Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

[Near]

Welcome to the Stable Diffusion thread. Feel free to discuss and post images generated using Stable Diffusion here. Prompt discussion and feedback is also very much welcome.

I am featuring this thread because discussion elsewhere on the Internet is dominated by pornography and very unsavory types. Please enjoy the death of art and the beginning of the AI era responsibly.

All NSFW should be spoilered and clearly labeled. "Ecchi" should also be marked as such.

PSA: Monitor your GPU temps and increase cooling and/or undervolt them if you need to. There have been claims of frying/breaking/worsening GPUs due to high temps. There has been a claim of display kernel errors, but that might be user error. It's reccomended you add "--medvram" to your command arguments for stable diffusion if you haven't already

What is this?
Stable Diffusion is an open source AI Art model. Some may have noticed that it took the Internet by storm. Its creations have been by far the most accurate and photorealistic yet, and blow other non-open source models such as DALL-E 2 out of the water. Stable diffusion is capable of both img2text and img2img art generation. The user's way of generating particular art is done by inserting positive prompts (things they want the art to contain) and negative prompts (things that should be omitted).

As an example of the kind of art it can generate using different kinds of models, here is a few examples from /g/




How do I get it?
/g/ has WIP guides for how to locally install Stable Diffusion. There's multiple guides depending on what kind of hardware you own. Long story short, if you are on an Nvidia GPU and on Windows, you're going to have the easiest time installing it. The recommended GPU to generate art is minimum a GTX 970 and its Radeon equivalent. Most recommended is anything from Pascal series and upwards.

Local Install:
Nvidia GPU: https://rentry.org/voldy | https://github.com/AbdBarho/stable-diffusion-webui-docker
AMD GPU: https://rentry.org/sd-nativeisekaitoo | https://rentry.org/sdamd | https://rentry.org/ayymd-stable-diffustion-v1_4-guide
CPU: https://rentry.org/cputard

Guides
NovelAi: https://rentry.org/sdg_FAQ
Rob's NovelAI tips: https://rentry.org/robs-novel-ai-tips
NAI Diffusion: Various Tips & Tricks: https://github.com/TravelingRobot/NAI_Community_Research/wiki/NAI-Diffusion:-Various-Tips-&-Tricks
Dreambooth: https://github.com/ShivamShrirao/diffusers/tree/main/examples/dreambooth
Inpainting/Outpainting: https://rentry.org/drfar
Upscaling images: https://rentry.org/sdupscale
Textual inversion: https://rentry.org/textard
Training: https://rentry.org/informal-training-guide
Hypernetworks: https://rentry.org/hypernetwork4dumdums

In-Depth Usage Info
/g/ wiki: https://wiki.installgentoo.com/wiki/Stable_Diffusion

Tools & Resources
Index: https://pharmapsychotic.com/tools.html
Artist Styles: https://pastebin.com/HD7D6pnh (embed) (embed)
Prompt sharing: https://lexica.art/ | https://krea.ai/
Wildcard list: https://github.com/Lopyter/stable-soup-prompts
Dataset: https://pastebin.com/HpBPTmbF (embed) (embed)
img2img skeleton: https://app.posemy.art/

The most popular model right now is the NovelAI model. It has been leaked and is available to download via magnet link.

I just want to generate anime tiddy, this is hard!
If you're having issues with the installation or lack the hardware to generate art efficiently, feel free to use the NovelAI model directly from their site. Keep in mind however that it is paid. It's designed to be a coherent experience with as little hoops to jump through as possible. The cost of the cheapest subsciption is $10USD. and then it uses a separate currency token that is spent on generating images.

NovelAI - The AI Storyteller

GPT-powered AI Storyteller. Driven by AI, construct unique stories, thrilling tales, seductive romances, or just fool around. Anything goes!

novelai.net

Ok but does it generate porn?
Type nsfw in positive prompts and you will quickly find out. And no don't post any here.

How fast can my card generate this stuff?
VRAM is the key player in Stable Diffusion. The more you have it, the better. You want to have as many it/s as possible.
You can refer to this chart for a rough idea of GPU performance while generating art.


Happy painting!

 

[Off Brand Soup]

Some of the results for inpainting are truly breathtaking.

 

[Near]

Simple Wikihow dog prompt
img2img

 

[Off Brand Soup]

For anyone following the voldy guide who doesn't want anime shit, for the love of god download the hugging face model, skip novelAI. Everything else is specifically for various types of 2D porn.

 

[Slav Power]

Let me know when a sensible GUI that's piss easy to set up shows up, something like Cupscale for ESRGAN, because I don't feel like running a bunch of commands and installing specific versions of Python and dealing with all this shit just to have some AI fun. Also don't tell me to use Docker, it's retarded and doesn't think I have virtualization instructions on (I do have them on) and won't even start because of it.

