly, don't be afraid to edit. When I get a good gen but there is an extra limb (or anything else) you can roughly paint over the bad part with the background then use the 'enhance' feature. Lastly, I up res using a different site like www.aiseesoft.com/image-upscaler (or any other), because AI can only gen relatively small images.

Don't be afraid to experiment.
Hope that helps.

I use a variety of stable diffusion based models personally.

Obviously in the following, change () to {} if you're using NAI.

My current thing is using the following tags in my prompts:

nsfw, mind control, (empty eyes), expressionless, glowing eyes

Adding the following to unwanted content/negative prompts:

happy, sad, angry

And using a wildcard script to generate the details of the subject through a pool of descriptors I created containing things such as hair style, hair color, clothing, and pose. Eye color seems to not work very well with the glowing eyes tag which overrides it most of the time.

Additionally, I've found that adding ((sketch)) tends to generate empty eyes more consistently. (And I'm personally very fond of the sketchy lineart anyway)

i.imgur.com/LGr6F3k.png
i.imgur.com/4ZSsObW.png
i.imgur.com/XqZav7H.png
i.imgur.com/XUFp6jB.png
i.imgur.com/nEEznjG.png

Using NAI, I've found that generating the original image at 512x512 and then "upscaling" it using img2img provides the best results. The rest is just a lot of trial and error regenerating certain parts.

For example, I generated this image using the following prompts (sorry, don't have the negatives on hand):
i.imgur.com/a1t6g0v.png
masterpiece, best quality, high definition, anatomically correct, nsfw, {{no clothes}}, {topless}, {bottomless}, medium breasts

I then re-generated certain parts (hands, expression, general anatomy) with img2img at 512x512. Once I found what I liked, I generated the final 1024x1024 image a few times until I got this:
i.imgur.com/21SeaSB.png

Then once I'm done that, I would do some touchups manually in photoshop to fix easy things like the back. This can be done by drawing things out yourself, or frankensteinng multiple attempts together... usually a mix of both.
i.imgur.com/1bh0RhZ.png

There's still some things in my image that I'd fix if I was planning on posting it like the eyes, but that's been my general pipeline so far.

I'll copy a post on the subject I posted on another forum.

If you want to get into AI porn you have basically three options:
1) Use a paid service. I think novelAI is pretty significantly the best porn one. Pro: It requires basically no setup. Con: Its the most expensive and has the least flexibility.
2) Set Stable Diffusion up on your own computer. Pro: Everything is local, and you have 100% complete control. You are a golden god at the front of AI research. Con: Setup is the most difficult, it requires pretty good NVIDIA video card (you really need at least 6GB VRAM), and its theoretically possible to screw your computer up. Also you can't use your computer for games or anything while its running.
3) Set Stable Diffusion up on a cloud service. Pro: You can use your own computer like normal while generating images and can even generate them on your phone, you also have 100% control over image generation. You are a AI research specialist utilizing the infinite power of the cloud. Con: Setup is still a pain, you want to pay a couple buck a month, a couple extra hoops to jump through every time you start it up.
0) Use a free service somewhere. Don't even bother, they are going to be trash compared what you can do with 1-3. Generating these images is energy and equipment expensive. The good stuff isn't free.

(Note that you probably want to use the NovelAI model anyways if you are using Stable Diffusion).

I'm personally using #3 with a paid paperspace account ($8 per month). The service has some issues sometimes, but the difference between paid and free is pretty significant, and getting a 16GB video card to play with is great. Note that you *can* use the cloud stuff without paying, its just more difficult and has more hoops to jump through. The guide I used to set it up is here by the way.
rentry.org/865dy

Honestly, all this porn image generation stuff is fucking amazing and feels intoxicating. There are so many tools and so many settings and so many different things you can do. I've messed with it for like 40(?) hours the last two weeks and I've only scratched the surface.

If you want more info the 4chan Diffusion General threads on the subject (on /h/ or /g/) have a massive amounts of links and information and are the only way I could figure any of this stuff out. The stuff there also has a ton of info on prompts and stuff that works.

As far as actually generating stuff goes you want to make sure you are including a bunch of negative tags every time to avoid bad stuff happening. Yes, all of it, every single time.
Posted below is a example of default negative tags:
Negative tags:
lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, deformed,disfigured,poorly drawn face,mutation,mutated,extra limb,poorly drawn hands,extra hands,missing limb,floating limbs,disconnected limbs,malformed hands,out of focus,long neck,long body


If there is anything other negatives you don't want for a particular prompt obviously add it for that prompt.


Generating hypnosis stuff in particular is tough in general, AFAIK none of the models are particularly good at it although you can massage it using the right tags or inpainting.

As a femdom hypno conniseur, I find getting content to be difficult as hell, at best on occasion I can get the girl to hold a watch, but thats about it, also a shame i'm a spiral eyes guy all the way, because the glowing pink eyes dont really do it for me, so I guess what i'm trying to ask here is if anyone knows a consistent way to get some decent spiral eyes on these ai generated girls

At least right now, you're gonna be hard-pressed trying to get that level of detail without manually manipping it yourself. Interactions between people and other people/objects is also pretty difficult to do. Your best bet is to draw a rough scene and hope that you can guide the AI to what you want to generate.

Thanks, I just tried getting into it, this seems like good tags I can employ