Lumen GI and Reflections feedback thread

[jblackwell]

@TheKJ

I wasnt specifically aiming at you

I understand, maybe I miss understood your point. I don’t want to point my finger at you either.

I just want to say that I appreciate you all keeping things civil on this channel, genuinely. It’s good to have an environment where people can air their thoughts and frustrations about the technology and still be respected for it.

@TheKJ

I don’t know what the most expensive part of Nanites’ is algorithm, is it the clusters? It is the monitoring of the camera’s positions to provide unnoticeable pop?
Because I would trade in some more visible pop from nanite meshes(As a togglable option) if it meant more performance.

That is a bit of a complex question, and I’d need to pull up ProfileGPU to be sure, but I feel like I should mention that the new forms of nanite don’t always have a flat render time. I got to speak about this with a connection who knows Brian Karis, and he described that Nanite’s programmable raster feature is not guaranteed to cost a flat amount of time. Whenever you have WPO-enabled materials like foliage for example, there is a cost to evaluating that logic that (as I understand) does scale with total triangles rendered. Add to that that foliage is a bad occluder, and you also have overdraw to contend with that’s bottlenecking your performance. The same is true for masked materials, and a handful of other cases. The Electric Dreams demo has absurd nanite costs because it’s filled with extremely high-poly geometry that’s constantly deforming.
Never mind the overdraw that comes as a result of stacks of very thin geometry that can’t use preserve area- again, massive overdraw costs. What I’m essentially saying is that Nanite should be a relatively flat cost in cases of opaque geometry that occludes well-otherwise, it will vary, and the same sort of performance optimization we’re all used to will be called for.

[jblackwell]

@Yaeko

Meh, did anyone see someone from the Lumen Team in this thread since 5.1? (or have they been pushed away?)

I really liked how we could give Feedback and Bugs to them directly…

I saw Krzysztof.N on either here or an ancillary channel talking about lightmass and lumen reflections, but that’s about it.

To be honest, I’m not too surprised. I could definitely see them being on here more in the early days, but this channel is labeled ‘Lumen GI and Reflections feedback thread.’ It started off with the more narrow scope of identifying needed features and fixing bugs, but now it’s become a channel that has everything from UE newbies looking for help, to people complaining/identifying technical frustrations, to occasionally completely non-lumen related discussions about game and light transport/what the next-gen looks like.

I think it’s partly because even though the feedback thread is what they created, people needed a lumen general thread for all the things that weren’t directly dev feedback. I actually wish they’d create a new standing channel dedicated specifically to discrete bug identification and feature requests, so they wouldn’t have to sort through this very long channel history to find what matters to them.

Also, we generally don’t @ them too often, which I partly assume is a courtesy because they are very busy, and there aren’t really any killer feature requests I can think of that they don’t know about (reflection denoising has been on the request list for what…a year?)

[Yaeko]

@jblackwell

identifying needed features and fixing bugs

Hmm, then I will just drop my wishlist and what I know that is still not working:

• per Instance Custom Data (just black in Lumen scene, the last time I checked.)
• an option to not recalculate parts of the level that didnt change (In my game, like 98% of the area lumen deals with is static… so technically it wouldnt need to recalculate that all the time.)
• Distance GI (it was faked in the PS5 trailer with the valley, but we still dont have it in 5.3 afaik.)
• if possible, integrade HWRT reflections into Software Lumen (if there are benefits to performance, which I assume due to dedicated RT hw that otherwise is unused. And of course, the obvious visual improvements for reflections that would come from this.)
• a function to make “holes” into the Distance fields that Lumen uses, since we still have issues with overocclusion and dense plants/bushes. (I posted somewhat of a workaround in this thread, but its annoying to do.)
• move SSGI Data into Lumens permanent Data, so that it isnt regenerated just because I look away, and causing very visible artifacts for large lights. (in general, prevent lumen from “losing” information all the time, even if it costs 100MB VRAM… make it an option. It really doesnt matter if my game needs 3.5 or 3.6 gb vram at 1440p.)

Thats just the things I have in mind right now, that are still unfinished, missing, or “buggy” in 5.2/3.

Also, we generally don’t @ them too often

I really try to avoid @ ing them, since I dont want them to “get annoyed” and never come back. (which is also why I was concerned about the negativity here over the last weeks )

[TheKJ]

I would @ them in a heartbeat (I wish I knew who). Remember they’re paid to make UE5 better. We’re not Engine Devs(or at least most of us), they are.

