AI-generated images thread

[Bartholomew Gallacher]

Each of the images are amazing! But this one is by far my favorite... especially from the neck up. Even the mouth/eyebrows remind me of a skin that I'd used in SL for a long time before mesh obsoleted it.

Thanks, it's always a nice exercise to work on such things to get better. Here's some quick introduction on Controlnet.

 

[Cat Fratica]

Nicki Minaj - One generation of dozens - ALL covered in pink roses!?
Nicki Minij - Without flowers, or any resemblance to Nicki Mibaj...

I'm confused.

 

[Noodles]

I wanted to branch out from Automatic1111's tool, and have not tried Auto Diffusion yet, but I did decide to try this one out.

AI Runner with Stable Diffusion | AI Art Editor and Custom Chatbots by Capsize Games

AI Art using Stable Diffusion on your own PC

capsizegames.itch.io

I noticed it has that, controlnet option, so I decided to mess with it with some SL screen shots. With a weird bran fart moment. Unfortunately, it doesn't look like this tool auto saves results like Automatic1111 does, so I don't have the images.

The first one I did, it came out, ok-ish, but the conversion was too old. So I just, ran it again, adding "Young" to the prompt. That ended up looking too young. I also decided to add "smiling", but it looked weird after a couple of runs, I took it out.

The other thing is, the results I am getting, are just becoming.... More and more kind of grosely ugly. Like, I wasn't expecting super model results, but I was getting more and more towards troll. I even started adding negative prompts like "crooked teeth" to no avail.

Then I realized..... Each run, wasn't building on my original base image. It was building, on the last run. So my old slightly wrinkled woman, became a weird wrinkled child, then a weird wrinkled woman with an odd smile, which just kept becoming weirder and wierder and it kept making the teeth just, worse and worse.

Whoops.

Also, I feel like there must be something with NSFW content in this tool, it has a toggle button, its not clear which was is on or off, but I tried bothz but both times I tried converting an SL nude shot, it worked, ish, but the result was this weird pencil drawing. Like this was the creator's idea of a funny joke. It was based on the original image, because it had a similar pose.

 

[Jopsy Pendragon]

, I feel like there must be something with NSFW content in this tool,

I tried generating two guys wrestling, curious if it would actually generate something vaguely nsfw... but what I got was straight out of the deepest heart of 'Body Horror' land.

Disturbingly conjoined bodies, fused or juxtaposed in ways that looked just real enough to really drive home the 'wrongness' of it. Shoulders and torsos had faces as if they were heads. Necks that turned into a leg or torso instead of a head. Hands and feet (deformed of course) randomly swapped. Heads growing out of heads. -So- very many extra limbs some of which connected one shoulder or hip to another. Almost like it was trying to render a gnarly tree instead of a two of people.

 

[Argent Stonecutter]

I just fed it some "archy and mehitabel" and this is what I got for the bit starting "dance mehitabel dance..."

Creepy.

 

[Noodles]

I tried generating two guys wrestling, curious if it would actually generate something vaguely nsfw... but what I got was straight out of the deepest heart of 'Body Horror' land.

Disturbingly conjoined bodies, fused or juxtaposed in ways that looked just real enough to really drive home the 'wrongness' of it. Shoulders and torsos had faces as if they were heads. Necks that turned into a leg or torso instead of a head. Hands and feet (deformed of course) randomly swapped. Heads growing out of heads. -So- very many extra limbs some of which connected one shoulder or hip to another. Almost like it was trying to render a gnarly tree instead of a two of people.

Oh AI can absolitely not handle two people interacting in ANY way.

You can sometimes get two people next to each other. Often they will be clones of each other.

Don't ask me how I know, but AI also does not understand how a penis works, at all. For ANY function including "just being there."

 

[Jopsy Pendragon]

Don't ask me how I know, but AI also does not understand ...

You mean "thumbs" right? Because the few times it's tried for me, that's what it puts there. Fingernail and all.

 

[Bartholomew Gallacher]

Don't ask me how I know, but AI also does not understand how a penis works, at all. For ANY function including "just being there."

Well most models won't even include one, because they are not trained on it. Some do, these are normally marked/described as horny. But that's no big problem, because that's why Loras are a thing.

LoRA stands for LOw Rank Adaptation, and these are very small changed to established models, which can contain a big variety of stuff, e.g. like water color painting style, manga styles, but also nurse uniforms, straightjackets or even fictional characters. There's a real big variety on it. Since Loras can be trained on 50+ images, and it doesn't take too long, also the download size is normally in the range of 10-50 MB, there are tons of it being available.

For example an Anime penis Lora, which you can download here: Anime penis LORA LoRA for AI Models . There are of course other Loras for that around as well, which might have a different image style.

NSFW penis picture ahead.

Spoiler

Stable Diffusion: What Are LoRA Models and How to Use Them?

Learn how to use LoRA models in Stable Diffusion to enhance your AI art and take your generations to the next level with character, concept and style LoRAs.

softwarekeep.com

 

[Noodles]

You mean "thumbs" right? Because the few times it's tried for me, that's what it puts there. Fingernail and all.

Yes. Or disembodied, or no balls, or upside down balls, or like, a head, with a head, with a head, which is really getting intonthe weirdness zone.

