Unreal Engine 5, [UE5 Developer Availability 2022-04-05]

@Shifty Geezer
Also i never said that you don't need good lighning. My argmuent is that people are to much fixed ond lighning and forget the geometry and if you have the choice you should always go geometry frist.

I have brought a lot of examples where you have Ray Tracing, Screen Space lightning or Lumen lighning and you see barly a difference. If you see the last years game Test, the most disapoitment was always ray tracing. Yea it looks a little bit nice but it was not a deal breaker for games like Cyberpunkt an other games.

Also your Cornell Box, it is a room which have one lightning source and simple geometry and of course you will see the difference between lighning because the room was designed to bring out the difference of the varius lighning technics. And even with Ray Tracing the Cornell Box is not looking realistic because real world opjects have small scetters, ruff surfaces, screctes and so on. I doubt that some normal user can see the difference between a good lightning and normal lighting. But the people will easyly see the difference between a medium and high poly.

The same is with Toy Story. Toy Story have Ray Traced Scences but it looks like in a movie, because the surface are to smoot too gald bo be realistic.

@OleghSh Why it is not possoible to compress also color data stored at vertex level? The information ant the repeating of the information is not chaning and it should the same compression rates possible for both?

@Andrew Lauritzen I agree for detailed user which we are it is obius. But if you show your picture to normal users. Who will see difference? What do you think will users attract more? The detailed statue in the left or that the shadows looks a little bit more realistic?
 
Graphics are produced via a multitude of techniques and properties, and to pick geometry out and put that ahead of everything else is very misguided IMO.
I agree, although I will say based on a couple years of experience with Nanite now it's like any "upgrade": at first you are like "sure, that's looks a bit better" but once you get used to it, when you go back to the old thing it seems a lot worse. It's increasingly harder for me to look at games and not just immediately notice the large triangles and jagged-edged curves now. It definitely depends on the content how obvious this is, but even in relatively urban settings there's always areas or objects that benefit from significantly more geometry detail.

As an aside though, the "HW/accelerated/whatever vs SW raytracing" argument is a bit silly. If anything, the discussion is more triangle RT vs tracing [other things], but even then we've already moved past that into a big sea of hybrid techniques. Pure triangle RT is no more appropriate than pure rasterization was IMO, as triangles are not an ideal representation for sampling large, diffuse areas or volumes.

Neither is "triangle RT vs Nanite" a particularly interesting long-term discussion as we clearly need "both". For DXR-style RT to be practical for stuff like shadows and mirror reflections we need to be able to build BVH's faster, or efficiently traverse non-triangle leaf nodes (intersection shaders in their current form are far too slow). We probably need better control of the traversal process as well. The current form is a good proof of concept and has been used reasonably in some games, but the limitations are still very apparent (outside of static tech demos).

Furthermore almost all of the things that are a problem for Nanite are even more of a problem for triangle RT, foliage being a primary example. I promise you the long term solution for foliage blowing in the wind with RT is not going to be intersection and any-hit shaders. Basically every game has to make massive sacrifices to represent *last-gen* levels of foliage into any sort of DXR data structure, with several choosing to just skip foliage entirely. The problems with foliage are fairly fundamental and probably deserve a rethink of how we render foliage in the first place. We as an industry will find solutions and compromises of course, but it always strikes me as a bit funny when people use foliage as an example of something that Nanite doesn't currently address but seemingly ignore the even larger issues with foliage and triangle RT... and insert skinning and whatever else in there as well.
 
Last edited:
@Andrew Lauritzen I agree for detailed user which we are it is obius. But if you show your picture to normal users. Who will see difference? What do you think will users attract more? The detailed statue in the left or that the shadows looks a little bit more realistic?
My point is that I don't think that's a good question because we perceive the statue to have detail in a large part *because* of the lighting on it. As a counter-point if I removed shadows entirely from a scene, even if you keep the BRDF I'm pretty sure 100% of users would prefer anything lower-detail geometry but with shadows. We need both more detailed geometry, and more detailed lighting to go along with it to truly bring visible improvements.