 

[Thought precriminal]

I've messed around with Dreambooth (training the model on a specific thing, i.e. you own face) over the past two weeks. You take a bunch of photos of your face/body and then upload them to the Dreambooth notebook which will then train the Stable Diffusion model on your face, associated with a keyword, so that you can then use that keyword in prompts to generate images with your face in them.
Conclusion: it is flawed but can still produce some impressive results.
There are a couple of different ways to go about this. One requires a lot of computing power, and apparently the only way to use it is to rent computing from a service like vast.ai or runpod (actually very cheap) unless of course you own a PC with the required specs. The other can be done for free on Google Colab. Apparently the former will produce better results than the latter, but even on the Discord server people don't seem to really know what works best yet; it's still early days.
Follow these tutorials if you want to learn how:

 

[Off Brand Soup]

Trying to make our favourite lad Ralph.

As of right now my prompts are
Positive:

fat, man, obese, (sunglasses), ((wearing a cap)), full body, yelling, angry, solo, sitting down, shirtless, breasts, hairy belly, drinking

Negative:

woman, beer glass, beard

() gives more weight to words incase you didn't know.

 

[Near]

KICK THAT HOT DOG BITCH OFF THE PLANE




NIGGER

 

[Shiawase]

IMG2IMG fun with Windows 3.11 bootscreen:



This abstract set goes out to @Bango Bongo in the Congo:

Spoiler: Win3.11 in the congo

Let me know when a sensible GUI that's piss easy to set up shows up, something like Cupscale for ESRGAN, because I don't feel like running a bunch of commands and installing specific versions of Python and dealing with all this shit just to have some AI fun. Also don't tell me to use Docker, it's retarded and doesn't think I have virtualization instructions on (I do have them on) and won't even start because of it.

Spoiler: This youtuber Soiboi was very helpful to get me going

This is best I have found so far. Shows how to set up the image editor Krita with an easy SD interface. Even for me it was not too complicated to set up or use. (Yes it includes ESRGAN and GFPGAN functionality without an additional setup process.)

4GB+ VRAM:

<4GB VRAM:

 

[Near]

img2img
man, fat, obese, fully body, short hair, yelling, (((angry))), solo, black sweater, yellow shirt, ((((hairy stomach))))



Suuuuuweeeeeeee!

 

[MarvinTheParanoidAndroid]

Ok but does it generate porn?
Type nsfw in positive prompts and you will quickly find out. And no don't post any here.

I can only imagine what butchered stinkditch roast beef this AI thinks is a vagina.

 

[Near]

I can only imagine what butchered stinkditch roast beef this AI thinks is a vagina.

Given it goes through all the boorus and other shit, it's actually not too bad most of the time.

 

[Slav Power]

Okay, managed to get it set up. I have no idea if it's working 100% correctly but it works.

As an example, an img2img result of my avatar unedited with the prompt "A photo of a border collie dog in the center of the photo looking at the distance, white background"

 

[Near]

Okay, managed to get it set up. I have no idea if it's working 100% correctly but it works.

As an example, an img2img result of my avatar unedited with the prompt "A photo of a border collie dog in the center of the photo looking at the distance, white background"

View attachment 3742513

It works, but consider using tags like a booru site.

 

[Slav Power]

It works, but consider using tags like a booru site.

Care to explain? I'm just supposed to pull booru tags as prompts to generate images? Also which models should I download? Right now I only have the standard Stable Diffusion v1.4 checkpoint, and I don't know what else to get from the NovelAI leak? I don't really have disk space for all of it.

 

[Monolith]

For anyone following the voldy guide who doesn't want anime shit, for the love of god download the hugging face model, skip novelAI. Everything else is specifically for various types of 2D porn.

You're not kidding, waifu diffusion gives me this for "among us."

 

[Stalphos Johnson]

Care to explain? I'm just supposed to pull booru tags as prompts to generate images? Also which models should I download? Right now I only have the standard Stable Diffusion v1.4 checkpoint, and I don't know what else to get from the NovelAI leak? I don't really have disk space for all of it.

You only really need 3 files from the leak, which should go to about 5gb, animefull-final-pruned or animesfw-final-pruned, animevae.pt, and the hypernetwork modules. You can find a lot of tips for using SD here, which I've found somewhat helpful. That site has a lot of additional models, but they're all porn so I haven't touched them. Not a lot of sfw models floating around unfortunately.

 

[Near]

Care to explain? I'm just supposed to pull booru tags as prompts to generate images? Also which models should I download? Right now I only have the standard Stable Diffusion v1.4 checkpoint, and I don't know what else to get from the NovelAI leak? I don't really have disk space for all of it.

I don't know what the default 1.4 SD checkpoint likes, but I know the Novel AI models like their booru tags. Since the model was trained on booru sites and such, it appears to be the most familiar with tags.

 

[Null]

I find it fun to try and answer a topic with a prompt. It's very difficult to try and persuade it to have two people interacting. It could just be because I'm on AMD and running AMD for this is more time consuming than grinding up every african child in Patrick's basement.

speaking of, the man of the hour.



this was very hard to get anything looking right with the meat or utensiles and in that picture the hand is very fucked up. I took it to img2img and tried to repair the hand.



Easier said than done tbh

 

[Near]

I just ordered a 3070ti because of this fucking autism. I was gonna wait, but fuck it, I'm having too much fun. 7it/s is gonna be pretty baller.