dusty stu
Well-Known Member
- Jan 24, 2018
- 1,778
- 1,654
AI generated art experiences for NSFW content
- Thread starter bobjones9792
- Start date
baloneysammich
Active Member
- Jun 3, 2017
- 998
- 1,544
Some tips/answers to posts i didnt see or read wrong before:
To use the same characterlock the Seed. You'll get the same character or one close and through iterations you'll get a lot of that character
The checkpoint/safetensors ENTIRELY determines what the ai can produce. Babes makes OUTSTANDING cartoonish characters like the readhead i posted before. PYROS is a good bj model. There are more porn models being produced daily.
Civitai dot com is an excellent place to get models (checkpoints/safetensors).
If you find an ai art picture you like you can pull it's settings, seed, prompt and negative prompt in Automatic1111 in the PNGinfo tab.
You can combine models (checkpoints/safetensors) in checkpoint merger. For instance the BABES model doesnt have any sex images in it. So you combine it with pyros or some other porn model and it can. You can combine 3 or more models with a lot of variation. This takes shockingly little time (at least on my system).
You can train your own models or embeddings. I havent done this. It is extremely time intensive (especially models). Embeddings are a little different for instance you could take 50 pictures of some celeb and create an embedding then use it with a defined keyword to include that celeb in images you create. Again i havent done this as it is very time intensive.
If you just generate 50 images of a prompt you'll end up getting 50 similar images. If youre searching for something good it is better to generate 2 or 5 or 10 at a time til you find one you like then lock the seeds.
Save good prompts that work. Change one word or sentence at a time and produce a few images and note the difference. It can be a much different change than you expected.
Use descriptive prompts but not overly long ones. Split the prompt up. ie beautiful blond girl in a fuzzy sweater sitting in a chair in a room with carpet..... the ai cant handle long complex prompts it isnt human. A better way would be "a beautiful blond girl in a fuzzy sweater, 1girl sitting in a chair, room with carpet" etc.
Sampling method matters but it's hard to determine how. I use ddim as a standard and try others.
You can search for extensions and enter their url in the Extensions tab and it will auto install them. Image Browser is a great one.
I suggest clicking the EXTRA box on the txt2img page and raising the variation strength.. for example to .1 or .17 etc.
You can train your own models/checkpoints.
on image size and sampling steps. Keep these low at first for rapid generation until you find the character/etc you like .. THEN run the same prompt with higher sampling steps.
Dont stress image size. You can resize it in the extras tab and do a lot of other things such as get a wider angle etc. (i havent used this much). As with all graphics if you double the size of the image you quadruple the amount of memory required. 512 stressed my 1070 hard. I think the largest ive gone with my 3080ti is like 700 x 700. For raw 4k images we're going to need gpu's with more ram than our SSD's have. You're artists.. worst case you toss it in photoshop and upsize it etc etc.
Stable diffusion is NOT daz maya etc. You do not have precise control. It absolutely will not create that exact image or model you have in your mind. It is more like finding something you like and guiding it towards creating something you like more with that character. Ive used it to create some absolutely AMAZING art. But if i want precise control of a scene.. back to daz.
Note if DAZ owners are smart theyll integrate an ai into it... imagine describing a video scene and having the ai auto pose it for you......
Happy to help anyone im probably more obsessed with stable diffusion than i have been with daz and that is a LOT
Im not an expert but im a bit obsessed with it so im happy to help anyone who asks if i can.
To use the same characterlock the Seed. You'll get the same character or one close and through iterations you'll get a lot of that character
The checkpoint/safetensors ENTIRELY determines what the ai can produce. Babes makes OUTSTANDING cartoonish characters like the readhead i posted before. PYROS is a good bj model. There are more porn models being produced daily.
Civitai dot com is an excellent place to get models (checkpoints/safetensors).
If you find an ai art picture you like you can pull it's settings, seed, prompt and negative prompt in Automatic1111 in the PNGinfo tab.
You can combine models (checkpoints/safetensors) in checkpoint merger. For instance the BABES model doesnt have any sex images in it. So you combine it with pyros or some other porn model and it can. You can combine 3 or more models with a lot of variation. This takes shockingly little time (at least on my system).
