You guys seen this? Looks awesome.

Ok, so whats the catch?

Off the top of my head, I see two (related) issues that aren't addressed fully by the video.

Firstly, how are the artists supposed to generate that much content? As they note at 2:55, they have more polygons in one cubic metre of dirt than in any game that doesn't use procedural generation. I assume they're expecting most of the content to be procedural or scanned? (I note that all the trees on the island look exactly the same. I'd rather have a variety of less ultra detailed tree, than one ultra detailed tree copied everywhere...)

And secondly, at 21,062,352,435,000 polygons in that island... this is going to get to the user's computer how exactly? And once there, it's going to get pumped to the graphics card how? There are some I/O bandwidth issues that appear to have been totally glossed over.

Still, looks pretty cool, even if the graphics shown could be made without this breakthrough technology (pre-rendered?). I'd be interested to see how that point cloud stuff could be integrated with physics, to allow a wind to blow the tree branches about, for example (calculation of forces between the atoms allowing the branches to bend and spring in a realistic way?)

Cool

Holy ****. I would like to see this in action.

Hoax until proven otherwise.

Their website claims that they 'convert polygon models to voxels' and that the speed is made possible by a special 'search algorithm.' Everything about what they claim and how they claim it wreaks of hoax, including and especially hoodwinking the Australian government for a grant.

I am very curious what Jon thinks of this, all the same.

I have some experience with point cloud rendering.

It's not immediately clear what rendering method they are using. According to the video it is a software solution with continuous LOD adjustment, which tells me that it is probably not a surface reconstruction approach. I expect they are using a splatting algorithm. If true, this video isn't very interesting; it is a very modest step forward from previously-published results (e.g. QSplat.)

There are problems with this approach, but none of the ones Giraffe mentioned. The main issues are animation and self-shadowing, which are not easy to do without visiting the entire point cloud. Content creation isn't a big problem, because modern game art is already created at an extremely high resolution (Mudbox, ZBrush - tens of millions of polygons.) This just changes the last compile step. I/O also isn't much of a problem... current splatting algorithms are bottlenecked by fillrate and compute (from tree traversal, which has favorable cache characteristics. n.b. rendering multiple instances of the same model amortizes the traversal cost.)

I bet John Carmack will have a marketable solution before these people do.

So if I am understanding this correctly, the incredible speed is possible because they take something complex but repeat it a zillion times, because using this method, having multiple instances of the same object is not costly at all. So a huge drawback of this is that there won't be the possibility for a lot of unique content, and in fact, it doesn't seem to mesh at all with procedural generation.

Look at the paths through the grass. There are only 6 or 7 unique tiles — hyper-detailed tiles — but tiles. And they repeat endlessly like a bad set of textures.

gah, let me delete youu

Well, broadly speaking, the speed (and detail) is possible because you can abandon rendering the object if you think the extra work won't pay off. That's something you can't do with polygonal models.

This would work fine for procedural content. You wouldn't need to generate a level of detail unless someone is looking at it.

Edit: Remember that I can't comment specifically on this implementation. Everything I'm saying is very generally about real-time point cloud rendering.

So the only major drawbacks are animation and lighting, then? What about space?

That goes without saying, I think.

This thing comes around every few years. It's an investment scam.

Meh.

Maybe that would explain why he's using a really stupid fake voice throughout the whole thing...

This reminds me of the Phantom gaming console.