Google’s Imagen AI: Outrageously Good! 🤖

Dear fellow scholars, this is two minute papers with Dr. Zsolnai. I cannot tell you how excited I am by this paper. Wow. Today you will see more incredible images generated by an AI. However, not from OpenAI, but from Google. Just a few months ago, OpenAI's image generator AI, called Dall? E2 took the world by storm. You could name almost anything. Cat, Napoleon, a teddy bear on a skateboard on Times Square, a basketball player dunking as an explosion of a nebula. And it was able to create an appropriate image for it. However, there was one interesting thing about it. What do you think the prompt for this image must have been? Hmm, not easy, right? Well, it was a sign that says deep learning. Oh yes. This was one of the failure cases. Please remember this now. We always say that in research. Do not look at where we are, always look at where we will be. Two more papers down the line. However, we didn't even make it two more papers down the line. We barely made two months down the line. And. And here it is. This is Google Research's incredible image generator AI Imagen. This technique also looks at millions of image and text description pairs and learns what people mean when they say, this is a guitar or this is a panda. But the magic happens here. Oh yes. It also learned how to combine these concepts together and how a panda would play a guitar. The frets and the strings seem a little wacky, but what do we know? This is not human engineering. This is panda engineering or robot engineering. Something for a panda. We are living crazy times indeed. So what is different here? Why do we need another image generator AI? Well, let's pop the hood and look inside. Oh yes. Two things that will immediately make a difference come to mind. One, this architecture is simpler. Two, it learns on longer text descriptions. And hopefully this also means that it can generate text better. Let's have a look. What are you seeing? What I am seeing? That is not just some text. That is a beautiful piece of text. Exactly what we were looking for. Absolutely amazing. And these are not the only advantages. You know, I am a light transport researcher by trade. I promise to myself that I'll try not to flip out. But hold on to your papers and holy mother of papers. Look at that. It can also generate beautiful refractive objects. That duck is truly a sight to behold. My goodness. Now I will note that Dall E2 was also pretty good at this. And if we plug in DeepMind's new Flamingo language model. Would you look at that? Is this really happening? Yes, that is an AI commenting on a different AI's work. What a time to be alive. We will have a look at this paper too in the near future. Make sure to subscribe and hit the bell icon. You really don't want to miss it when it comes. And you know what? Let's test it some more against OpenAI's amazing Dall E2. See how they stack up against each other. The first prompt will be a couple of glasses sitting on the table. Well, with Google's Imagen. Oh my, these ones are amazing. Once again, proper refractive objects. Loving it. And what about Dolly too? There's one with the glasses with an interesting framing, but I see both reflections and refractions. Apart from the framing, I am liking this. And the rest? Well yes, those are glasses sitting on the table, but when we say a couple of glasses we probably mean these and not these. But that's really interesting. Two AIs that have a linguistic battle here. Imagine showing this to someone just 10 years ago. See how they would react. Loving it. Also, I bet that in the future these AIs will work like brands and products today, where people will have strong opinions as to which ones they prefer. The analog warmth of imagen or the three year warranty on Dall E4. And wait, you are all experienced fellow scholars here, so you also wish to see the two tested against each other a little more rigorously and will do exactly that. The paper checks the new technique against previous results mathematically. Or we can ask a bunch of humans which one they prefer and wow. The new technique passes with flying colors on both. And once again Dall E2 appeared just about two months ago. And now a new follow up paper from Google that tests really well against it. This is not two more papers down the line, not even two more years down the line. This is just two more months down the line and a year before we had Dall E1. And see how much of a difference OpenAI made in just a year. Now I am almost certain that this paper has been in the works for a while and they added comparisons against Dall E2 at the end. But still, a follow up paper this quickly. The pace of progress in AI research is absolutely incredible. Bravo Google. What a time to be alive. So does this get your mind going? What else would you use this new technique for? Let me know in the comments below. This episode has been supported by Lambda GPU Cloud. If you're looking for inexpensive cloud GPUs for AI, check out Lambda GPU Cloud. Get this, they've recently launched an Nvidia RTX A6000 with 48 gigabytes of memory and hold on to your papers because Lambda GPU Cloud can cost less than half of AWS and Azure. 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. Our thanks to Lambda for their long standing support and for helping us make better videos for you. Thanks for watching and for your generous support and I'll see you next time.