OpenAI’s new AI image generator pushes the limits in detail and prompt fidelity

Submitted ⁨⁨1⁩ ⁨year⁩ ago⁩ by ⁨Voyager@psychedelia.ink⁩ to ⁨technology@lemmy.world⁩

https://arstechnica.com/information-technology/2023/09/openai-announces-dall-e-3-a-next-gen-ai-image-generator-based-on-chatgpt/

On Wednesday, OpenAI announced DALL-E 3, the latest version of its AI image synthesis model that features full integration with ChatGPT. DALL-E 3 renders images by closely following complex descriptions and handling in-image text generation (such as labels and signs), which challenged earlier models. Currently in research preview, it will be available to ChatGPT Plus and Enterprise customers in early October.

Like its predecessor, DALLE-3 is a text-to-image generator that creates novel images based on written descriptions called prompts. Although OpenAI released no technical details about DALL-E 3, the AI model at the heart of previous versions of DALL-E was trained on millions of images created by human artists and photographers, some of them licensed from stock websites like Shutterstock. It’s likely DALL-E 3 follows this same formula, but with new training techniques and more computational training time.

Judging by the samples provided by OpenAI on its promotional blog, DALL-E 3 appears to be a radically more capable image synthesis model than anything else available in terms of following prompts. While OpenAI’s examples have been cherry-picked for their effectiveness, they appear to follow the prompt instructions faithfully and convincingly render objects with minimal deformations. Compared to DALL-E 2, OpenAI says that DALL-E 3 refines small details like hands more effectively, creating engaging images by default with “no hacks or prompt engineering required.”

source

Comments

Sort:hotnew top

simple@lemm.ee ⁨1⁩ ⁨year⁩ ago
That avocado image is insane. I’ve yet to see any image prompt AI get text and composition anywhere near this level. Mindblowing to know there were zero edits. I really want to try this now.

source
MysticKetchup@lemmy.world ⁨1⁩ ⁨year⁩ ago
One of the things I’ve noticed from testing these generators is that they look very good when someone is showing off cherry-picked photos or you are going in with little to no expectation of what you’ll get. The more you have a specific image in mind or the more you want certain details the worse it gets. And you quickly realize that when you generate things from similar prompts over and over the model gives you the same results but slightly adjusted.

Obviously art generators look bad for artists right now, but I think once the new toy factor wears off people will realize they aren’t as good as they seem. There’s a lot of smoke and mirrors involved and once you’ve seen a good amount of AI photos it gets easier and easier to pick them out of a lineup. They’re closer to advanced stock photo generators, in terms of what you can actually get from them. Companies’ race to the bottom means that this is going to have effects on artist jobs, but I think the next “revolution” in art is going to be having human art as a selling point the same way stuff like fully orchestrated music or hand-crafted things are.

source
- thehatfox@lemmy.world ⁨1⁩ ⁨year⁩ ago
  I wish more people realised this. It’s much harder to create very specific images with the current image generation tools than most people seem to think, which is creating an inaccurate view of the technology in the public eye.
  
  The generator will create something inspired by the prompt it is given, but it can be very hard to make it match the output the prompt writer imagines when writing the prompt. There are various tools that can refine and narrow the generator’s output, to try and control things like posing, composition, style etc and to redraw details. But even then it’s often pot luck as to the output. The generated images aren’t necessarily bad, just not what was wanted.
  
  I think the comparison to stock photo images is apt, current image generators are great for creating themed but somewhat generic images. The tools are going to continue to advance, and they are useful in for some applications already. But they are still a long way off from truly replacing human artistry.
  
  source
  - lloram239@feddit.de ⁨1⁩ ⁨year⁩ ago
    The crux with that argument is that the artists is the only one that cares about specific output, meanwhile the art consumer doesn’t. When somebody plays a game or watch a movie, they don’t know what to expect, that’s part of the fun, they just care about it being good. So as long as the output is good enough for the consumer, whatever the artists thinks about it, really doesn’t matter, assuming they still have a job to begin with.
    
    source
- lloram239@feddit.de ⁨1⁩ ⁨year⁩ ago
  
  And you quickly realize that when you generate things from similar prompts over and over the model gives you the same results but slightly adjusted.
  
