Polygons, voxels, SDFs... what will our geometry be made of in the future?

[Frenetic Pony]

Dreams is definitely making me believe polys are a dead end technology. The amount of detail in this hurricane scene is damned insane. I'd be wowed if this was running on a PS5. The complex lighting, insanely complex cloud structures, the crazy amount of detail along the ground, and yet it's all running on a PS4 like it's nothing.

 

[jlippo]

Dreams is definitely making me believe polys are a dead end technology. The amount of detail in this hurricane scene is damned insane. I'd be wowed if this was running on a PS5. The complex lighting, insanely complex cloud structures, the crazy amount of detail along the ground, and yet it's all running on a PS4 like it's nothing.

That scene is pretty much perfect for some old school adventure game.

 

[JoeJ]

Dreams is definitely making me believe polys are a dead end technology.

What do you think now, after seeing UE5 demo?
If single pixel triangles are just as fast as atomic splats, which of the two is less crazy?

 

[Shifty Geezer]

Dreams is getting a lot more detail for the memory footprint. Given finite storage, it seems the future will be determined by the technique that is most efficient in data density.

 

[Arwin]

True but the art creation pipeline matters a lot too. So how things develop there will also be interesting. UE5 is brute forcing the assets that can come from movie VFX pipelines. Dreams basically ‘only’ has dreams as a pipeline. And of course there is still a lot missing there like transparency, soft body physics, etc. Still the basics are stunning and can create some amazing looking graphics already. I wish I could share some more of ryan47’s latest videos, his modeling remains Pixar level and the animation he is starting to do is right up there too.

So they will probably coexist and Dream’s solution will probably remain a minority approach for a good while yet, although I wonder if the triangle approach will run into storage limitations - Dreams seems much more efficient there. But I do wonder that if objects can be streamed into memory fast enough, there might soon barely be a limit to what size of level you can create in Dreams (though that holds for all approaches basically).

I am glad we’re seeing some exciting things happening in the field again though. I was showing the AC Odessy tour mode to my gf and my mother yesterday, and it was very impressive to see the Akropolis in its original glory, being able to walk around etc. This will probably be taken to a whole new level soon and then to combine that with VR ...

 

[JoeJ]

I would not wonder if UE5 already uses atomic pixel splats instead single pixel triangles.

 

[chris1515]

I would not wonder if UE5 already uses atomic pixel splats instead single pixel triangles.

Yes Brian Karis said there is some similarity with Dreams.

 

[JoeJ]

Bingo

... was known long before Dreams, how it can beat rasterizer:

 

[chris1515]

Bingo

... was known long before Dreams, how it can beat rasterizer:

I like this too

https://perso.telecom-paristech.fr/boubek/papers/WPBGI/

or this

https://perso.telecom-paristech.fr/boubek/PBGI/

The UE5 with so much polygons is a no go for raytracing. I think you are onto something. Why trace ray against traingle when you have very complex geometry?

 

[JoeJ]

I remember Quake 1 already did it: Render points instead triangles for distant characters.

Why trace ray against traingle when you have very complex geometry?

I don't know if it makes sense performance wise, but independent of this it is just the same 'missing LOD RT support' problem that i have discussed so often.
I think this could work:

Green and red are discrete LODs, with a crack between them. Not sure if a blue skirt polygon is necessary to close this.
The LODs have a displacement map (per vertex) which projects them to lower LOD surface.
If a ray hits the surface, we calculate it's lod fraction between the discrete levels. Can calculate just from distance to camera usually.
We then displace the hitpoint from the fine surface to the coarse lower lod surface times this value, and launch other rays from there.

This gives us continuous lod without a need for cache trashing traversal shaders or complicated progressive mesh. No access to BVH is necessary, and DXR 1.0 is enough.

I have not thought a lot about it yet. What artifacts, etc. But could just work?

If so, can solve LOD and ofc. also insane detailed geometry. We just trace lower LODs and accept hard shadows do not capture fine details. Who cares, we want soft shadows anyways.

 

[MfA]

What do you think now, after seeing UE5 demo?
If single pixel triangles are just as fast as atomic splats, which of the two is less crazy?

They are both just rendering details ... what is the underlying surface representation is the more interesting question to me.

