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

[Frenetic Pony]

Works until animation is needed. Or that's at what I understood to be major limitation of nanite. Looks really awesome though. That detail level is insane.

Yeah... the issue isn't... totally getting geometry clusters to work with foliage. It's going to be getting the heirarchical lod to work with complex overlapping geometry instead of relatively continuous stuff, that can then move. You can see how trying it right now has 90% of the mesh just vanish when it gets far enough away. And then getting lumen to work with that as well. It's also why you don't see it for character models yet.

To take it further, it's also why I've been criticizing how the underlying technology choices are made by separate teams doing separate things, that then need a bunch of bodges to fit together. Well, they're a big company the full release date isn't set or anything, they'll be able to get it done, even if it's not pretty and even more bodge like.

 

[chris1515]

I put a timestamp the sequence if it is not working this is at 6:25, this illustrate better the vegetation problem and you see how lumen engine work with a Nanite tree.

 

[trinibwoy]

I’m probably missing something obvious but how does Lumen handle specular reflections, particularly for off screen objects? Isn’t it tracing against a much lower fidelity representation of the scene - the surface cache?

How would you render a mirror using Lumen?

 

[OlegSH]

Isn’t it tracing against a much lower fidelity representation of the scene - the surface cache?

It does cone tracing against local (lod 0) and global SDFs (merged SDF and lower detailed lod), then color is fetched from the Surface cache in hit points (whenever it's possible).
It doesn't shade in hit points, so it's good for something like specular occlusion since it uses a few cones for it (which is very nice), but ugly for more mirror reflections since they require more cones to trace and there is no shading in hit points so mirror reflection will look like this.

How would you render a mirror using Lumen?

It won't work because of the reasons above, also it won't work for any animated geometry (except probably for the animated objects made of the rigid Nanite Parts, such as the Golem in the demo).

 

[Clukos]

I wonder how nanite would work with vertex animated foliage, in that case you probably want to use a "regular" mesh instead. I noticed a bit of ghosting when moving nanite meshes around in the editor.

 

[DegustatoR]

How would you render a mirror using Lumen?

Via SSR fallback only at the moment with all the drawbacks associated with it. This is somewhat similar to what Northlight in Control does without RT reflections active I'd say.

 

[JoeJ]

I've seen UE has supports some planar reflection thing, so for a mirror this should work i guess.

 

[trinibwoy]

I've seen UE has supports some planar reflection thing, so for a mirror this should work i guess.

Yup it’s in the UE4 docs. Works by rendering the scene again from the perspective of the mirror (so not very scalable). Presumably it’s still available in UE5.

 

[DegustatoR]

Planar reflections have loads of limitations and is not a very good solution for anything other than large horizontal reflective surfaces. They also tend to be rather heavy in performance.

 

[JoeJ]

Planar reflections have loads of limitations and is not a very good solution for anything other than large horizontal reflective surfaces. They also tend to be rather heavy in performance.

But it's good for a bathroom mirror, and more importantly: Non euclidean stuff like portals Good game needs portals!

 

[3dcgi]

True. I did not mean they were synomous but had some similarities in common. Based on this read mesh shaders seem to offer more flexibility regarding programmability.

Primitive Shader: AMD's Patent Deep Dive | ResetEra

The patent describes the version of the primitive shader that is fit into the traditional graphics pipeline. There's also a version that launches like a compute shader that has existed from day one. The problem was it didn't quite support the Mesh Shader API. It's close, though when creating meshlets you need to account for the difference.

 

[SlmDnk]

 

[milk]

This question is one of the very critical open ML questions. How is it that humans learn from very few samples and computers require a lot and more samples.

the thing about humans is they always carry a bagage of "implicit samples" for lack of a better term.

If you give a human 5 example rocks and tell them to draw more, wether you like it or not, their whole baggage of what other rocks look like still influences ang biases his output, and it can't be avoided even if they try.

Whereas an AI is a complete blank slate aside from the specified training-set. That's both a flaw and a virtue depending of what you are after. For creating art that looks plausible for humans, that's a flaw, but for finding unexpected/unintuitive but highly-effective solutions to complex problems, that lack of baggage can be its greatest strengths.