  That’s quite true, however it’s worth keeping in mind that this is largely not due to a limit in the model itself, but a limit how that model interfaces with the human. Text just isn’t very good to get specific results, especially not when it lacks the incremental refinement that you can do in ChatGPT with follow up prompts.
  
  On the other side if I take StableDiffusion with ControlNet, instead of just a text prompt, I can generate far more specific results, as I can feed other images and sketches into the generation.
  
  but I think once the new toy factor wears off people will realize they aren’t as good as they seem.
  
  Quite the opposite, there is a ton of hidden potential still left to uncover. We have barely even started to train them on video or 3D data, integration of image models with newer language models is also a work in progress and integration into old-school image manipulation tools has just began as well.
  
  Worth keeping in mind that Dalle-1 isn’t even three years old. We are basically still in the Atari2600 days of image generation.
  
  Meanwhile Dalle-3 comes along and can produce this level of quality with a complete generic prompt: “A fan-art of Guardians of the Galaxy Vol. 3” on the first try.
  
  I think the next “revolution” in art is going to be having human art as a selling point
  
  The big problem for artists is that AI art drives the value of art down to zero. It’ll be hard to convince anybody to pay hundred of dollars for something when AI can produce something similar in 30sec for free. Worse yet, AI can take any existing image and remix it. The whole idea of a singular static images feels quite restrictive once you played around with AI art for a while, as everything is just a few clicks away from being something different.
  
  I think the idea of AI art as just generators for stock images doesn’t capture the magnitude of the changes that are coming. We are straight up heading into Holodeck territory where you tell the computer what you want and you get it. The AI generators won’t be a tool for the artists, but go right to the users. There won’t be an static image that comes out the other end, the AI will be the medium of media consumption. Just like people today can flip through TikTok, future people will flip through a AI generated stream of content custom made for them.
  
  Wanna play some 2D game with snow and ice? Tell the computer, a couple seconds waiting, and boom here it is. First try. Want lava instead? Done. How do you compete with that as a human when AI can pull that out of thin air in seconds?
  
  source
  - MysticKetchup@lemmy.world ⁨1⁩ ⁨year⁩ ago
    Nah that Guardians of the Galaxy art is exactly what I’m talking about. It makes basic mistakes even a child could point out and looks more long a knockoff. And refining it is just rolling the dice to get a better result, whereas an artist you can actually give feedback they can understand.
    
    The game assets look a little better, but if you look carefully you’ll notice that they don’t tile correctly. It’s 90% there but the last 10% is the hardest part and it’s important especially for large projects and not just single static images. Not too mention they look generic as fuck, you’re not going to get the next Hollow Knight or Darkest Dungeon with an amazing original style from AI, you’re only going to get existing styles mashed together. The more specific the vision for the artstyle the harder it will be to generate it.
    
    Also the idea of a Tiktok feed of AI generated content is exactly why I hate AI art. Sure, go ahead and use it to help existing artists generate cheap assets that would otherwise be random brush strokes. But replacing them? The idea that AI generated slop will have anything close to the quality and meaning of even cheap art is ridiculous. Why would anyone want that when they could have actual art made by real people, more of which exists today than anyone could go through in their entire life?
    
    source
    -> View More Comments
- GenderNeutralBro@lemmy.sdf.org ⁨1⁩ ⁨year⁩ ago
  Getting good results usually requires a prompt that looks like the title of a generic product in Amazon: just an endless stream of keywords.
  