More detailed lighting without more detailed geometry gets you 90's raytracing demos and more "bumpy shiny", while more detailed geometry without more detailed lighting gets you more expensive POM. (These are obviously both reductive but I hope you see my high level point.)
 
I also don't argue about that. But the question is do you need realy ray tracing shadows or are old shadow technics ok for this specific game?

My point is that medium - good lighning and high poly mesh is better than medium poly mesh with extrem good lightning. Like you said, if you have seen nanite and look back on older games you realy ask why it has ray tracing but not better geometry? Cyperpunk is one of this games.

Like the foilage. Better bring foilage inside the game, instead of ray tracing when both is not possible. I think there would be a lighning apporximation which will work, maybe it looks not so good like ray tracing but you get foilage inside. Like the shadows VSM and CSM
 
Last edited:
Why it is not possoible to compress also color data stored at vertex level? The information ant the repeating of the information is not chaning and it should the same compression rates possible for both?
Because indexed geometry be it stipes or triangles can be arbitrary in size, topology, etc. It would not be as convinient and efficient as working with power of 2 large and uniform textures with automatic MIPs and great support in HW, so compressing textures is way more convenient for both SW and especially HW than compressing tons of small arbitrarily sized meshlets or indexed geometry.
As a sign of evidence, progress on geometry and per vertex attributes compression (which usually comes down to interpolating some attributes across large patches of geometry) is quite moderate and slow in comparison with texture compression.
 
My point is that medium - good lighning and high poly mesh is better than medium poly mesh with extrem good lightning.

Do you have examples in mind of those two scenarios? During normal gameplay I definitely notice low fidelity geometry issues more easily than inaccurate lighting but both stand out at times. Low resolution shadows are probably the most noticeable lighting issue for me. Low poly or low density geometry is much worse though. E.g. garbage piles simulated by textures.

Like you said, if you have seen nanite and look back on older games you realy ask why it has ray tracing but not better geometry? Cyperpunk is one of this games.

Same thing applies to lighting. After a few years of GI what you consider acceptable dynamic lighting now will look like garbage. Of course this doesn’t apply to static baked lighting which already looks very good today for obvious reasons.
 
@Shifty Geezer
Also i never said that you don't need good lighning. My argmuent is that people are to much fixed ond lighning and forget the geometry and if you have the choice you should always go geometry frist.
Okay, that's not really what you said. You started with - "It is not llightning. It is the geometry which interacts with the light. You she the Bolts and the fine shadow of these bolts? This makes the Sceen realistic. You don't need ray tracing to look good. You need a medium light source and good gemotry and you will have astonishing effects." I appreciate you aren't a native English speaker so perhaps phrased yourself poorly but the arguments since have been about the need for realistic lighting over geometry, and not the need for RT-level lighting versus simpler models which is now seemingly your true point.

However, it's still the lighting that defines realism first and foremost. What you're now saying is it doesn't need RT as long as a suitable approximation lighting system is used. That's something we can agree on, but it's still the lighting being realistic that makes the whole graphics realistic. We've been able to produce really great stuff through baked lighting; Naughty Dog have some amazing results in game with baked lighting. But it's limited in other ways and that's one of the things RTRT provides, more robust high-quality lighting where the fakes can break in a fair number of ways or be otherwise limited.
 
I agree, although I will say based on a couple years of experience with Nanite now it's like any "upgrade": at first you are like "sure, that's looks a bit better" but once you get used to when you go back to the old thing it seems a lot worse. It's increasingly harder for me to look at games and not just immediately notice the large triangles and jagged-edged curves now. It definitely depends on the content how obvious this is, but even in relatively urban settings there's always areas or objects that benefit from significantly more geometry detail.

As an aside though, the "HW/accelerated/whatever vs SW raytracing" argument is a bit silly. If anything, the discussion is more triangle RT vs tracing [other things], but even then we've already moved past that into a big sea of hybrid techniques. Pure triangle RT is no more appropriate than pure rasterization was IMO, as triangles are not an ideal representation for sampling large, diffuse areas or volumes.

