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

Well, I watched the whole video. I liked it, but I missed seeing more practical material. For instance, they spend a lot of time saying how polygons are not suited for this and that, and how polygons have a lot of drawbacks related to a proper representation of matter and performance-wise (UV data, interactivity and dynamic simulation, etc.), but they don't show how their solution would be better in most of those aspects.

I agree with them in polygons not being the best solution and that a volumetric solution is preferable and tech should move forward in this path, but if you say that UV is the devil because there's a lot of data in play, then please explain how those billions of voxels are not as bad (if not, worse) to manage in real time, and if you say that polygons limit the interactivity and dynamic simulations, please show some examples with your tech that feature a higher level of interactivity and dynamic simulation, not just carving stuff.

However, I feel that this might be a good step forward and I understand the present compromises, so I will watch closely their progress.

I don't think that's voxels. Here's a demo from two years ago basiclaly showing the same thing -

The voxels are dynamically scaled, and you see them rotate on the wheels as well as move around the scene less than one voxel at a time. I think they're using a different representation such as SDF and layering a voxelised visualisation on top.
There's a part in this video where they explain precisely that (scaling and rotation), but they insist on using the term "voxel" and even say that their solution is more advanced that SDFs, point clouds, etc.
 
...if you say that polygons limit the interactivity and dynamic simulations, please show some examples with your tech that feature a higher level of interactivity and dynamic simulation, not just carving stuff.

Dennis Gustafsson has been doing impressive things lately with voxels. The regular structure is beneficial for ray tracing, global illumination, physics simulation, etc.


Regarding Atomontage, there is still clearly a long way to go for the technology. There was a punch line toward the end of the presentation where a 20-second animation loop of a crocodile in empty space was shown to use 22 GB of memory. The demo had no voxel compression or disk streaming, granted, but Atomontage has been talking for years about compressing down to less than 1 bit per voxel, and it is disappointing to have nothing to show here; that sentiment applies to much of the talk, which is mostly aspirational. Voxels are definitely useful, and might well be the future, but with funding and a team they need to start demonstrating the advantages and applications without constant caveats.
 
Last edited:
I'd like to summon @sebbbi into this conversation. :D

Regarding their criticism towards SDFs and bearing in mind what you and the guys at Media Molecule achieved with SDF, do you still think that if there's a volumetric way to go voxels are better than SFDs?
 
The more I see footage from Dreams, the more I think this path should be further followed, researched and improved.

I love how all the details are carved into surfaces, real geometry, nor textures nor normal maps:
upload_2019-1-28_16-9-14.png

upload_2019-1-28_16-18-33.png
upload_2019-1-28_16-17-15.png

Fuzzy/furry surfaces look great, as well:
upload_2019-1-28_16-10-54.png
upload_2019-1-28_16-19-43.png

A more realistic looking approach:
upload_2019-1-28_16-13-8.png
upload_2019-1-28_16-13-48.png

---------------------------------------------------------------

Full video here:
 

Attachments

  • upload_2019-1-28_16-16-11.png
    upload_2019-1-28_16-16-11.png
    659 KB · Views: 13
So i was wrong in the other thread they would have no PBS support. I really like the fuzzy stuff like grass, and the game seems a wonderful toy for creative kids.

The more I see footage from Dreams, the more I think this path should be further followed, researched and improved.

... which was not possible for long time because PC APIs have had no 64bit atomics support. But VK has it finally :) Not sure about DX.
 
upload_2019-2-7_20-18-8.png

The detail in that tree character is fully geometrical. I don't think there's a current game with a polygonal character with the equivalent amount of polygons needed to represent the same level of detail (I mean, without normal maps).
 
That’s what I mean, it’s volume but not geometrical ;)
When I talk about "geometry" I'm just talking about the volumes (polygons also create "volumes", don't they?) that make things in the 3d world, be it polygons, SDFs or whatever. I'm not debating the use of the word per se, I'm just making a point that the amount of "non-2d" detail achieved in Dreams can be very high.

And now, regarding the use of the word itself, "geometry" is used for SDF representations, as well.

EDIT: at any rate, @London-boy , "geometrical" or "geometric"? Now this conversation got me thinking... :-?
 
Last edited:
I think eloyc is right here LondonBoy. Geometry is often use as a synonym for poligonal mesh in the context of 3d rendering, because, traditionally, polygons are the only kind of geometry engines support most of the time. But if we are gonna debate semantics, I believe geometry applies to anything with a position and shape in 3D space (hell, even in 2D space) and thus, dream's volume based primitives are just as much geometry as a triangle based model.
But we all got what you meant.
 
I think 'geometry' is still correct. It is still a discrete approximation of surface, like triangles as well.
It can approximate volumes as well, but only to some degree. And for rendering it shares the same limitations about transparency, because splatting is very similar to rasterization. It's basically the same, just it lacks the bloat of connectivity that triangles have using shared vertices.
This lack is what makes it so interesting: In the same fashion as the REYES method for offline rendering it can fake DOF, motion blur and limited transparency much more easily than large triangles.
It also allows to render diffuse geometry like a bush or grass with high detail and efficiency.
It can fake continuous LOD pretty well.
At distance it becomes a very interesting option to handle LOD, when you no longer can make a good approximation with decimated triangles! (maybe the less obvious killer feature)
It is fully compatible with traditional rasterization. Can be mixed without any hurdle or problem.
It is not compatible with DXR. But a workaround could be to intersect rays with compute, or use a custom intersection shader. Surely a hurdle. For RT volumes are a better fit.
 
But we all got what you meant.
It was a good attempt at a joke and I actually looked up whether it was right or not, updating my school-level understanding of 'geometry'. ;) Dreams isn't Geometry as covered in typical school - all triangles and trig and Cartesian coordinates - but Geometry as a discipline is about describing shapes and volumes with numbers, however that's done. SDF et al are just different forms of geometric representation.
 
Also, I raised the question about geometric/geometrical. I'm not sure if these 2 existing words are equally valid for geometry (I'm not talking about SDFs, specifically).

So, "geometric detail" or "geometrical detail"? Or both?
 
Back
Top