Discoart / DALLE - AI generated art

[Eric]

In our ceaseless quest to make robots that can replace us entirely and force us to reconsider how we are going to exist as a society with careers and capitalism and what not, we are finally on the cusp of replacing those pesky artists! /s

You have probably seen the DALLE and discoart Images on Twitter. It is capable of replicating various artist styles and motifs with textual and image prompts:





Now if you ask for nice things it often gives you nice things back. If you ask for things you shouldn't it will give you nightmares as punishment. I'm just saying be careful what you wish for.

For this atrocity I provided a black and white silhouette and asked for some starwars.



Here I provide an image of a grassy field and asked for "A beautiful painting of a stargate in a field of grass in the style of Thomas Kinkade":

This world has a lot of gates though most seem to be half buried. To keep out the Ghoo-ha-oould no doubt.

Here I try again without a image prompt:
create(text_prompts='A beautiful painting of a stargate in a grasy field by Thomas Kinkade, Trending on artstation.'




I don't think this one was done baking but I also don't think we need any more Thomas kinkade pieces, generated or original, so lets move on.

'A beautiful painting of a stargate by Beeple, Trending on artstation.' - I really like this one.


I'm not the one to ask for help when it comes to what the heck is going on here. I am more like a monkey with a gun rather than the intended audience for using this tool, but maybe they shouldn't have left guns in the monkey enclosure? If you have a prompt you want me to run though post it! Or give it a try for yourself! You might need to click on the resources button on the topish right of the page and enable GPU hardware acceleration as this is obviously a GPU intensive project. Images take about 25 minutes to bake but it shows you progress images as it works which is quite entertaining!

Google Colab

colab.research.google.com

 

[Andy]

At the other end of the spectrum, I've found what might be the dumbest use of this technology: generating an image of a thing that was described (but not captured) in a TikTok video, and then using it as the lead image in a news article summarizing the TikTok.

(I'm not going to link to the article, because I don't think they deserve to be rewarded for this.)

That beeple image is pretty great. I didn't know these took that long to create--I thought it was more or less instant.

 

[Eric]

At the other end of the spectrum, I've found what might be the dumbest use of this technology: generating an image of a thing that was described (but not captured) in a TikTok video, and then using it as the lead image in a news article summarizing the TikTok.
View attachment 142
(I'm not going to link to the article, because I don't think they deserve to be rewarded for this.)

That beeple image is pretty great. I didn't know these took that long to create--I thought it was more or less instant.

That’s funny, that was almost exactly what i was predicting it was going to be used for last night. I imagined it being used a little more proficiently though. Like those modern tech and social news sites that do more long form thought peices and usually have some sort of generic simple illustrations at the top that are so vague that its hard to tie it to the article itself. You could fill a library for editors to choose from in no time with this tech. Especially if you ask for it in an artists style like Matisse and avoid the blurry half baked stuff.

I think the slow speed may be because this one is tuned to exist in the free tier of google collab and the render size is a bit larger than some of the other faster free sites I have since seen. The project I linked is a fork of DiscoDiffusion which I just read could be run locally! I am going to see if i can get that going on my gaming rig since it sounds like someone got it running in windows.

GitHub - MohamadZeina/Disco_Diffusion_Local: Getting the latest versions of Disco Diffusion to work locally, instead of colab. Including how I run this on Windows, despite some Linux only dependencies ;)

Getting the latest versions of Disco Diffusion to work locally, instead of colab. Including how I run this on Windows, despite some Linux only dependencies ;) - MohamadZeina/Disco_Diffusion_Local

github.com

 

[Eric]

I played with this much faster and lower quality app which leverages DALLE mini for its generation. i had read discodiffusion is more powerful but both are similar in broad strokes it seems.

Craiyon, AI Image Generator

Craiyon is an AI model that can draw images from any text prompt!

www.craiyon.com

New avatar?

“Bigfoot boarding the Myst rocket at night”


I used to say that if I could go back in time I wouldn’t change anything.

 

[Andy]

@Eric if you're looking for inspiration, I sent a friend this thread and he came back with this suggestion:

 

[Eric]

I absolutely love the prompt and am accepting requests!

So starting with the basics, I fed the goosebumps art as the image prompt and used the text prompt of

create(text_prompts='A painting of The Beastie Boys by Tim Jacob's, Trending on ArtStation.',




Kinda of a mild fear and loathing vibe in this one. And everyone remembers the 4th member of The Beastie Boys, Ad-tinyhead.

I really want their wish to be true too though so I'm trying to tweak the prompts to get there. Maybe there isn't much Tim Jacobus tagged art in their data snapshot? He seems to stick to air brush for most of the goosebumps so changed it up:

create(text_prompts='An airbrushed painting of The Beastie Boys in a phone booth by R.L. Stine.',

...Is that Chandler?


I don't think I can offer two images for merging the way we want here, but I still think we might be able to make this happen so I'll continue running variations on this prompt.

