by powersnail on 8/20/22, 7:23 PM
DALL-E is still highly probabilistic in its judgement. For instance, in this article, it keeps putting "fire" in the the background on something that is likely to be on fire, rather than lighting up the person.
I have a similar experience. In my own experiment, I can't get DALL-E to turn off the street lamp at a bus stop in the darkness. I've tried "no light", "broken street lamp", etc.; no use. Any mention of "street lamp", the scene will include a working street lamp.
It's just more probable that a scene with a lamp in the darkness must have that lamp providing light, and this is something that DALL-E will not break out of.
by randymy on 8/20/22, 8:51 PM
Worth noting that DALL-E automatically “rejects attempts to create the likeness of any public figures, including celebrities". So, you wouldn't be able to get an image that included the 4 Liverpudlians. It does allow you to create fake faces. Might be fun to try and recreate Miles Davis Tutu, Aladdin Sane, Piano Man.
by Michelangelo11 on 8/20/22, 6:31 PM
Man, after seeing Stable Diffusion's output, DALL-E's looks just janky. Like watching a propeller plane after seeing a jet.
Crazy how fast the tech is moving.
by cowmix on 8/20/22, 6:40 PM
After getting access to the beta, combined with all these HN posts -- I've determined DallE2 is neat but no where as great as the initial samples made me believe.
by nprateem on 8/20/22, 7:37 PM
An upvote for whoever can give me a prompt to generate an image of someone who's been massaged so much their body has been flattened, as if they were made of dough or jelly or something.
I spent ages on this earlier getting nowhere. I'm starting to think DALL-E is better if you don't really know what you want and you're just fishing for ideas.
by soneca on 8/20/22, 6:41 PM
Have anyone given a prompt to Dall-e of designing a company website and included “make it pop!”?
Maybe the AI will finally get what designers always complained about annoying clients.
by yummybear on 8/20/22, 6:41 PM
I love seeing people experiment with this technology. You can feel we’re on the cusp of something great - whatever it is, we’re just not quite there yet.
by teddyh on 8/20/22, 6:27 PM
>
The question is, when the music blows up and the artwork becomes a signature, like the Rolling Stones' Tongue & Lips, who will own the copyright?That’s what trademarks are for.
by w0mbat on 8/20/22, 9:13 PM
How do you know that the album covers are not part of the corpus of images that DALL-E was trained on in the first place?
by phonescreen_man on 8/20/22, 8:52 PM
Interestingly related, I just used AI image generation to create my EP cover.. first I tried running luciddrains dall e 2 PyTorch implementation using the prompt “death by Tetris EP album cover 2022” unfortunately I am using a Mac Pro so the gpu was not able to work. Then I tried imagen PyTorch implementation and used same keyword. This time it was working with the CPU unfortunately 2 days in we had a power outage so I had something but nothing complete. So I fed the generated image into the google dream generator and got my album cover!
https://willsimpson.hearnow.com/
by fimdomeio on 8/21/22, 8:45 AM
There are a lot of articles focusing on how close does DALL-E match some prompt, but I wonder if this is a suboptimal way to explore the medium.
What if you can get a lot more out of it by embracing the unexpected responses. Can it be a tool for exploring lateral thinking? You provide a prompt computer responds with images that are a prompt to human.
A baby swiming next to a dolar bill outputs a distorted person face inside a dolar bill with some baby features, could be the start to a rabbit hole of prompts and images where you'll end up with something completly different than your initial expectations.
by tsimionescu on 8/20/22, 9:15 PM
It's interesting that the prompts that would do badly in a Google image search also seem to be the ones that make poor prompts. Basically, it seems that rather than describing a scene, you have to try to give an analogy for some image(s) that it might have in its training set - which is why, I believe, "banana in the style of Andy Warhol" produces a much higher quality result than "Outline of prism on a black background in the middle of scene splits a beam of light coming from the left side into rainbow on the right side".
by xwdv on 8/20/22, 6:32 PM
Although AI artists will destroy a lot of jobs, it will also create demand for new jobs for people who specialize in “paint overs” – taking a high concept output created by AI artists and touching it up to perfection.
Or perhaps even beyond just a paint over, and into the realm of recreating an entire AI artwork but with a human touch to get details just right.
Looking forward to it.
by waveywaves on 8/21/22, 8:26 AM
DALL E works really well if you are specific enough. When you don't get the intended result, it helps to identify the element which wasn't generated properly and then improve your description of the same.
"Two men, one of whom is on fire, shaking hands on an industrial lot." can be rewritten as, "Two men, shaking hands, standing on an industrial lot. Person on the right is on fire. Camera is 30 metres away."
You can go into more specifics of the framing and the angle from which you want the picture to be take. By default, DALL E will give you the most realistic generations to your prompts unless you mention "digital art" or a particular art style. I have gotten the best results when generating art instead of photos.
by spike021 on 8/21/22, 3:31 AM
I haven't gotten to try it for myself, but I've read a few of these blogs that take you through generating examples or even look-alikes to older art pieces.
It surprisingly reminds me a lot of when I traveled to Japan without knowing really any Japanese. I needed to communicate not only with friends who don't know much English either, but also other people (like restaurant wait staff, train station staff, etc.).
I used Google Translate often, but many times I or my friend(s) (or the other people) would need to re-write our statements a few times until the translation result clicked well enough in each other's languages to be understandable.
by google234123 on 8/20/22, 8:16 PM
Is the issue with faces a deliberate choice by the devs?
by pjgalbraith on 8/20/22, 10:52 PM
by _HMCB_ on 8/21/22, 12:51 AM
Very cool. But it just goes to show the impact of human creativity. The conceptual aspect.
by bryanrasmussen on 8/20/22, 7:22 PM
No Smell the Glove cover, this is a black day for rock and roll!
by remote_phone on 8/21/22, 2:03 AM
If you gave those same instructions to humans I’m sure the output would be just as varied. I’d be interested to see a comparison between dall-e and humans.
by wodenokoto on 8/20/22, 7:32 PM
I'd love to see what it had come up with if simply prompted for "Album cover for Nevermind by Nirvana"
by kaffeeringe on 8/21/22, 1:18 PM
I wonder, how much energy is being burned for these kinds of experiments.
by dsign on 8/20/22, 9:22 PM
It's going to leave all those artists without a job, you just wait!!
by NonNefarious on 8/20/22, 10:06 PM
Went to use my invite, and OpenAI demands your PHONE NUMBER.
No excuse for it. Screw that.
by sgt101 on 8/20/22, 6:35 PM
Look, it's trained on these images.
It's really great and cool and all - but it's retrieving things that it was trained on.
Show me something original it did.
by machinekob on 8/20/22, 6:31 PM
Is i do smth with DALL-E auto top hacker news post i saw like 20 post like that in past 2 weeks.
by system2 on 8/20/22, 7:04 PM
DALL-E still seems very useless. Reminds me of the hype of Cardano.
by andreyk on 8/20/22, 6:29 PM
I wonder how long the novelty of DALL-E will persist. HN seems to upvote anything titled "I did X with DALL-E". This is a fun post, but it's not that interesting or surprising. Still worth a look don't get me wrong, but personally didn't learn anything new from it. (eg recreating the famous pink Floyd cover with "Outline of prism on a black background in the middle of scene splits a beam of light coming from the left side into rainbow on the right side" unsurprisingly worked somewhat well).