Why don't indie devs use AI-generated images as art?

[what am i doing]

[snipped huge images]

The comic wasn't really funny. AI-generated shit on the other hand:

 

[Dexter]

Soon there probably won't be a need to use the tile by tile method, as that can still leave visible changes where one tile fades into each other. Every few days there's a new optimization that reduces VRAM usage, I can now generate 1664x832 px images on 6 GB VRAM. Emad from Stability AI did a Q&A on Reddit and announced that firstly the hardware demands will go down further pretty soon (they'll release the model in native 16 bit precision, among other things), and secondly they have a more hi-res version ready, that uses 1024x1024 training data natively, instead of 512x512.

I can already create up to 1600x1600 images on a 3080Ti with 12GB VRAM and the new 4090Ti are rumored to have up to 48GB VRAM, the problem (aside from it obviously lasting longer) is that all the Training was done at 512x512 with the current model and choosing anything much larger will absolutely fuck up your composition and/or makes things appear triple or quadruple, for instance the kind of results you can expect for a simple prompt like "John Berke Sci-Fi" with 512x512, 1024x1024 and 1536x1536:

Kind of feels like playing Infinity Engine Games on 4K displays all zoomed out

 

[V17]

Soon there probably won't be a need to use the tile by tile method, as that can still leave visible changes where one tile fades into each other. Every few days there's a new optimization that reduces VRAM usage, I can now generate 1664x832 px images on 6 GB VRAM. Emad from Stability AI did a Q&A on Reddit and announced that firstly the hardware demands will go down further pretty soon (they'll release the model in native 16 bit precision, among other things), and secondly they have a more hi-res version ready, that uses 1024x1024 training data natively, instead of 512x512.

I can already create up to 1600x1600 images on a 3080Ti with 12GB VRAM and the new 4090Ti are rumored to have up to 48GB VRAM, the problem (aside from it obviously lasting longer) is that all the Training was done at 512x512 with the current model and choosing anything much larger will absolutely fuck up your composition and/or makes things appear triple or quadruple, for instance the kind of results you can expect for a simple prompt like "John Berke Sci-Fi" with 512x512, 1024x1024 and 1536x1536:

Kind of feels like playing Infinity Engine Games on 4K displays all zoomed out

If you do it like I say in the paragraph below that, generate a 512x512 image, upsample and re-generate it in higher resolution using img2img, you circumvent this problem.

 

[mkultra]

Dexter that's pretty cool, never tried that - SDUpscale. Must look into!

 

[Üstad]

How to upscale the simplistic images we've drawn? And which tool we need to use MidJourney or Stable diffusion?

 

[Nathaniel3W]

How to upscale the simplistic images we've drawn? And which tool we need to use MidJourney or Stable diffusion?

For upscaling images, I use Real-ESRGAN. I've heard that there are better tools for cleaning up faces and keeping a photorealistic style. But I'm more interested in a hand-drawn look, and I've heard that Real-ESRGAN is better at preserving specific styles.

I haven't used MidJourney. I like Stable Diffusion a lot.

 

[Davaris]

Its amusing they built the database for their AI to train on by copying artists' work from the internet, but now they are trying to make it difficult to copy from them.

If you have Firefox:
Click on the padlock icon next to the web address in your browser.
In the menu that opens up, click on the padlock with 'Connection secure'
In the next menu click on 'More information'
Then click on 'Media' and you will see a list of files you can save.
The images are of type webp, so you will need to convert them if you want to post them here. I don't know what program does that.

If you have Chromium. I don't know. It might be a similar process.

I have seen a program that downloads everything from webpages and websites, but I can't remember what it is called.

Is there a way to search for images by text for MidJourney? Stable difusion has a site with this feature but I cant find one for MidJounrey.

I don't know. If you mean search the history of their showcase webpage, perhaps you can do that on the Wayback Machine website.

Could you post the link to the Stable Diffusion website that lets you search? I'm curious about that.

 

[Üstad]

Could you post the link to the Stable Diffusion website that lets you search? I'm curious about that.

https://lexica.art/

 

[Dexter]

How to upscale the simplistic images we've drawn? And which tool we need to use MidJourney or Stable diffusion?

Assuming you're talking about the AI Tools talked about in this thread (and not outside Tools like Gigapixel or Cupscale or whatever which might in part even give better results), someone wrote a Guide about SDUpscale since I posted if you have Stable Diffusion installed locally, maybe it explains it better: https://rentry.org/sdupscale
Also Inpainting/Outpainting if you want to improve or "Reroll" parts of your image or expand it outwards: https://rentry.org/drfar

Dall-e and MidJourney are Closed Source. The only way you as a person can use or interact with MidJourney is afaik joining their Discord and posting a command along your keywords and it'll generate an image for you that you can download. Stable Diffusion is Open Source and you can either use it by inputting keywords in any of the various Online sites offering it as a Service, or you can install various versions of it including various UIs to make it more useable on your own computer at home and use it yourself (preferably if you have an NVIDIA card with at least 4GB of VRAM), there's a guide on how to do that too: https://rentry.org/voldy

Most developments and knowledge gained is obviously going to happen on the Open Source version as people build upon it, there's Optimizations, bug fixes new features, Plugins and other improvements happening almost every day.

