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

[Outland Cow]

I come back every few weeks/months, am in awe what is possible and then proceed to have absolutely no usecase whatsoever for it. I wish I had this stuff twenty years ago, when I needed graphics assets for vidya I was writing. It would have been a major motivator to actually continue.

You could always just use it for fun.

 

[Baraadmirer]

You could always just use it for fun.

This. I generally make what I create free to use because I find it ridiculous to monetise images I can make in 30 seconds. Still don't understand why people make Patreons over these.

 

[Rembrandt]

This. I generally make what I create free to use because I find it ridiculous to monetise images I can make in 30 seconds. Still don't understand why people make Patreons over these.

Where there's demand there will always be supply. When it comes to image gen especially there's also a budget/skill gap that not everyone can overcome. If you don't have the VRAM for it, you're not going to do any image gen on your own, which means depending on other people to train LORAs. And some people are just too lazy/dumb, so they'll not even resort to prompting their own artwork in third party services.

 

[macrodegenerate]

It's been a month or two since I've really taken a day to checkout civit and play with all the new SD stuff, and I'm always amazed to see how much the meta changes.

Illustrious fine-tunes have completely taken over the meta. I'm starting to notice that all of the Pony LORA bakers are just rebaking everything in Illustrious. Pretty cool.

Someone finally made a LORA of an artist I was trying to make for a solid half a year. Works amazingly well in Illustrious.

I would really recommend trying out Dynamic Prompting for Forge/Comfy. I gave it my first real try today and spent an hour writing a Jinja2 template using wildcards, and now I can generate characters with a vast variety of different characteristics, genders, and apply varying Loras appending their trigger words easily.

I can generate a cowboy, wearing a fezz, with random pants, and a variety of random different effects based on LORAs. it's awesome.

I wonder what ASCAS is up to right now. He really started bringing attention to Illustrious on the farms before it was big. Probably having a blast, even though most of the LORAs he likes haven't been converted to Illustrious. I wonder if he ever finished his.

 

[Baraadmirer]

Yeah, I'm seeing a lot of updated Illustrious versions for LoRAs I'm following. I'm assuming everyone's training for SDXL with Illustrious as the base model now? No idea if any training parameters aside from that has changed.

 

[macrodegenerate]

Yeah, I'm seeing a lot of updated Illustrious versions for LoRAs I'm following. I'm assuming everyone's training for SDXL with Illustrious as the base model now? No idea if any training parameters aside from that has changed.

The weird thing is I'm finding is some of the concept LORAs I like are becoming obsolete. The concepts seem to already be baked into Illustrious. Or maybe they were just washed out of Pony during training.

Also Pony v7 is half way done training according to the guy baking it. Really weird choice training it on Aura flow , but we'll see.

 

[Eventide]

Is Midjourney any good? I haven't really been following AI Art tools, but I was thinking of purchasing a subscription for funsies.

Weirdly enough, I'm actually an artist But I doubt AI is going to be replacing sculpting any time soon.

 

[The Spoils of War]

Is Midjourney any good?

It's pretty good but expensive. All of your images will be public unless you shell out more money.

 

[Eventide]

It's pretty good but expensive. All of your images will be public unless you shell out more money.

Is there an AI service/app you'd recommend that isn't public? & preferably user-friendly.

 

[The Spoils of War]

Is there an AI service/app you'd recommend that isn't public? & preferably user-friendly.

NovelAI is great for image generation if you don't mind an anime art style. Everything is private, and you can just pay for credits instead of a subscription, but you'll have to learn how to prompt with Danbooru tags. I think that's covered in their docs.

 

[Baraadmirer]

Is there an AI service/app you'd recommend that isn't public? & preferably user-friendly.

CivitAI allows you to generate images at the cost of their site currency (which you can pay for with real money, but otherwise can gain slowly with posting images, reacting to other people's images, etc.) From what I can tell you can choose to publish your images on the site, but it's not mandatory.

 

[Slurred]

Is there an AI service/app you'd recommend that isn't public? & preferably user-friendly.

Bing Image Creator is free and uses OpenAI's DALL-E 3. You can also use DALL-E 3 from within ChatGPT (free or paid) by asking it to generate images.

 

[macrodegenerate]

If you are willing to learn about how the stuff works, you could always head over to runpod and spin up a pod with webui-forge. It's less than a dollar an hour for a decent GPU. Plus you have a lot more freedom, and if you're paranoid you can download your stuff over SCP/SFTP which encrypts your images over the net.

 

[Stoneheart]

is there a good was to make connected series of pictures? i want multiple pictures in the same style using the same models.

