Károly Zsolnai-Fehér breaks down OpenAI's DALL-E 3, which claims better prompt following, higher image quality, and tight ChatGPT integration for consistent characters and text rendering. No research paper exists yet, and the examples shown are likely cherry-picked.
Big day today. Dall? E 3, the third version of the legendary text to image AI is coming. No, we can't try this yet. There is no product or paper yet, only the initial announcement. Now, it seems that it does three things better than all the other techniques out there. So what does it do? Well, well, first it listens. Now I hear you asking, doctor, what do you mean by that? Well, well, with other techniques, important parts of our prompt often get lost. However, not here. When we enter a detailed prompt, it really tries to take all of it into consideration. If I write the prompt, I want to get that mustache to be noteworthy. And this one I would say is. And some of these prompts can get bloody long. And I just kept looking and looking. But I think I might see everything that was requested here. Even for things that are really difficult to paint or sometimes even imagine. Try to imagine a whirlwind of porcelain fragments, harmonious fluid forms in a dreamlike atmosphere, and so on. Excellent. 2. I am wondering, can this really compete? I mean, midjourney and stable diffusion have become so incredibly good. I can imagine that even Dall? E3 might have a hard time competing. Now we have a little indication here. Oh yes, the legendary prompt from Dall? E 2. An expressive oil painting of a basketball player dunking depicted as an explosion of a nebula. And now look, this is what it looks like with version 3. More detail, more definition, more life. I find it a great deal better. Great job, Little AI and three. Now hold on to your papers, fellow scholars, for better integration with ChatGPT. You see, we don't even need to write the prompts directly. We can even ask it to make up a new character. Let's call him Larry the hedgehog. And we get Larry. Larry is amazing. So let's get some more images. And now you immediately see that getting several images created about the same character is not a trivial matter completely at all. We've looked at a bunch of papers that tried to pull that off and now it is here. But every hedgehog needs a house, so it can imagine that too. That is a beautiful house. And look at that text. Fantastic. I really want that. Because we have been promised proper text support in text to image. And at least for me, it never quite worked out without significant cherry picking. I hope this is gonna be better here. And then finally we can ask for stickers of the Amazing Larry and even a bedtime story with him. I have a 7 year old daughter and just thinking about the amount of fun we are going to have with this gives me goosebumps Love it. Thank you so much. Now one more thing. There is no paper here. I can only comment on what I see in the announcement and the cases that we see here are us not the average but the best case scenarios. However, we will be able to try this and find out soon. No doubt. What a time to be alive. I was also happy to see that it will not create an image in the style of a living artist. And when I watch their video, what do I see? Proper and I mean proper scholarly representation. That is what I see. Well done fellow scholars. If you're looking for inexpensive cloud GPUs for AI, Lambda now offers the best prices in the world for GPU Cloud compute. No commitments or negotiation required. Just sign up and launch an instance and hold on to your papers. Because with the Lambda GPU cloud you can now get on demand H100 instances for just $1.99 per hour and yes, $199 and they are one of the first cloud providers to offer publicly available on demand H100 access. Did I mention they also offer persistent storage? So join researchers at organizations like Apple, MIT and Caltech in using Lambda Cloud instances, workstations or servers. Make sure to go to lambdalabs.compapers to sign up for one of their amazing GPU instances today.