LoRA stands for LOw Rank Adaptation, and these are very small changed to established models, which can contain a big variety of stuff, e.g. like water color painting style, manga styles, but also nurse uniforms, straightjackets or even fictional characters. There's a real big variety on it. Since Loras can be trained on 50+ images, and it doesn't take too long, also the download size is normally in the range of 10-50 MB, there are tons of it being available.

I have tried different base models and used embeddings, maybe I need to try looking into LORAs a bit now.

 

[Bartholomew Gallacher]

I would not have too high hopes so on this specific topic for Loras, since this seems to be quite niche. Of course you could try and create your own genital Lora, there are myriads of tutorials outside on how to do that.

 

[Noodles]

I still only sort of get the difference in all these add on deals.

The base model I get, but then in Automatic1111 anyway, there are Embeddings and there are Hyperneteorks and then there are LORAs apparently, which I am not sure, maybe its the same as one of the others.

 

[Argent Stonecutter]

Yes. Or disembodied, or no balls, or upside down balls, or like, a head, with a head, with a head, which is really getting intonthe weirdness zone.

I can guarantee that someone's totally into that.

 

[Lexxi]

 

[Lexxi]

 

[w4ffl35]

Hi, I'm the creator of AI Runner - I can speak to some of the things you mentioned

I wanted to branch out from Automatic1111's tool, and have not tried Auto Diffusion yet, but I did decide to try this one out.

AI Runner with Stable Diffusion | AI Art Editor and Custom Chatbots by Capsize Games

AI Art using Stable Diffusion on your own PC

capsizegames.itch.io

I noticed it has that, controlnet option, so I decided to mess with it with some SL screen shots. With a weird bran fart moment. Unfortunately, it doesn't look like this tool auto saves results like Automatic1111 does, so I don't have the images.

If you want to automatically export images to disc, just got to `File > Settings > Image import / export` and check `Automatically export images`. There are also metadata options you can mess with there as well.

The first one I did, it came out, ok-ish, but the conversion was too old. So I just, ran it again, adding "Young" to the prompt. That ended up looking too young. I also decided to add "smiling", but it looked weird after a couple of runs, I took it out.

The other thing is, the results I am getting, are just becoming.... More and more kind of grosely ugly. Like, I wasn't expecting super model results, but I was getting more and more towards troll. I even started adding negative prompts like "crooked teeth" to no avail.

To get better looking people, I highly recommend using the Next Photo model paired with EasyNegative and veryBadImageNegative embeddings. Try generating at 512x768 and you'll get some very realistic looking images of people depending on the prompt you use.

You can set up custom model, embedding etc. paths by goijng to File > Settings > Import / Export preferences > Paths - choose any folders you wish to store custom models etc. in

Then I realized..... Each run, wasn't building on my original base image. It was building, on the last run. So my old slightly wrinkled woman, became a weird wrinkled child, then a weird wrinkled woman with an odd smile, which just kept becoming weirder and wierder and it kept making the teeth just, worse and worse.

Whoops.

Also, I feel like there must be something with NSFW content in this tool, it has a toggle button, its not clear which was is on or off, but I tried bothz but both times I tried converting an SL nude shot, it worked, ish, but the result was this weird pencil drawing. Like this was the creator's idea of a funny joke. It was based on the original image, because it had a similar pose.

You can generate NSFW with this tool as long as you turn the filter off (the button should not be highlighted if you want NSFW). I haven't done anything to mess with prompts so you'll get what you put into it - it can generate very realistic people including NSFW.

Controlnet however needs some work. You'll be happy to know that I'm working on v2.0 of the app and it should go live in a day or two. It features a huge list of updates including fixes to controlnet implementation, a revamped interface, loads of bug fixes and more. I'm extremely active in the development of this app. patches go out almost every day and larger updates go out weekly.

If you check the help menu in the app you'll find a link to Discord where I'm very active in answering questions like these, so feel free to join if you run into any other issues.

Thanks for your support and I hope you'll find the app useful, especially with the upcoming version.

 

[w4ffl35]

Another thing I want to mention: the upcoming version of the app allows for easy generation of things such as people just by clicking a single button. Here are some examples that I generated by clicking "person" - these were all generated at 768x1024 without upscaling using the model and embeddings mentioned in my reply above.

The tool I created in the upcoming version allows for mixing and matching certain elements (like ethnicity, gender etc) to generate more specific images, or just click "person". I'll have more info on it once this launches. again I'm about 2 or 3 days away from the 2.0 launch.

 

[Noodles]

Hi, I'm the creator of AI Runner - I can speak to some of the things you mentioned

Thanks for the info. It seems like a pretty nice easy to use tool BTW. I had wondered if there might be a setting to auto save images. When Inhave used other tools it was sometimes easier to just run prompts them go back later and delete out the really bad results.

 

[Essence Lumin]

Hi, I'm the creator of AI Runner - I can speak to some of the things you mentioned

I'm impressed! It isn't often a program creator makes an account on this little board to help people out with their questions.

 

[Argent Stonecutter]

How does it do for "a humanoid fox painting a picture of isembard brunel"?

 

[w4ffl35]

How does it do for "a humanoid fox painting a picture of isembard brunel"?

lol i don't know, but i can tell you that the thing I mentioned above can create animals as well and when i blend it with a text prompt like "nightmare fuel" or "devil monster" it creates some interesting things