 

[manux]

the thing about humans is they always carry a bagage of "implicit samples" for lack of a better term.

If you give a human 5 example rocks and tell them to draw more, wether you like it or not, their whole baggage of what other rocks look like still influences ang biases his output, and it can't be avoided even if they try.

Whereas an AI is a complete blank slate aside from the specified training-set. That's both a flaw and a virtue depending of what you are after. For creating art that looks plausible for humans, that's a flaw, but for finding unexpected/unintuitive but highly-effective solutions to complex problems, that lack of baggage can be its greatest strengths.

I don't have a link handy, but I have heard ML professionals talk about even human babies picking up things with a lot less samples than our current computer algorithms use. Recent studies are showing babies pick up things earlier than we thought before. Some speculate this could have something to do with the structure of brain. Humans are hardwired for things like learning language or interpreting facial expressions. Caveat being there is huge variance between individuals so maybe there is someone who literally cannot learn language or interpret emotions.

 

[milk]

I don't have a link handy, but I have heard ML professionals talk about even human babies picking up things with a lot less samples than our current computer algorithms use. Recent studies are showing babies pick up things earlier than we thought before. Some speculate this could have something to do with the structure of brain. Humans are hardwired for things like learning language or interpreting facial expressions. Caveat being there is huge variance between individuals so maybe there is someone who literally cannot learn language or interpret emotions.

Yes, to use the AI ML analogy, humans are like AI that come with some pre-installed NN models out of the box, hardwired even, built evolutionarily. Some shared with other animals, by the way. The idea of humans psychology as a "blank-slate" at birth, while ideologically convenient, has proven tima and again to be a naive hypothesis.

Erm... anyway, nanite...

 

[Lurkmass]

As great as Nanite or Lumen is shaping up to be, UE5's game framework hasn't fundamentally changed all that much compared to UE4. Their current game framework is very much still single threaded which is by and large down to it's object oriented design. A relatively small amount of actors in a scene can cause you to be bounded by the game logic performance. Unreal Engine's current game framework architecture will continue to impose limitations on the complexity of dynamic objects for games in the future ...

 

[Rootax]

As great as Nanite or Lumen is shaping up to be, UE5's game framework hasn't fundamentally changed all that much compared to UE4. Their current game framework is very much still single threaded which is by and large down to it's object oriented design. A relatively small amount of actors in a scene can cause you to be bounded by the game logic performance. Unreal Engine's current game framework architecture will continue to impose limitations on the complexity of dynamic objects for games in the future ...

Any talk / news about improving performances in UE5 vs UE4 at iso functionality ?

 

[Dampf]

Yeah, aside from Lumen and Nanite and the nicer UI, not much has changed which is a shame. I mean they are still using DX11... just why? Hopefully their DX12 implementation is much improved, it had so many issues in UE4. And no DX12 Ultimate support.

Speaking of AI, I was fully expecting AI accelerated workflow, stuff like ML upscaling textures and ML accelerated physics.

But it's early access so there's a good chance all the features I mentioned will be implemented in the final version.

 

[zed]

I mean they are still using DX11

Well give them a chance DX12 only came out 6 years ago.
Yeah a lot of UE needs to be overhauled, but why worry about that when we can bolt something else new and shiny on top of the broken foundations.

 

[Lurkmass]

Any talk / news about improving performances in UE5 vs UE4 at iso functionality ?

Some developers will implement their own custom framework which is usually based on ECS to bypass Unreal Engine's game framework. Improving the performance of the default game framework would require a change to it's architecture which would involve a substantial rewrite of the engine and Epic has not explicitly revealed that they plan on doing this either in the future. The only thing you can do is to avoid the default game framework as much as possible since it's not very fast at updating the game logic or the actors. I have a feeling that Unreal Engine being this tightly knit to the OO model won't be sustainable for very long and it'll make it harder in the future to support multi-threading on the engine if they don't respond to the issue sooner ...