Midjourney AI: A League Above DALL-E 2!

Dear fellow scholars, this is Two Minute Papers with Dr. Zsolnai. Today we are going to have a look at the incredible capabilities of Midjourney. A text to image AI that creates fantasy images that are so good it's frankly almost unbelievable. It is a league above any other AI I have seen so far. Here, I'll show you five of my favorite examples in in a moment. Now, what is this? Text to image means exactly what it says. We write a short description and the AI paints our image. But things are changing now. We may not even need to be writing our own prompts. What? Why is that? Well, the GPT4AI also just came out. This can be used with ChatGPT, OpenAI's amazing assistant and one. Hold on to your papers, fellow scholars, because we can ask it to break down the key elements of food photography. What kinds of food should we use? What should the background look like? What about lighting, composition and so much more? Photography is an art after all. So what did we get here? Did we get some sort of tutorial or some ideas? Well, don't forget this is a chat assistant, so we can ask it to tabulate all these ideas. And oh yes, we get a table with detailed instructions for each photo. These instructions can be dumped into a prompt which ChatGPT will happily do for us. And then what do we do with this? Of course, we give this to Midjourney and create these beautiful hamburgers. Look, these instructions were truly amazing. We can go for a retro style top view, or if this made us super hungry and we wish to get over the top, we can do this too. All this with the help of two AIs combined. Wow. The same concept can also be applied to interior design as well. The lighting and the mood on this one results in a stunning picture. I don't know about you, but I still cannot believe that an AI made this. Or two AIs if you will. Now too, we can also create not only images, but videos with these methods. Look, if we have made a character that has the right look but not quite the right age, or we wish to age this character over time, this video is not to be used as is so. So it should not be confused by techniques that are specifically tailored for that. But if we stop it at individual frames, these can be used as images. 3. It is now able to do something that none of the previous text to image AIs were able to. Interestingly, none of the previous techniques were able to synthesize human hands with five fingers. Also, they were often not quite the right shape. And can that really be? Is it finally solved? Look. Let's see this one. Wow. Now those are not only five fingers everywhere, but the detail and the variety of these images are absolutely incredible. I will now officially allow it to give a big thumbs up to itself. 4 Midjourney is currently on version 5 in and I found it really surprising that improvement has still not plateaued. For instance, in version 4 it imagined holograms and people to look like this, which is not too great. But here are the same prompts on version five. Now that's what I'm talking about. And just imagine what version 7 will be capable of. Two more versions down the line. But we are not done yet. Not even close. 5. We have a few more amazing results. And look who this one is coming from. Yes, that's right. The legendary computer graphics researcher Donald Mitchell. Among many other things, he is the author of the legendary Mitchell Natural Valley filter which contributes to creating beautiful high quality images in light transport simulations and and more. Here he asked for some robot sushi and I love how imaginative these results were. These works somehow both pass for robots, but they also seem like edible lunch at the same time. Bravo. And what better way to wash it down than asking for a cocktail from an alien planet? Once again, they certainly have a strong alien feel. And I have to say some of these also look drinkable. But be careful. Who knows what they contain. Loving it. And one because I cannot resist. My favorite, perhaps everyone's favorite macro photos. Close up photography of typically small subjects. I have found this apple pie and this Dragoneye. Not only the best AI made macro images I've ever seen, but but these are so good they might even rival a journeyman photographer's images. What a time to be alive. I will note that here is the point where you usually get some deeper analysis in the video. However, it is not possible here. This work likely builds on previous papers, but it does not have a paper itself that I know of. However, if a technical document describing it appears, I will be very excited to have a look and and show it to you. Also. I expect that you are now very excited to try this. Can you? Yes. If you wish to try it, the first 25 images are free. After that they are asking for a paid subscription. So which image from these was your favorite? Have you done your own experiments? Let me know in the comments below. And for now, let the experiments begin.

Weights and Biases provides tools to track your experiments in your deep learning projects. What you see here is their amazing sweeps feature which helps you find and reproduce your best runs. And even better, what made this particular run the best. It is used by many prestigious labs including OpenAI, Toyota Research, GitHub and more. And the best part is that Weights and Biases is free for all individuals and academics and open source projects. Make sure to Visit them through wnb.com papers or just click the link in the video description and you can get a free demo today. Our thanks to Weights and Biases 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.