They already have the extremely complex algorithms for Lumen and Nanite in their head. Simply hearing an idea can spark a new idea in a programmer(I know, I am one, just not very advanced). If we present an idea that sounds to crazy good to implement to us. They might have enough knowledge about the code’s architecture to say "Yeah, I know where I would need to start on that in the code.

Developers like us are their customers and you can never have enough innovation when it comes to ideas and Coding. Because that is what coding is about.

TAA = everything is blurry—TSR 80% runs circles around TAA and gives extra Performance

My response, my TAA hatred had begun (no hate toward you lol)

TAA and temporal solutions aren’t just blurry though, It has immense ghosting/smearing as well, every iteration does, including DLSS.

Except one indie game called bright memory infinite. DLSS looks shockingly good in that game. Idk how but I wish every game with DLSS looked like that?
I thought I could fix TAAU in Tekken 8 with my “special tweaks for clarity”…no….all TAA is bad. That’s what the Tekken 8 demo showed me. In fact it convinced me to not build my game with Temporal filters in mind. I used Tekken 8 as a testing ground since its very fast paced like my game. As for Lumens artifacts. I’m just going to have to find a way to reduce that. Its sharper than TAA blur. FXAA does the trick. Might have to hire someone to put in SMAA.

@ZacD

Native resolution is already dead, consoles wanting to output 4k and DLSS2 proving upscaling can look good enough killed native. The industry is moving on, sorry.

Response to ZacD, Everyone can take a look

With DLSS2 implementations like in Bright memory infinite. I could finally understand the DLSS crave going on but that is the only game where upscaling doesn’t make the final image look like smeary paint.
Consoles are for causals. We can put any tricks we want on those. People buy consoles because they don’t care. PC gamers build PC’s because they notice all the crap in console ports. But PC gaming is turning into crap with all this forced upscaling and TAA.

I intend to break this industry of this idiotic TAA trend. I’m going to build a game that looks so much better than the crap being pumped out today. It going to require money, which pay the required innovations. It’s the modern day devs that have failed and cheated this industry.

Its that weird race for “better graphics” while the hardware isn’t ready for it yet.

Hardware is ready for it, but people aren't being paid to be innovative anymore

30 series and competitor equivalents are ready at their recommended resolutions. We’ve had insane, dynamic graphics without procedural GI or tools like nanite that were optimized in pre-dlss-day games.

I treat the PS5 and Series X like current gen desktop 3060 Ti. That Ti power is just for the “4k” upscaling solution that Sony and Microsoft require from my studio. I’m going to be brutally honest with gamers when it comes to the console port.

What this industry needs is more money being spent on the right things and right research.

[jblackwell]

@Yaeko

per Instance Custom Data (just black in Lumen scene, the last time I checked.)

Could you elaborate a bit on what you mean, please?

@Yaeko

an option to not recalculate parts of the level that didnt change (In my game, like 98% of the area lumen deals with is static… so technically it wouldnt need to recalculate that all the time.)

In many ways, this is what lumen does already, the core of the surface cache is to stably accumulate lighting without needing to trace buckets of new rays every frame. Would you like an option to have it do so more aggressively, and maintain a bigger surface cache over a longer range?

@Yaeko

Distance GI (it was faked in the PS5 trailer with the valley, but we still dont have it in 5.3 afaik.)

That I can directly speak to actually. In Land Of Nanite, the distant GI was achieved via a reflective shadow map derived from the directional light. It wasn’t exactly faked, but it was only a single bounce of indirect light and only from the directional light. I did a fair amount of testing on it, and it…wasn’t that great for what it cost. Are you just talking about SWRT, or do you also have issues with HWRT’s far-field?

@Yaeko

a function to make “holes” into the Distance fields that Lumen uses, since we still have issues with overocclusion and dense plants/bushes.

I was under the impression that this was already a working feature. I’ve seen it work very well in levels that I’ve tested (although I don’t remember how exactly it was enabled). Foliage and thin geometry seemed to let light through to an extent, which avoided the incredible over-darkening otherwise experienced with SWRT.

