Was blown away by this video https://developer.nvidia.com/gtc/2019/video/S9367

Here is the tech home page http://www.lightmass-dynamics.com/#Intro

I can't imagine how QNM works, can you?

There have been groups talking about their “unlimited detail” since the 1980s.

Most use a variation on fractals for the detail rather than textures, and variations on parametric surfaces as opposed to modeled polygons. People have been doing deep Mandelbrot fractal zooms since the 1970s, possibly earlier, and many space-filling systems and beautiful volumetric patterns have been discovered. Every few years a company rediscovers that you can have infinite detail both in models and textures, and puts their marketing budget on it.

Others turn the entire world into a voxel system, much like you see with Minecraft at a distance, and the “unlimited detail” is just how the blobs of the world are visualized. A company called Euclideon did this about 5-10 years ago, spending most of their money on marketing rather than implementations. Their marketing campaigns were highly mocked by data visualization and graphics professionals.

The big issue with all these systems has been that they completely saturate all the CPUs thrown at them with rendering the massive point clouds or deep fractal processing, rather than doing other useful work like actually playing a game.

It is entirely possible that this variation will be different from the hundreds that tried and failed before it. I'm not holding my breath.

“unlimited detail” was more about SVO. This is different.

Very interesting.

No idea how it works or what it is doing at all, but like many of those AI approaches it looks eerie, even if very impressive.

Now, i think we can put those details on screen also with our ancient tech, avoiding this problem.

The main limitation i see on the long run is not processing power, but storage. This could be solved with streaming all those details from servers while the player traverses the game, beating the traffic of Youtube?

A much better option would be procedural generation of close up details. And maybe that's where ML becomes truly interesting for us.

Curious if this shown work goes into this direction. Make some photos from rusty ship, train NNs, and then those (currently bored) tensor cores create such texture and displacement as a background task… would be awesome.

However - knowing nothing about ML i'll try to do it the old school way… using ancient vector fields, fluid simulation and noise function crap :D

@swiftcoder

swiftcoder said:

This first part just looks like billboard clouds, and the last part seems like lossy-compression of BVH. Not sure there's much going on here beyond using machine learning techniques to “learn” how to compress data for existing techniques?

What is BVH?

probabilistic distribution of a primitive's state similar to how the quantum wave function mathematically describes a quantum state of a system

trsh said:
What is BVH?

Bounding Volume Hierarchy. It's a tree of bounding boxes (mostly). Traversing the tree lets you find a triangle (or other primitive) at a certain location, within a certain region of space, along a certain ray.

I do not get this part - how can ML help to build better trees? How can they use non-box-shape as bound for RTX? (looks like K-Dops)

swiftcoder said:
This first part just looks like billboard clouds

Yeah, but ‘intelligent’ billboards :D … hmmmm… can they cleverly transform volume data to a given direction to give displaced billboard cloud easy to process?

trsh said:
similar to how the quantum wave function mathematically describes a quantum state of a system

Huh? At this point people usually loose me. Resulting in conclusion like:

frob said:
It is entirely possible that this variation will be different from the hundreds that tried and failed before it. I'm not holding my breath.

They really should try to talk more grounded and seriously. Who buys into odd correlation with quantum physics? That's just philosophy but no information.

Unfortunately their whole web page is like this, leaving us in just doubts. Even Bruce Dell did better here.

Yeah, you probably right.! But emailed them. And they say, that the WEB will be updated soon with more info.

JoeJ said:
I do not get this part - how can ML help to build better trees?

Many types of trees have arbitrary cutoffs and shapes.

There are some regular trees, think quadtrees, octreees, and balanced binary trees, where everything is designed to be balanced and uniform.

But many tree structures and data graphs are irregular in practice, and can be improved by finding good, better, or best partitioning systems.

Often those three are at odds: “Good” is often cheap using basic greedy algorithms. “Best” is usually hard with an expensive optimization pass, such as a search of all permutations “Better” often involves probabilistic solutions, often called AI for marketing purposes, to get something in between.