You can train your own models or embeddings. I havent done this. It is extremely time intensive (especially models). Embeddings are a little different for instance you could take 50 pictures of some celeb and create an embedding then use it with a defined keyword to include that celeb in images you create. Again i havent done this as it is very time intensive.
If you just generate 50 images of a prompt you'll end up getting 50 similar images. If youre searching for something good it is better to generate 2 or 5 or 10 at a time til you find one you like then lock the seeds.
Save good prompts that work. Change one word or sentence at a time and produce a few images and note the difference. It can be a much different change than you expected.
Use descriptive prompts but not overly long ones. Split the prompt up. ie beautiful blond girl in a fuzzy sweater sitting in a chair in a room with carpet..... the ai cant handle long complex prompts it isnt human. A better way would be "a beautiful blond girl in a fuzzy sweater, 1girl sitting in a chair, room with carpet" etc.
Sampling method matters but it's hard to determine how. I use ddim as a standard and try others.
You can search for extensions and enter their url in the Extensions tab and it will auto install them. Image Browser is a great one.
I suggest clicking the EXTRA box on the txt2img page and raising the variation strength.. for example to .1 or .17 etc.
You can train your own models/checkpoints.
on image size and sampling steps. Keep these low at first for rapid generation until you find the character/etc you like .. THEN run the same prompt with higher sampling steps.
Dont stress image size. You can resize it in the extras tab and do a lot of other things such as get a wider angle etc. (i havent used this much). As with all graphics if you double the size of the image you quadruple the amount of memory required. 512 stressed my 1070 hard. I think the largest ive gone with my 3080ti is like 700 x 700. For raw 4k images we're going to need gpu's with more ram than our SSD's have. You're artists.. worst case you toss it in photoshop and upsize it etc etc.
Stable diffusion is NOT daz maya etc. You do not have precise control. It absolutely will not create that exact image or model you have in your mind. It is more like finding something you like and guiding it towards creating something you like more with that character. Ive used it to create some absolutely AMAZING art. But if i want precise control of a scene.. back to daz.
Note if DAZ owners are smart theyll integrate an ai into it... imagine describing a video scene and having the ai auto pose it for you......
Happy to help anyone im probably more obsessed with stable diffusion than i have been with daz and that is a LOT
Im not an expert but im a bit obsessed with it so im happy to help anyone who asks if i can.
You should check out ControlNet extension if you haven't already, it's capable of providing exactly that - precise control. Hell, on SD subreddit there were examples of people using it to dynamically change the lighting in the image.
dusty stu
Well-Known Member
- Jan 24, 2018
- 1,778
- 1,654
How it's coming after 3 weeks of practice:
Models:
Stable diffusion 1.5
always start 512x756 res
xyz plot to test settings is the best
IRL:
Elegance -or- Liberty
20-25 steps, DPM++ 2m Karras
1.5x highres fix, r-ersgan4x
Anime:
OrangemixHardcore -or- Protogen.Anime.2.2
20 steps, DPM++ 2m Karras -or- 30 steps Euler A
1.5x highres fix, r-ersgan4x anime6b
my favorite LORAs:
hyperpreg (more consistently pregnant, weight at like .2 for 9-months pregnant. works with irl and anime)
thatsmybush (natrual pubic hair, .8. works with any model)
sagbobz (stretch marks & droopy tits, like .1 or .2. irl models mostly. Anime just doesn't do sag!)
Biggest disappointments:
URPM, takes too much work to make sex look good. Sex scence in general are way too much trouble to be worth it.
any MLP model to make ponies, just looks like trash
trying to use anime character LORAs in IRL models (Hinata Hyuga etc)
trying to start with any res other than 512x756
waiting 1-3 hours for the highres steps to finish, only for some prompt to go wonky, wasting huge amounts of time
my first overnight 'highres fix' success:
Models:
Stable diffusion 1.5
always start 512x756 res
xyz plot to test settings is the best
IRL:
Elegance -or- Liberty
20-25 steps, DPM++ 2m Karras
1.5x highres fix, r-ersgan4x
Anime:
OrangemixHardcore -or- Protogen.Anime.2.2
20 steps, DPM++ 2m Karras -or- 30 steps Euler A
1.5x highres fix, r-ersgan4x anime6b
my favorite LORAs:
hyperpreg (more consistently pregnant, weight at like .2 for 9-months pregnant. works with irl and anime)
thatsmybush (natrual pubic hair, .8. works with any model)
sagbobz (stretch marks & droopy tits, like .1 or .2. irl models mostly. Anime just doesn't do sag!)