Neither is "triangle RT vs Nanite" a particularly interesting long-term discussion as we clearly need "both". For DXR-style RT to be practical for stuff like shadows and mirror reflections we need to be able to build BVH's faster, or efficiently traverse non-triangle leaf nodes (intersection shaders in their current form are far too slow). We probably need better control of the traversal process as well. The current form is a good proof of concept and has been used reasonably in some games, but the limitations are still very apparent (outside of static tech demos).

Furthermore almost all of the things that are a problem for Nanite are even more of a problem for triangle RT, foliage being a primary example. I promise you the long term solution for foliage blowing in the wind with RT is not going to be intersection and any-hit shaders. Basically every game has to make massive sacrifices to represent *last-gen* levels of foliage into any sort of DXR data structure, with several choosing to just skip foliage entirely. The problems with foliage are fairly fundamental and probably deserve a rethink of how we render foliage in the first place. We as an industry will find solutions and compromises of course, but it always strikes me as a bit funny when people use foliage as an example of something that Nanite doesn't currently address but seemingly ignore the even larger issues with foliage and triangle RT... and insert skinning and whatever else in there as well.

Explains why vendors are chasing RT acceleration so hard (intel/nv and soon amd probably). Very intresting.
 
Neither is "triangle RT vs Nanite" a particularly interesting long-term discussion as we clearly need "both". For DXR-style RT to be practical for stuff like shadows and mirror reflections we need to be able to build BVH's faster, or efficiently traverse non-triangle leaf nodes (intersection shaders in their current form are far too slow). We probably need better control of the traversal process as well. The current form is a good proof of concept and has been used reasonably in some games, but the limitations are still very apparent (outside of static tech demos).

Furthermore almost all of the things that are a problem for Nanite are even more of a problem for triangle RT, foliage being a primary example. I promise you the long term solution for foliage blowing in the wind with RT is not going to be intersection and any-hit shaders. Basically every game has to make massive sacrifices to represent *last-gen* levels of foliage into any sort of DXR data structure, with several choosing to just skip foliage entirely. The problems with foliage are fairly fundamental and probably deserve a rethink of how we render foliage in the first place. We as an industry will find solutions and compromises of course, but it always strikes me as a bit funny when people use foliage as an example of something that Nanite doesn't currently address but seemingly ignore the even larger issues with foliage and triangle RT... and insert skinning and whatever else in there as well.

Points splatting certainly seems like it's being given a second look. Obviously Dreams works with dynamic objects all the way back on a PS4. But more recently I've seen it used for distant foliage impostors in an upcoming title, mipmapped even, and it's looking pretty seamless. Just as relevantly is Unity's new hair system which seems to be using point splatting and looks to be running both very fast and do better for thin geo representation than UE5's system.

Given that you can trace it just as fast as primary vision that seems like something to look at. I wonder if you can get it to represent arbitrary geometry as well, sharp corners and all.
 
I don't think I could classify lighting or geometry as more important than the other. I do think lighting is much further on the photorealism curve than geometry though Nanite goes a long way to equalize it. Baked lighting has also tipped the scales quite heavily though. IMO current game design doesn't provide a very compelling argument to abandon it.
 
Last edited:
Points splatting certainly seems like it's being given a second look.
Yeah, that or just something volumetric is clearly needed in the distance. Artists already spend a bunch of time trying to "undo"/work-around the solid object shading of leaves: messing with normals, shadow falloffs, etc. to try and soften it all out to fake partial occlusion and transmission. Treating them like participating media at some distance likely makes a lot of sense. The trick, as-always, will be that if you can get up very close to them we do need to transition back into some representation that can handle individual leaves and so on, at least for primary visibility (shadows TBD...), and do that without obvious visual pop.

I wonder if you can get it to represent arbitrary geometry as well, sharp corners and all.
I believe this was explored a bunch for Nanite in addition to voxels and other things (see Brian's talk below). There's a bunch of problems and inefficiencies with those representations when it comes to solid geometry, but that of course doesn't mean they don't make sense for stuff that is more "on the line" like foliage.

 
Someone posted this on Twitter.

Made from just smashing some market place megascans together in UE5.

FTBh-i-Ja-QAEYDth.jpg


My brain can't cope.......
 
Back
Top