ymgve posted:

When you give «simple» prompts like that, I think it might just «remember» some of the training data images more or less. Would be interesting to see the devs comparing the outputs to training images with matching descriptions and see if there are some near matches.

Ogdred Weary posted:

I made some "average somethingawful.com users" with MidJourney

Tunicate posted:

dalle is less inspired on that

Vlaphor posted:

I find it kind of neat when it brings up fairly accurate (or at least, recognizable) results.



Edit: Though, throw in a few more words and it looks accurate and weird. Beautiful snowy winter though

A Strange Aeon posted:

The yard sale reminds me of that one image that looks like reality but then you're hard-pressed to describe anything that's actually in the picture. Like maybe I can make out a blue garment but primarily it just has like the essence of a yard sale.

Vlaphor posted:

I think I might skip the waterslides here....



This is a case where the lower resolutions really come into play. Since you can't make anything out clearly, your brain fills in the details. The yards in those pics look exactly like yards, at this resolution and size.

quote:

We learn, for example, that park employees tested the unfinished Cannonball Loop – a water slide that ends in a full-blown upside-down loop – after Mulvihill offered anyone who dared to try it $100. We also learn that many of the later testers emerged from the slide bearing bloody gashes. The cause? Human teeth stuck to the slide's interior walls that were left behind by previous testers.

LifeSunDeath posted:

GABA ghoul posted:

Seize the means of art production from capital. Dall-e 2 should belong to everyone

pixaal posted:

Isn't it open source?

GABA ghoul posted:

OpenAI never released their dall-e 1 or 2 model to the public. I doubt they gonna do it until the tech is completely obsolete years down the line. Dall-e mini is an open source attempt to recreate the dall-e model, but they don't really have the same resources for training nor is it as good in general.

TIP posted:

it's definitely possible to replicate the training data

Wopzilla posted:

Cable Guy posted:

I gotta say I was somewhat dubious so ran the same string....



DallE expermenting with cubism and Pollock now...?

TIP posted:

post

played with Dall-E mini a while looking for insights into how it builds things up into a picture

earlier in the thread someone posted a tweet where some guy says AI generated art is the fleshlight of creativity, I copied his whole tweet into mini dall-e and got this:



every time I generated it I got a very similar set of photos with a couple Starry Starry Nights

then I saw someone else in the thread used just "AI generated art is the fleshlight of creativity" without the rest and got nearly identical results, including the Starry Starry Nights

this led me down a rabbit hole of trying to figure what part of that phrase was triggering it to recreate Starry Starry Night so regularly

so I tried each important word individually and determined that it seemed to be combining colors and shapes from AI, art, and creativity to arrive at Starry Starry Night







put them all together and



and interestingly enough the input "starry starry night" does not work



I also tried adding back in the less important words like of, the, and is and it was interesting the way that they could change the outcome, driving it towards or away from Starry Starry Night





"art is the creativity of AI" seemed to produce Starry Starry Night the most out of any entry using words from the quote, but I think I also figured out the absolute shortest input to reliably get Starry Starry Night



the fact that "AI art creativity" so reliably produces a specific painting makes me nervous at the thought of trying to use any of these systems commercially

we're gonna need a second AI to judge the outputs against the training materials to make sure we're not duplicating copyrighted designs

TIP posted:

I made a big effort post last month about it because I spent a while figuring out that phrase and I thought the way it gets generated out of the elements that each of the individual words produce was interesting

quote:

post

played with Dall-E mini a while looking for insights into how it builds things up into a picture

earlier in the thread someone posted a tweet where some guy says AI generated art is the fleshlight of creativity, I copied his whole tweet into mini dall-e and got this:

you can also just use "starry art" but that's less surprising

quote:

This dude boasting to us about how good his fleshlight is and he thinks it's ok because he doesn't think we're real people

Cable Guy posted:

I can't find the post but I used a goon's reply to that tweet...

... and the result was uh... illuminating

Rinkles posted:

I wonder how this shakes out legally.

Vlaphor posted:

I think I might skip the waterslides here....

quote:

The company announced in a blog post that it will expedite access for customers on the waitlist with the goal of reaching roughly 1 million people within the next few weeks. With this "beta" launch, DALL-E 2, which had been free to use, will move to a credit-based fee structure. First-time users will get a finite amount of credits that can be put toward generating or editing an image or creating a variation of an image. (Generations return four images, while edits and variations return three.) Credits will refill every month to the tune of 50 in the first month and 15 a month after that, or users can buy additional credits in increments of $15. Artists in need of financial assistance will be able to apply for subsidized access, OpenAI says.

mobby_6kl posted:

More DALL-E 2 invites incoming:

Where is my approval, assholes?

Understandable that they're going for the credit system to manage the resources, and great that it'll give you some every month. It's fun but I'm not paying money to generate pics of Ted Cruz wearing a bikini

Dia de Pikachutos posted:

For some reason "*in* a cosy home office" wasn't giving very good results?