@Yaeko

move SSGI Data into Lumens permanent Data, so that it isnt regenerated just because I look away, and causing very visible artifacts for large lights.

My question for that might be how to move SSGI data into the surface cache (I believe that’s what you’re referencing)?
Screen traces from lumen exist to cover the mismatch between the Gbuffer and the lumen scene, because SWRT’s caching is inheritly inaccurate compared to the Gbuffer. Caching screen traces would require some way of representing their lighting in world-space, and as Krzysztof.N mentioned in a post a while ago, one of their big limitations to improving lumen ATM is they can’t figure out how to make the surface cache better represent the Gbuffer scene.
In addition, the 2022 SIGGRAPH presentation on lumen mentions that the team tried not to over-cache screen traces because they liked having small-scale lighting be much quicker to respond to change than large scale lighting could be.

These were just my thoughts on the requests you’re looking to have fulfilled for lumen. I’d love to hear your thoughts, and just get to better understand what people are looking to get out of the system.

[Yaeko]

@jblackwell

Could you elaborate a bit on what you mean, please?

Per Instance Custom Data is, when you use a single HISM and change material properties of the instances, per instance. Lumen cant do that, so the HISM is black in the scene, while the same HISM would work in Lumen if it wasnt created with PICD.

Would you like an option to have it do so more aggressively, and maintain a bigger surface cache over a longer range?

It still must be recalculating a lot, otherwise performance would be better.

What I would like is: if there is no change around a probe, then skip it, since in most games the vast majority of probe never changes.

Probably not that easy to do, but could bring a lot of performance.

Are you just talking about SWRT

Only SWRT, since HWRT is kinda dumb (for now - until “RT-Cores” have become the standard even in potato-tier hardware etc.).

Lumen has a bit of an issue, where it “cuts off” in the distance, and doesnt even do something, which kinda makes it very visible (especially in large interiors) and limits the size of stuff you can build, even if we now have infinite shadow draw distance because of VSM.

which avoided the incredible over-darkening otherwise experienced with SWRT.

Still doesnt work for many meshes, but there is a workaround where you can subtract one mesh from the MDFs shape… it would be nice if we could just tell the engine: “make actual physical holes into that Mesh Distance Field” - which is possible, but needs manual intervention right now.

Lumen hates small bushes with dense leafs (if they arent floating in mid air), since it creates a dense MDF and Lumens own capabilities in “dealing with this” arent enough.

See this post for a visual representation: Lumen GI and Reflections feedback thread - #819 by Yaeko

[Miguel1900]

Just point something quickly, in my opinion:

I think Unreal 5 is an engine made for the next 10 years (I read it somewhere), so I think it’s not thought for the current gen. A different thing if that they are trying to make it ‘compatible’, but not ‘native’ for this gen. For example, a 3060 it’s a quite weak GPU for this technology and its state (I wish it could be better, running GI at 540 FPS, of course, but no). And it’s not comparable with consoles, as they are closed system, with much higher optimizations.

So, for me: the current tech is almost UE4 (or UE5, but using UE4 workflows). UE5 with Lumen is something like prototipying tech. Developers can start developing with it, but to release games after some time, when the tech if affordable for gamers, not only for devs.

[TheKJ]

a 3060 it’s a quite weak GPU

No, it’s not weak. Not for 1080p gaming. On optimized games, it runs well above 120FPS. That’s without “nextgen” features.

The 3080 performs the same as a 3060 playing 1080p when the 3080 is playing at 4k. It’s scales. It’s as simple as that.
This is the same case for AMD and Intel GPU competitors.
Global GI, optimized meshes/shadows will get you 60fps on these systems native resolution. It just takes care which doesn’t exist in most modern studios.

And it’s not comparable with consoles

Um? The 3060 has enough power that Sony and Microsoft based their “Next Gen” console’s computing power(teraflops) on a 3060. (Exclude series s)

Consoles like the PS5 /Series X are both RNDA2. They do get optimized pipelines(so do RNDA2 gnus on pc) but they also have different tech inside than a 3060. Consoles have 16bit instructions sets instead of AI cores which accelerates things like TSR’s calculations.
Consoles are just current gen 1080p GPUs.

Ehh... Little comment

Current gen is still defined by the 30 series and competitors since 40 series above 4070 are a joke to target consumers.