 

[Andy]

OMG that's amazing

 

[Eric]

Trying to learn some of the adjustment that can be made I wondered a bit from the prompt with another goosebumps cover:

create(text_prompts='An airbrushed painting of The Beastie Boys hatching out of eggs by Goosebumps.',
       init_image='42_egg_monsters_from_mars.jpg',
       steps=500,
       skip_steps=50)

DiscoDiffusion (b. 2021)
Brass Monkey, 2022
Satanic Computer Rituals on Canvas

I like the little egg man and this time I'm not talking about Columbo.

 

[Andy]

How random is it? Like if you ran that same query again, how close to the same result would you get? What about running it on another computer? Does it seem like that would get the same results, or does it seem to use data from its installed environment?

 

[Eric]

I’m of the understanding that its not even possible to get the same result again. The skip step parameter will help you control how close to the image prompt you want it (apparently by defualt you skip the first 100 steps as it changes so much there as to be unrecognizable from the image prompt in those first steps, egg man is with less skipped steps but if increased we would look more like the image prompt). There is another parameter that can control how aggressively it tries to reach the text prompt over the image prompt. There are also weights which can be assigned to the words in the text prompt. And then theres like dozens of other parameters all of which are playing in the same data set for the machine learning which had like 4 million images in it. You could theoretically curate your own data set for it as well but would be like diving headlong into a data science degree.

I think part of how it can make images from nothing is that it basically makes generates a noise layer on top of it and then uses image processing algos to sharpen the noise into shapes multiple times with the same noisy image and then AI decides which of those it thinks is closer to the text prompt sort of like natural selection. My understanding is that because its using noise and making like 250-1000 decisions per image the chaos of the noise would make it nearly a statistical impossibility to get the same image.

The way the google collab sheet is setup it just keeps running so it throws each completed result into a data set that you can pull the results from. I rarely let it make more than one image per prompt without changing the prompt though because they take so long. Like the Craiyon results in my previous post it should eventually give a whole group of similar results. If you adjusted the parameters tight enough we might be able to back it into corner of the results we want but each image will be unique.

 

[Eric]

Here is a gif of someone's collected batch. The results are so similar they must have worked from the image prompt of the girl while skipping the first 30% of the steps to get a fairly homogenous result.


This is what it looks like as it refines the image. This is after that skip-step parameter so it starts past a lot of the early steps which cause radical changes so at this point it just refines what is there though it can add new details.

 

[Andy]

Ok that makes sense. Those gifs are cool as hell

 

[Eric]

create(text_prompts='A beauitufl painting of colorful Corals in the ocean with sunrays beaming through the water by Lisa Frank, Trending on artstation',

 

[Eric]

Sadie came up with the artist pairing of the century after seeing the coral picture.

create(text_prompts='A beauitufl painting of colorful aliens by Lisa Frank and HR Geiger, Trending on artstation',
       steps=300,
       skip_steps=50)

 

[Eric]

create(text_prompts='A xenomorph by HR Giger and Lisa Frank,  trending on artstation',
       steps=300,
       skip_steps=40)

 

[Andy]

LISA FRANK AND HR GIGER IS THE PAIRING I NEVER KNEW I NEEDED!

I especially like that it knew to add both of their signatures.

 

[Andy]

Also, I am getting a tidal wave of DALL-E stories from Google news now, and just saw this one.

I feel like BoingBoing buried the lede here (shockingly). The opening images are hilarious, but also scroll down and see the photographer recreations. I thought we had a year or so until we had serious copyright implications, but they're here now.

"Using DALL-E to make images of Muppets in Mad Max | Boing Boing" https://boingboing.net/2022/07/05/using-dall-e-to-make-images-of-muppets-in-mad-max.html

 

[Eric]

It seems art, whether it be by paint, photograph or 3d render, we are going to have some interesting ownership claims to grapple with. I always have found that concept interesting and perhaps incompatible with captialism though. Like how a singular music artist can invent an entire genre once it has enough people copying their style. Where we draw the line on “inspiration” versus outright theft or unoriginality is intriguing to me.
I can’t help but think of Flatmo’s influential style and influence on Humboldt County and then the issues that occurred back in 2015 when Lost Coast Brewery paid another artist to replicate the style of the logo. I suspect soon the “other artist” will be a DALLE descendant.

When the Shark Bites

Humboldt's best known artist and biggest beer company clash over the area's most recognizable illustration

m.northcoastjournal.com

 

[Andy]

Oh my god, someone's feeding news headlines through it.

For Sale Domain: prompt.press

👉 Prompt.press #DomainsForSale #MakeOffer #BuyNow #PromptPress 💡 Prompt.press - A premier domain for AI-driven prompting services! ✨Ideal for: - AI Prompting Platforms - Content Creation Tools - Creative Writing Assistants - Copywriting Servi

prompt.press

 

[Eric]

Now that's what I'm talking about! These are still not a good idea for actual news articles in my opinion but since they are sticking art at the top of the ones they don't have photos for anyways it's probably inevitable.

I think I want to pair this with GTP3 AI text generation to make up the headlines too. Then I can live an entirely fictional world "based on a true story".

This Elon musk buying twitter one is fantastic. I love that the twitter bird has turned into some sort of fat blue turkey/dodo thing.
Elon Musk to Acquire Twitter