0:00 [...] [Music] Are we screwed? I mean 3D artists, whether aspiring or already working within the industry. The bombshell paper just dropped along with a live demo. It's called Sparc 3D, or Sparse Representation and Construction for High-Res 3D Shape Modelling, a new research project from Nanyang Technological University. In short, it seems like a much more robust take on photogrammetry more or less, but here's the twist, it works from just a single image. And not only that, it can reconstruct the parts that aren't even visible in the source image. This tech has some serious disruptive potential as far as our industry goes, so I thought we should take a moment to have a look at it. And what could it mean for us all?
0:46 Oh, by the way, if you're enjoying the content, feel free to hit that like button and subscribe! So here's what the public demo looks like. There's a page where you can upload an input image, and the authors say they're planning to release the code once they get all the necessary permissions, but for now it's just this demo. Basically you can start with any image, a photo for instance, and you upload it here, and then you just hit generate, wait a little bit in the queue, the queue is getting longer today due to higher demand, and then boom, you get a 3D model, or actually two alternative reconstructions to choose from. The quality though! Just look at that. Pretty damn impressive! Actually, that was the first thing that stood out to me. If you've ever worked with 3D reconstruction
1:32 or photogrammetry, you'd probably notice it too. The surface doesn't look wobbly. Like, at all. There's none of the typical reconstruction noise you often see in photo scanned models. Then on top of that, there's the elephant in the room. It kind of knows or rather guesses what the object looks like from the other side. From all possible angles, really. I wonder if you agree, that's what strikes most about it. Actually, this is something that just wasn't possible before. First off, I've got to ask, what would you use it for? If you want, you can pause the video and drop the answer in the comments. Would you treat it like photogrammetry 2.0? Maybe use it to recreate cool objects from photos to drop into your scenes as set pieces or props? I can already imagine one use case that's likely to shake up the 3D art ecosystem,
2:19 taking AI-generated images and designs from tools like Midjourney ChatGPT, Grok, Copilot, you name it, and turning them into actual 3D props and assets. Let's see how it plays out, but honestly, it's starting to look almost too easy. You just feed it a single image and it fills in the blanks. Parts that weren't in the image at all. The underside of the object, the top view, hidden geometry, whatever's missing. And I can already see where the hype with all its healthy and unhealthy properties can enter the picture. So what's so worrying about all this? Well, isn't the resulting model kind of a mess? To some extent, yeah, it's still not exactly production-ready. But here's the thing, the surface quality
3:06 feels like a whole generation ahead this time. Flat surfaces are actually flat now, if you know what I mean. It's still a bit of a dense polygon soup, there's no proper topology flow and all of that, not to mention it doesn't support textures and colors, but the way the surfaces connect, it doesn't feel lumpy or broken anymore. It's smoother, more coherent. I'm pretty sure that a few research papers down the road, and with Nanyang Technological University pushing like this, it's only going to get scarier. Or more exciting, depending on how you feel about neural network-generated geometry. Or to take a step back and look at a broader picture, even without neural networks, there's still major progress happening. Take Google's new scene reconstruction tech, for example.
3:51 It reverse-engineers breakcasting and reconstructs geometry using purely algorithmic methods. No neural networks involved. As Dr. Károly Zsolnai-Fehér from Two Minute Papers pointed out, this is another step forward in this field. It really feels like all these advancements are happening all at once. But coming back to Sparc, what really sets it apart from traditional image-based tools like photogrammetry is its ability to guess the geometry on the other side of the object. Not just that, but it also handles thin, intricate parts surprisingly well, I'd say. Take this little antenna, for example, that's something older tools usually struggle with. Now the mesh is still a bit crumpled in places, mostly due to how it calculates surface geometry. There's no hard edges as such, as you can tell.
4:38 Ultimately, this feels like a stepping stone toward better topology in future versions. That being said, it's already a lot more impressive than most of the similar tools we've seen before. The combination of production speed and decent quality pretty much guarantees we're heading for an avalanche of half-baked models, what some might call "AI slop" flooding the market. Because, yeah, it's honestly not that hard anymore to mass-produce models that, while kind of sloppy, are still good enough. Let's be honest, in some cases they're actually pretty good. Whether they're based on photos or whatever else, it's now just a few clicks away. So the industry needs to brace for impact, not just in terms of quality control, but also when it comes to legal and ethical concerns.
5:23 You know, what's the copyrighted status of models generated this way, should the marketplaces like Sketchfab or Unity Asset Store set clearer policies? Remember that time when one Fab user uploaded over 38,000 AI-generated assets? Yeah, that caused a bit of a scandal. Infinite scroll, infinite regret, my god. Many of them will be based on designs generated in Midjourney or whatever AI tool is trending at the time. I'm talking about those AI-generated input images being turned into actual assets. It just seems like a no-brainer workflow at this point, a kind of low hanging fruit, if you know what I mean. Alright, so we've got this cool new way of reconstructing 3D geometry from just a single image.
6:09 What's next? Where do we stand? And what should we make of it? Well, here at Creative Shrimp we think this is exactly that, a tool. A pretty impressive one, sure, but still just a tool. Think of it like photogrammetry. It helps create assets. Even if it does make asset creation way easier, which remains to be seen, tools like Blender are still right in the center of it all. Blender stays the hub, the place where everything comes together and becomes usable. You'll still be doing retopo, cleaning up messy geometry, prepping stuff for animation, kit-bashing parts together, and so on and so on, building the freaking scenes and worlds, doing the real creative work. And from there it's up to you. You get to build your own worlds and tell your own stories. So yeah, that's the optimistic take. As a Blender user, as a 3D artist
6:56 you're still valuable, and you probably will be for quite a while. Oh yes, so many questions. Actually I want to hear what you think of all this. Do you feel threatened by it? Or are you in the other camp, seeing AI-based mesh generation as a chance to level up your 3D workflow perhaps? If you've got a strong opinion, don't hold back by the way, it's completely fine. If you found this video interesting, I would appreciate if you can hit like and subscribe. That's it from me, Gleb Alexandrov at creativeshrimp.com, keeping you updated on the latest in computer graphics. Are we good or are we screwed?