 

[MfA]

Some twitter rumblings from Unreal dev Karis says it's not vector displacement mapping, one of the repliers suggests it might be (variable resolution) geometry images then because he blogged about it previously ... that seems a bit too fine a distinction to me, but could still be true.

PS. can tiled resources have compression per tile? So as to leave some tiles uncompressed.

 

[eloic]

Another great video by Badrobo:

The ground is actually made of individual leaves, no textures. Yeah, some elements appear a bit blurry, but damn, the geometric density is insane!!

 

[JoeJ]

Some twitter rumblings from Unreal dev Karis says it's not vector displacement mapping, one of the repliers suggests it might be (variable resolution) geometry images then because he blogged about it previously ... that seems a bit too fine a distinction to me, but could still be true.

Well, we have to admit there is not much of a difference between vector displacement and GIM, as GIM also is just vector displacement but requiring a bijective parametrization of the model.
IIRC, Karis denied those things, saying he looked it up and spend some time on them, but ended up with something else.

To me it looks like hierarchical mesh reduction, giving a LOD hierarchy of triangle clusters. It would be similar to progressive meshes, but without the morphing to get continuous LOD. With their fine detail, a discrete switch of cluster LOD + TAA is good enough.
Looking at quixel models, they have large UV charts. Maybe they get away with preserving the chart boundaries so the whole LOD hierarchy can use the same texture. Though, such approach can not deal with closing holes and genus reduction without artifacts, so not sure.
Maybe at some point they switch to point hierarchies to get rid of such adjacency problems, but i'm no longer convinced about this. If it's all triangles it might scale better to low power devices with reduced LOD. Or not.

Lots of time has passed but we still don't know anything...

They are both just rendering details ... what is the underlying surface representation is the more interesting question to me.

I'm mostly interested on if they do front to back draw to have occlusion culling built in, and if they do this fine grained (e.g. drawing small tiles to LDS), or coarsely (drawing ranges of depth and updating Z pyramid).
A rendering detail, but more interesting than points vs. triangles.

The ground is actually made of individual leaves, no textures. Yeah, some elements appear a bit blurry, but damn, the geometric density is insane!!

Dreams mostly lacks some close up detail, but in the distance there is much more to see than in UE5 demo. No clear winner for me yet.

 

[eloic]

Geometric microdetail in MK characters (Dreams):

https://twitter.com/x/status/1306020404289470464

 

[SlmDnk]

The future is definitely volumetric, as Atomontage puts it.

I've lately been dreaming of fully voxelized games with physics and ray-traced lighting. Such a game could be based on the movie Tremors, for example. Those "graboids" coming from under the ground to get you...

Anyways, when will NVIDIA bring their first "voxxelerator" to the market?

 

[eloic]

The future is definitely volumetric, as Atomontage puts it.

I've lately been dreaming of fully voxelized games with physics and ray-traced lighting. Such a game could be based on the movie Tremors, for example. Those "graboids" coming from under the ground to get you...

Anyways, when will NVIDIA bring their first "voxxelerator" to the market?

I do hope the future is indeed volumetric and not polygons. The potential shown by Dreams (and other games, at a lesser extent) is amazing.

 

[milk]

I've lately been dreaming of fully voxelized games with physics and ray-traced lighting.

Did you mean you've been dreaming about the game Teardown?

 

[eloic]

Did you mean you've been dreaming about the game Teardown?

I saw that video a few months back. I do like the tech, but in order to get visually pleasing graphics according to reasonable realistic standards (instead of just blocky style), the amount of voxels should be considerably ramped up.

On the other hand, solutions such as the one used in Dreams can achieve prettier/more conventional results, even if it has other limitations, so I'm not so sure that a volumetric approach with standard voxels will get very far. Take a sphere as an example. In polygons, you need a lot to make it seem truly round with no visible polygon edges. In voxels, you'd need A LOT more. In Dreams, we get perfect curves with no visible edges (yes, the flecks may be visible depending on many factors, but still...).

 

[SlmDnk]

Did you mean you've been dreaming about the game Teardown?

Yes and no.

Yes, because I've been following Teardown for quite a while now and it looks like a good step in the right direction. No, because the voxels in it are far too large - or far too few in numbers, as @eloyc put it.

There's also this project: https://twitter.com/ProgrammerLin/media