YouTube video commentary (LARRUZ / Dr. Zsolnai) about a newly published text-to-3D technique that reportedly produces much higher-detail 3D geometry from text prompts. The speaker contrasts this paper with the previous "dream fusion" approach, stating the new method yields ~8× higher resolution while also running ~2× faster, demonstrates varied and imaginative outputs (e.g., "car made out of sushi", personalised pet models), and notes planned integration into NVIDIA's Picasso framework and the Adobe suite. The presenter also mentions that an even newer Gaussian-splatting-based technique already surpasses this work and teases a follow-up episode.
Questions this video answers
November 3, 2023.
Eight times (8×) higher resolution.
About two times (2×) faster.
NVIDIA Picasso and the Adobe suite.
Weights & Biases.
Chapters
I am very excited about this paper. So what does it do? Magic. According to the authors, that is not text to image where we write a text prompt and get a beautiful photo or painting, but something else. Text to 3D. Hmm. Now that would be fantastic. I would say it qualifies as magic. We write something and out comes 3D geometry that we can put into a video game or any kind of virtual world or teleconferencing. Anything you wish. Dear fellow scholars, this is two minute papers with Dr. Zsolnai. Now wait a second. The previous dream fusion technique was capable of doing that. And this was a paper that was published just a year ago now. Yes, but this one has two problems. Problem number one, it takes a long time. And problem number two, we still get coarse results. Now, does this paper really have the magic they promised? Well, let's have a look together. Whoa. Look at the difference. I immediately see that this has so much more detail. It has eight times higher resolution than the previous method. It. And it gets better. We noted earlier that they are not only coarse, but slow too. The coarseness is gone. Fantastic. But I imagine that this is now probably even slower. In return, we don't get all this quality for free, do we? That cannot be true. Wow. It is not only free, it is even twice as fast as as the previous technique. I also love the variety of objects it can conjure up. We can even ask for more imaginative things like a car made out of sushi. That's good. Now clearly, as all of you fellow scholars see here, this is not the quality of geometry that you would put into those fancy AAA quality games. But at this point I. I feel that we are just two more papers away from that kind of detail. Now there is plenty to love here and somehow it gets even better. We can even write a prompt and if we would like to refine the results, we can rewrite parts of the prompt and get a very similar result. It works great with this baby bunny. But I'll note that if we modify the prompt too much, we. We might get an entirely different scene. So this one requires a light touch. And it gets even better. We can have a collection of photos of our cat and then ask the AI to create a virtual version of this cat. But riding on a bike? Just imagine meeting your own cat in a video game. Love it. So when can we try it? Well, Nvidia is planning to include it in Picasso, their generative AI framework, which is hopefully going to be widely available soon. It will be part of the Adobe suite too. So quite Honestly, this will be everywhere real soon. What a time to be alive. If you cannot wait to give it a go and you wish to get notified when it becomes available, that is a possibility through the link in the description. And one more thing. I have surprising news. This technique has already been surpassed. Yes, you heard it right. The pace of progress in AI and computer graphics research is so incredible that there is already something even better than this out there. I'll give you a hint. It is a Gaussian splatting based technique. I can't wait to tell you all about it in the next episode. Subscribe and hit the bell icon to not miss out on it. We are sponsored by Weights and Biases, experiment tracking, Model evaluation and production monitoring for your Deep Learning products and LLM apps. It is the best. Everyone is using it. Try it out now at WNB Me Papers or click the link in the description below.