 

[Rembrandt]

is there a good was to make connected series of pictures? i want multiple pictures in the same style using the same models.

What have you tried so far? What software are you using and what model?

 

[Stoneheart]

What software are you using and what model?

all the normie stuff. and i want the model to be created with ai and used over multiple outputs.
im open to pretty much all suggestions of programs as long as they have some for of cheap/free test phase or allow you to buy small amounts of token.

I just want to make my Drachenlord/lightbringer crossover comic....

 

[Rembrandt]

all the normie stuff. and i want the model to be created with ai and used over multiple outputs.

You can make a LoRA and use it as many times as you want, even for photorealistic characters. Drachenlord and Lightbringer should be entirely doable.



Making image gens that are consistent in their look and style isn't that difficult nowadays, attempts were made to streamline the process and I found some good guides online. If you'd like to make models (LoRAs?) of Drachenlord and Lightbringer, this could be done in two ways - training your own LoRAs or making prompts that roughly resemble them and reusing those instead.

For local training this guide covers a lot of information, starting from making a dataset (in your case it'd be finding 20+ images of Drachenlord online and cropping them to the right resolution), how to curate it (telling a tool to do it for you) and then training it all locally (NAI 1.5, PDXL or Illustrious, software is linked w/ presets, zero thinking required). If you don't have a computer for it then you can give any of the online services a try, though be wary they'll most likely hold your credit card info hostage and change up the payment plan without you noticing. Civitai will charge you a bare minimum of $5 for their service (500 Buzz to generate an image then a varying price for on-site generation), but it seems to be the safest choice.



As for a choice of checkpoints, depending on your preferred style you could either do realistic gens on Flux, Stable Diffusion, Pony (not recommended) or a cartoony/weeb style in Illustrious. Flux is the safest choice if you want real people to do things together on one image. Natural language works best with Flux and Stable Diffusion, you can spell out what you want and it'll roughly understand it, while Illustrious was trained on an anime porn website and works best with Danbooru tags.

I think that covers everything I know. If you want to have an easiest time, then just pay Civitai and use Flux.

 

[jONALD]

I dislike Pony Diffusion, it uses that dreaded hive of autism that is e621 tag system and the "score_9" quality tags. Its great for all those niche details Danbooru never bothered to fucking tag but for some reason autistic gooners just can't be fucked to deal with hair. At least its good for nonstandard anatomy like aliens or demons.

 

[Nick Obre]

My lads, sorry if this isn't the place to ask, but I've recently seen things like this

These are shitposts, obviously, but aside from the potential for even more shitposting, I think that kind of thing could be very useful to create very brief bits of footage for fanedits (which is a hobby of mine) and things like that.

I don't ask for like a whole ass tutorial, but does anyone know what programs and models and so on I should be looking at if I wanted to do something like that?

The first idea I have is producing a usable shot of Smaug breaking out of Erebor, just without the gold, from footage of the second Hobbit movie. I don't know if that's even feasible, but any guidance is appreciated.

 

[⋖ cørdion ⋗]

I come back every few weeks/months, am in awe what is possible and then proceed to have absolutely no usecase whatsoever for it. I wish I had this stuff twenty years ago, when I needed graphics assets for vidya I was writing. It would have been a major motivator to actually continue.

Isn't that art in general? I spent most of my years around furries and other such communities that "used" art, but ignore that aspect and really, what is art for if you can't reliably monetize it? You got basically NEET asian american girls racking in thousands of dollars a month doing art on stream, rarely for commission sake, and they stream games on the sideline. I haven't seen anyone monetize simply making art outside furry and hentai platforms.

"You can create ANYTHING!", oh okay uh. Leon from RE4 in a mario costume! Haha.. Uh, Leon fucking that mouse girl.. heh, hot. Okay, what now? Back to my games. There truly is no use for AI art outside porn and OCs. My local barber did some funny xmas picture of a barbershop in the woods, cause he quite literally lives in the woods and has his shop there. Obviously nobody seethed, they went "oh that's AI teehee funny" and moved on with their lives.

'Art' never had much use in modern society if not actual physical art. I was gifted some etsy drawing of my cat which is clearly a template that was colored in but done so incredibly well. A novel, cheap idea. I can imagine someone using AI in 3 years to give grandma a funny picture of her own cat with its unique markings doing something unrealistic. That's what AI is for; novel shit. The same way you can ask chatgpt to write a small bio for your webshop, you can ask AI to put this keyed-out couch into a home setting for a product image.

There is no real other use for AI than these one-off gimmicks, exactly like 3D printing. "Dooood I can make anything!", okay, what'd you make? "Well I replaced this tiny hinge I broke". Cool?