Biggest disappointments:
URPM, takes too much work to make sex look good. Sex scence in general are way too much trouble to be worth it.
any MLP model to make ponies, just looks like trash
trying to use anime character LORAs in IRL models (Hinata Hyuga etc)
trying to start with any res other than 512x756
waiting 1-3 hours for the highres steps to finish, only for some prompt to go wonky, wasting huge amounts of time
my first overnight 'highres fix' success:
Last edited:
I agree on most of it.
I'm not much into anime, more interested by realism.
I haven't tried Elegance but I really like Liberty, it can produce very good looking and different pictures with imaginative prompts and with being very very sober with negative prompt. Very versatile model and finally easy to use if you prefer "natural language" to a list of words, and it works also with simple prompts (a descriptive phrase plus a few words separated by comma and I found that the less is better)
I switch sometime to delibarate, chilloutmix, protogen for anime look but pretty realistic, but then I have to swith to comma separated words.
Anyway better forget about sex scene for now. we agree, but for erotic realistic pictures it works really fine.
Did you tried embeddings/textual inversion ?
Personnaly none really convinced me, they give more or less always the same faces / attitude.
I tried to make my own based on a big set of pictures but it takes a huge time, you need to tag each picture witth the correct key word, so imagine ~100 pics, then strat the training that is taking hours and the result is disappointing .... Well some people succeed in it so I guess it can work too.
Did you try Controlnet ?
You must be registered to see the links
UI for Automative 1111:
You must be registered to see the links
It looks like an alternative to time consuming trainings. I had promising relsuts with it, promising means bad but approaching
Minimalistic documentation on how to use it but I gave you the 2 links to bookmark
baloneysammich
Active Member
- Jun 3, 2017
- 998
- 1,544
Are you on the Unstable Diffusion discord? They have a photorealism channel and a lot of incredible stuff is posted practically every few minutes. One really talented creator is GalaxyTimeMachine who created these:
You don't have permission to view the spoiler content.
Log in or register now.
You must be registered to see the links
he made. It does lewds too, I just really liked these.
dusty stu
Well-Known Member
- Jan 24, 2018
- 1,778
- 1,654
That looks scary 0.o
I think I'll just stick with making 1 person at a time for now
Bravo82Newtron
Newbie
- Oct 8, 2021
- 28
- 10
Sucks that my python ain't working. But I have a question. Do you guys have any suggestions on the best software for ai generating for hentai? Like, a list would help. Please and thanks!
- Mar 12, 2023
- 173
- 603
The sheer devotion of this medium is all the proof I need that AI porn will be the future of pornography.
There's just too many perverts working to make it so for it be otherwise.
There's just too many perverts working to make it so for it be otherwise.
dusty stu
Well-Known Member
- Jan 24, 2018
- 1,778
- 1,654
4 Month update:
Got a better GPU (huge improvement over using CPU to generate!)
Started training some of my own Loras. They look not half bad.
got a few saved prompts with my favorite models to really speed up generation.
Finally getting the hang of controlnet. It's god-tier good!
Got a better GPU (huge improvement over using CPU to generate!)
Started training some of my own Loras. They look not half bad.
got a few saved prompts with my favorite models to really speed up generation.
Finally getting the hang of controlnet. It's god-tier good!
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
VegitoHlove
Member
- Apr 27, 2018
- 376
- 953
Some of my works. My favorite checkpoints are dreamshaper and GalenaBlend. Hands are a bit fucky, but I've seen MUCH worse. These are made at different times during my playing around with AI. Just got Controlnet and Stylepile. So I'll learn how to do that. Also I love redheads.
