Have you guys seen OpenAI's new "Dall-E 2" art AI?
Printed From: Pixel Joint
Category: The Lounge
Forum Name: Diversions
Forum Discription: Get to know your fellow pixel freaks. Chat about anything to do with video games, comic books, anime, movies, television, books, music, sports or any other off topic bs you can think of.
URL: https://pixeljoint.com/forum/forum_posts.asp?TID=27363
Printed Date: 09 September 2025 at 11:57pm
Topic: Have you guys seen OpenAI's new "Dall-E 2" art AI?
Posted By: specialmin64
Subject: Have you guys seen OpenAI's new "Dall-E 2" art AI?
Date Posted: 03 May 2022 at 2:12pm
|
You can find lots of examples of Dall-E 2's creations on the "dalle2" subreddit. It's insanity; users input a simple text prompt (whatever they can think up), and in a matter of seconds, the AI creates like 10 versions of what they typed but in image form. I've seen AIs that do this before, but Dall-E 2 is in another league. It occasionally surpasses what you'd expect from real artists! Example: "Painting of the orange cat Otto von Garfield, Count of Bismarck-Schönhausen, Duke of Lauenburg, Minister-President of Prussia. Depicted wearing a Prussian Pickelhaube and eating his favorite meal - lasagna."  Imagine what technology like this will look like ten years from now, and the impact it could have on art. Is it a bad thing that will put artists out of a job, or a good thing that will just make art better and more accessible? Shout out to James Gurney's blog where I first read about this!
|
Replies:
Posted By: NancyGold
Date Posted: 16 June 2022 at 2:01am
|
1. The cat has eye problem. 2. Dall-E is not public, but there are opensource versions now. 3. It is more than just art. These things have limited self conscious and can operate objects in real world. Whatever they draw they can also process in other dimensions. |
Posted By: specialmin64
Date Posted: 16 June 2022 at 4:30pm
|
now that you mention it, that cat's face is jagged up... I could draw a wayyyyy better and more epic cat  .And yeah, I figure this is about more than just art. It's very disturbing to me how advanced AI appears to be getting, but no one else seems to care at all, as if the technology that produces programs like Dall-E 2 is a harmless novelty. I heard a spooky conversation between Sam Harris (podcast host) and Eric Schmidt (former Google CEO and AI guy) about the potential of AI that actually made my hair stand on end
|
Posted By: NancyGold
Date Posted: 17 June 2022 at 3:59am
|
Originally posted by specialmin64 now that you mention it, that cat's face is jagged up... I could draw a wayyyyy better and more epic cat .The cat is not "epic", because there were no constraints for it to be epic. Same way client commissions you artwork, and then complaints it did not met the expectation, because client thought of a completely different cat. And fixing the face is just a matter of future research I guess. I.e. AI should assume symmetry by default. That is common knowledge for us, but AI systems, despite all the lovely symmetric algebra behind, tend to mess the symmetry up for some reason.
|
Posted By: NancyGold
Date Posted: 17 June 2022 at 4:06am
|
Originally posted by specialmin64 now that you mention it, that cat's face is jagged up... I could draw a wayyyyy better and more epic cat .BTW, even the open source version knows what "epic" means. Here is just plain boring burning tree, versus the epic burning tree. So guess if you ask AI for epic cat, you will get one. And it will waste less resources than a human brain, while thinking about what makes cat "epic"  |
Posted By: NancyGold
Date Posted: 17 June 2022 at 4:17am
|
And here I made it absolutely clear I want the most awesome burning tree AI can imagine:  Now this dall-e mini is a reduced version, much simpler than the cutting edge research, and has a barebone training set.
|
Posted By: NancyGold
Date Posted: 17 June 2022 at 4:27am
|
Apparently it can get even more ridiculous!    |
Posted By: specialmin64
Date Posted: 18 June 2022 at 4:19am
|
well, I have to agree with Dall-E Mini, the last 2 image batches are EPIIIIIIIIIC!!!  |
Posted By: NancyGold
Date Posted: 18 June 2022 at 9:01am
|
Originally posted by specialmin64 well, I have to agree with Dall-E Mini, the last 2 image batches are EPIIIIIIIIIC!!! But you have to be explicit, because by default AI assumes you want a boring version.    |
Posted By: looloopaa
Date Posted: 20 June 2022 at 11:55pm
|
It's really interesting how these things are able to mash images together. I think they definitely operate better when trying to reproduce more painterly images due to the somewhat impressionistic results. Some of my friends have made some funny images out of it though, such as "Steve Irwin Catching a Garden Hose." |
Posted By: NancyGold
Date Posted: 21 June 2022 at 8:47am
|
Originally posted by looloopaa It's really interesting how these things are able to mash images together. I think they definitely operate better when trying to reproduce more painterly images due to the somewhat impressionistic results. Some of my friends have made some funny images out of it though, such as "Steve Irwin Catching a Garden Hose." More advanced models give photorealistic results, down to proper light source, reflections and diffuse interreflection.
|
Posted By: specialmin64
Date Posted: 21 June 2022 at 5:25pm
|
yeah Nancy's right, I think "mashing images together" is an understatement, especially for the more powerful ones like Dall-E 2 (as opposed to the far less powerful Dall-E Mini), which seems capable of understanding perspective, composition fundamentals and light/particle effects. Some are quite convincing at photo-realism- one of my favorite examples being thispersondoesnotexist dot com, which randomly generates believable human faces every time you refresh the page (it tends to generate a hideously malformed demon around the image's periphery now and then too, if you're lucky). Surely AI-generated pixel art isn't too far behind |
Posted By: NancyGold
Date Posted: 22 June 2022 at 3:21am
|
Originally posted by specialmin64 yeah Nancy's right, I think "mashing images together" is an understatement, especially for the more powerful ones like Dall-E 2 (as opposed to the far less powerful Dall-E Mini), which seems capable of understanding perspective, composition fundamentals and light/particle effects. Some are quite convincing at photo-realism- one of my favorite examples being thispersondoesnotexist dot com, which randomly generates believable human faces every time you refresh the page (it tends to generate a hideously malformed demon around the image's periphery now and then too, if you're lucky). Surely AI-generated pixel art isn't too far behind A person with Dall-E 2 access says it has no troubles generating pixel art and provides an example: https://www.youtube.com/watch?v=tZdHxkx4i4w |
Posted By: specialmin64
Date Posted: 22 June 2022 at 4:46am
|
well, guess I'll die, then. It's been nice knowing ye |