Since console’s get the optimized pipelines, it shrinks the millisecond budget below 16.33 when playing at 1080p. That extra ms headroom is for whatever upscaling method will bring that Native 1080p to “4k” such as TSR or FSR on console(Hell I think XESS can go on consoles). We really shouldn’t advertise 4k gaming on consoles but this is the rules Sony and Mircosofties have made to make “4k” or even “8K”(LMFAO) gaming a selling point.

as they are closed system, with much higher optimizations.— —
when the tech if affordable for gamers— —

I’m still having an extremely hard time understanding what you’re saying? I believe it is affordable if devs would just being realistic about these technologies? Nanite+VSM’s or Lumen. Pick your ms poision.

[Yaeko]

Meh, I am fine withy my ~40ish fps on a 3060, with TSR, Lumen, Nanite and VSMs @ 1080p. (If nvidia didnt “nvidia” the 4060, it could run it at 60… but we all know how that went, 3060 2.0)

Those “meh” gpus (60 for nvidia) never were good at running the newest stuff at max settings at their target resolution, for decades now.

Its true though, that many games waste performance for nonsense.

[TheKJ]

@Yaeko On a 3060 1080, don’t use TSR as for native 1080(as said by the inventor, it’s not meant for native resolution). USe FXAA or TAA( I personally like fxaa for reasons everyone should know by now) Keep VSM’s on medium. Lumen on High(as said by Epic targets 60fpsps) And every single post process on low or off. You should get 51fps on a busy scene, 60fps on a simple scene.

Max settings target 30fps.

[Yaeko]

you really need to start showing your scenes, your numbers are impossible in anything that looks good or is reasonably complex. (aka: an actual game, more complex than what Fortnite is.)

A 3060 can not do this, not even at 80% scaling. (which is what I am running, hence why TSR is enabled…)

Maaaaybe, if you have zero translucency etc. in the scene, no additional lights and so on - then, maybe. (and lumen cut down even further than just “high” settings.)

But thats something I consider “unreasonable” limitations to work with.

On a 3060 1080, don’t use TSR as for native 1080(as said by the inventor, it’s not meant for native resolution

It was completely fine, until epic messed up its performance in 5.2…

It was almost FREE (EDIT: Just checked, I pay 2 fps for the 5.0 TSR at 1440p - negligable, for the result.) to enable in 5.0, nowadays it costs actual ms for some reasons.
Idk if they stripped the RDNA2 features out of the engine, whatever they did, it did hurt performance big time. (it costs me 10 fps in 5.2 x_X… which is why I simply put it at 80% scaling and got back 20 fps, having gained 10 in total.)

[TheKJ]

@Yaeko

you really need to start showing your scenes, your numbers are impossible in anything that looks good or is reasonably complex. (aka: an actual game, more complex than what Fortnite is.

@Yaeko
I can’t, but when my project goes public, everything with be documented to finally get this performance argument over with. No secrets kept, I just want gaming to get better already(Graphics, Performance and Non-TAA wise). Go and do what I said in City Sample if you want something for complex. load up half a mile by half a mile.

I have City Sample 5.1 running near 1080p(windows bar), at 53-55fps on a mobile 6gb vram 3060(Mind you is 20 percent slower than a desktop 3060, 50+20%=60fps) in 5.1
No postprocess, FXAA, SWRT Lumen and reflections(scaled down but still looks photorealistic) with no flickering and scaled lumen settings scaled down a lot.

Still not good enough. I want bloom, I want Motion blur but NO. I’m stuck with 60fps bare-minimum because of Nanite+VSMs wasting performance.

DEAR EPIC DEV TEAM...

THIS IS WHY WE NEED MICRO-OPTMIZATIONS on Nanite, VSM, and Lumen. So I can add Motion Blur, etc for a 60fps UE5 workflow on current gen! STOP working on new features!
And give me performance tricks and enablers or your new Nanite-VSM-Lumen workflow is worthless to us and gamers.

5.2 has Nanite problems. I can only get 49fps with same settings I posted in the 2nd paragraph I wrote in this post. Looks like I won’t go to 5.2.

Note when played in Editor, City Samples VSM ms timings go from 2ms to 3.5 even with no AI. Probably because of the extra HLOD shadows. Idk. Then I’m back to to stupid 47fps. (All timing I gave were in editor btw)

It was almost FREE

You pay for that rotation & foilage smearing TSR hasn’t fixed yet.
The dude who made TSR told me the RNDA2 pipline is related to drivers. The Engine source code always checks for the 16bit instructions to enable the pipeline.