 

[Üstad]

How to upscale the simplistic images we've drawn? And which tool we need to use MidJourney or Stable diffusion?

For upscaling images, I use Real-ESRGAN. I've heard that there are better tools for cleaning up faces and keeping a photorealistic style. But I'm more interested in a hand-drawn look, and I've heard that Real-ESRGAN is better at preserving specific styles.

I haven't used MidJourney. I like Stable Diffusion a lot.

I was looking for something make a simplistic drawing turning into something detailed. I saw some examples of that on this forum but I can't find it and don't know which tool is used.

 

[Dexter]

I was looking for something make a simplistic drawing turning into something detailed. I saw some examples of that on this forum but I can't find it and don't know which tool is used.

That's the Img2Img Feature in Stable Diffusion, I think I had a few examples here: https://rpgcodex.net/forums/threads...i-generated-images-as-art.143986/post-8095279

Some more:

Spoiler

 

[Reinhardt]

can it generate top down tactical encounter maps for tabletop or kotc2?

 

[Üstad]

I was looking for something make a simplistic drawing turning into something detailed. I saw some examples of that on this forum but I can't find it and don't know which tool is used.

That's the Img2Img Feature in Stable Diffusion, I think I had a few examples here: https://rpgcodex.net/forums/threads...i-generated-images-as-art.143986/post-8095279

Some more:

Spoiler

Yes that's the thing what I was seeking. Thank you.

 

[Dexter]

can it generate top down tactical encounter maps for tabletop or kotc2?

It can certainly create facsimiles of maps and town maps and also knows what "dungeon maps" or "tactical encounter maps" are, if they'll make particular sense though is questionable:
https://lexica.art/?q=town+map
https://lexica.art/?q=dungeon+map
https://lexica.art/?q=tactical+encounter+map

Could probably spruce a self-drawn one up a lot though.

 

[Reinhardt]

interesting

 

[fizzelopeguss]

Digital art has always been lowest common denominator.

If you can draw, and want to continue making money drawing then just become a tattooist.

 

[Reinhardt]

there is one thing it do perfectly - monster portraits. shit is scary yo. proper body horror.

 

[Darkozric]

there is one thing it do perfectly - monster portraits. shit is scary yo. proper body horror.

Cats are not bad too.

 

[Non-Edgy Gamer]

This Stable Diffusion Krita plugin is better than photoshop:
https://github.com/sddebz/stable-diffusion-krita-plugin

I saw an image someone else made that I liked, selected the teeth, fixed them and then smeared some red on her face to make ketchup.

Just scribble something, feather select and tell the AI what it should be doing.

 

[Derringer]

ai is basically tracing but nobody gives a shit since all people do is trace and change enough shit so it doesn't look the same, that goes for music as well

 

[Non-Edgy Gamer]

ai is basically tracing but nobody gives a shit since all people do is trace and change enough shit so it doesn't look the same, that goes for music as well

A bit of a simplification. More accurate to say it's copying. You are asking the AI to copy something similar to hundreds, thousands or tens of thousands of things it's seen.

It's like asking an experienced artist to draw for you, only this artist is more likely not to understand you and draw a human centipede.

 

[Naveen]

After spending a few hours at lexica, I think people are misusing this technology by trying to get perfect-looking portraits of waifus. That's boring and completely uninteresting (and you get awful-looking hands anyway.) The cool shit is asking it to make weird stuff, like a roman mosaic of a bat

Blueprint of a laser gun

or an undiscovered one by Da Vinci:

 

[Davaris]

So close... I can't convince it to move the camera up into the sky, and look down on the figure like it was an isometric game.

The problem is when they advertise lead figures in online stores, they take photos of them like the above. They do not take photos looking down on them like they are in a game map.

 

[Bad Sector]

Stable diffusion, etc all that stuff sound interesting, though there are two issues i've found so far:

• They pretty much all seem to need Nvidia GPUs. Supposedly there are workarounds for AMD GPUs for some but seems to be ignored for the most part.
• The software itself might be open source but the data "weights" has some weird ass licenses that require logins, etc. IMO for this to truly be open, not only the "weights" but the training data also need to be open so that the models and algorithms can be improved on

What i'd like to see is some sort of program that i can download on my PC, with all the data locally that i can easily back up and be able to use as i like, preferably without weird licenses.

 

[Davaris]

Stable diffusion, etc all that stuff sound interesting, though there are two issues i've found so far:

• They pretty much all seem to need Nvidia GPUs. Supposedly there are workarounds for AMD GPUs for some but seems to be ignored for the most part.
• The software itself might be open source but the data "weights" has some weird ass licenses that require logins, etc. IMO for this to truly be open, not only the "weights" but the training data also need to be open so that the models and algorithms can be improved on

What i'd like to see is some sort of program that i can download on my PC, with all the data locally that i can easily back up and be able to use as i like, preferably without weird licenses.

Unfortunately the data is the most valuable part. It cost them 100s of thousands of dollars to train it. Eventually the community will make their own data.

 

[rusty_shackleford]

with respect to some licensing issues: in USA it actually became settled rather quickly, AI generated content is not bound to the licenses of the data used to train it