  The examples from simple prompts are cherry-picked.
  
  source
AlmightySnoo@lemmy.world ⁨1⁩ ⁨year⁩ ago
I will be convinced when they learn to draw hands correctly.

source
- Devorlon@lemmy.zip ⁨1⁩ ⁨year⁩ ago
  Here’s an example image from the article.
  
  https://cdn.arstechnica.net/wp-content/uploads/2023/09/plategirl-980x560.jpg
  
  source
  - Chozo@kbin.social ⁨1⁩ ⁨year⁩ ago
    
    from the article
    
    Well no wonder they couldn't find this example.
    
    source
    -> View More Comments
  - Quicky@lemm.ee ⁨1⁩ ⁨year⁩ ago
    Was the prompt “Woman from China”?
    
    source
    -> View More Comments
- tonytins@pawb.social ⁨1⁩ ⁨year⁩ ago
  Because real artists struggle with hands.
  
  source
  - thbb@kbin.social ⁨1⁩ ⁨year⁩ ago
    While there is some truth in this, humans and AI do not make the same type of mistakes with hands.
    
    Humans will rebuild the topological structure of the hand: 5 fingers protruding from a based and get the proportions wrong.
    
    AI will rebuild the image of a hand from the 2d appearance of a hand: a variable number of flesh colored, parallel stripes, and improvise from that.
    
    While both can get it wrong, the errors are not similar.
    
    source
- ZILtoid1991@kbin.social ⁨1⁩ ⁨year⁩ ago
  Making the context window likely helps with stuff, however it still has the issue of "background breaking".
  
  source
- lloram239@feddit.de ⁨1⁩ ⁨year⁩ ago
  Seems to be about 50/50, quite a few good looking hands, but still plenty of crocked fingers.
  
  source
  - AlmightySnoo@lemmy.world ⁨1⁩ ⁨year⁩ ago
    Yup that’s the thing with most of generative AI models, they have no implicit 3D modelling of the world. So depending on perspective, a real 2D image may give the impression that there are 2 or 3 fingers only but the model doesn’t know that that’s just because of perspective.
    
    source
PopOfAfrica@lemmy.world ⁨1⁩ ⁨year⁩ ago
Ive already accepted that my graphic design degree is worthless.

source
- 2nsfw2furious@lemmynsfw.com ⁨1⁩ ⁨year⁩ ago
  This is just automatic Photoshop - if all you were doing with graphic design was pasting blond hair onto a brunette, yes, this has really screwed you (or made your job a lot easier). If you’re actually doing any level of design… you’re safe for now
  
  source
PlexSheep@feddit.de ⁨1⁩ ⁨year⁩ ago
closedAI still sucks, even if their (closed!!) Tools are powerful.

source
autotldr@lemmings.world [bot] ⁨1⁩ ⁨year⁩ ago
This is the best summary I could come up with:

On Wednesday, OpenAI announced DALL-E 3, the latest version of its AI image synthesis model that features full integration with ChatGPT.

DALL-E 3 renders images by closely following complex descriptions and handling in-image text generation (such as labels and signs), which challenged earlier models.

Judging by the samples provided by OpenAI on its promotional blog, DALL-E 3 appears to be a radically more capable image synthesis model than anything else available in terms of following prompts.

While OpenAI’s examples have been cherry-picked for their effectiveness, they appear to follow the prompt instructions faithfully and convincingly render objects with minimal deformations.

DALL-E 3 also appears to handle text within images in a way that its predecessor couldn’t (some competing models like Stable Diffusion XL and DeepFloyd are getting better at it).

Microsoft’s Bing Chat AI assistant, also built on technology from OpenAI, has been able to generate images in conversation since March.

The original article contains 420 words, the summary contains 151 words. Saved 64%. I’m a bot and I’m open source!

source
avater@lemmy.world ⁨1⁩ ⁨year⁩ ago
always wanted to try it out but no way I’m giving my phone number to them, although I understand their approach to reduce bot accounts.

source
- Dkarma@lemmy.world ⁨1⁩ ⁨year⁩ ago
  Try stabil diffusion
  
  source
echodot@feddit.uk ⁨1⁩ ⁨year⁩ ago
What a time to be alive!

source
Voyajer@lemmy.world ⁨1⁩ ⁨year⁩ ago
Nice username!

source