EDIT: Fortnite 5.1 is pretty incredible looking. You literarily have every Biome imaginable including a city. I test it monthly. It has unacceptable performance for such a stylized (but populated at hell) environment. I get the same FPS in City Sample Mile by a Mile(remember no play in editor) and my attempt to optimize FN (51fps native 1080p, Lumen, no postprocess, FXAA on a mobile 3060).

This post shows why I have decided not to use Nanite for everything. Only per instances that show better performance over dithered (not temporally) LODS.

[Miguel1900]

Well… I still don’t agree with you. 60 series are low end gaming GPUs:

xx50, xx50ti, xx60, xx60ti, xx70, xx70ti, xx80, xx90, xx90ti… it’s the third card from the bottom (not even in the middle, which is the xx70), and I wouldn’t consider the 50 series as gaming cards. The minimum decent gaming cards are the 60ti’s (the gap is huge versus the 60 series).

Not sure how you are comparing the 3060 versus the 3080 (!), but here you can have a global idea about all GPU performances:

If you want to get 60FPS in a 3060 with Lumen enabled, I think you can wait eternally. When I had my 60 series, I was able to play quite well with high and ultra presets in games at 30-60 FPS, but I needed to low graphics in super demanding games. Unreal + Lumen can be considered a super demanding game right now; don’t pretend to execute it at its full capacity with a low end GPU, even if they are trying to make it very optimized and scalable.

This is only my opinion/advice, even if I think too that Lumen still needs to improve quite a bit.

Good luck.

[TheKJ]

@Miguel1900

Not sure how you are comparing the 3060 versus the 3080 (!)

You’re kidding right? I know the 3080 is more powerfull. Everyone Look below

Click Here for Full reply<-

@TheKJ

The 3080 performs the same as a 3060 playing 1080p when the 3080 is playing at 4k.

Take a game. It runs 73fps on a 3060 at native 1080p You play that same game at 1080p on a 3080, It performs 120+fps.

If you take that same game and change one setting(the resolution ) to 4k. It will perform at 73fps on the 3080.

@TheKJ

It’s scales. It’s as simple as that.

I know, I have done the game focused research on it. It’s what I live for.
This is the current gen scale for GPU iteration to resolution.

@TheKJ

Make no mistake, I believe the 30 series and any AMD/Intel equivalent competitors are the current gen for gaming. 1080p for 3060, 1440 for 3070, 4k on 3080, Plugin competitor’s vendors, etc. I think 30fps is unacceptable trash in this day and age and 60fps is a gold standard for a reason.

@Miguel1900

If you want to get 60FPS in a 3060 with Lumen enabled, I think you can wait eternally

I just explained how in this thread. I got 51-55fps on a mobile 3060 at near 1080p(like 98% 1080p). So not even target hardware(Mobile players will have to sacrifice?).
We need more micro-optimizations in the new features to have ms head room for postprocess like motion blur, bloom etc. We need more performance/scalability when it comes to Nanite and VSMs if they want to replace static LOD meshes. And less features relying on Temporal AA methods.

[Miguel1900]

Ok, I see your 3060-3080 comparison now. But what do you want to point with it? 70 FPS for native 4k seems good for a 3080, isn’t it?

If the 3060 is a mobile version, then it’s a super weak GPU, not only weak. And what TDP (power limit) does it have? It’s like thinking about running Crysis in a laptop during those days. A ‘low’ laptop may not be very recommendable to run our ‘nowadays Crysis’ (UE5 with Lumen).

Probably they can’t do miracles to gain more ms in those kind of cards. The only thing may be scaling methods.

[TheKJ]

what do you want to point with it? 70 FPS for native 4k seems good for a 3080, isn’t it?

Yeah, it’s above the 60fps standard. That’s not the case with a Games or projects using Lumen, Nanite and VSM’s, and or TSR at native resolution. I chose 73fps to make my example clearer.

From my testing, you have to choose 2 of the those features to run on current gen(30 series).
And balancing between Nanite and LOD meshes seems like the way to get out.

If the 3060 is a mobile version

I explained how much slower it is. 51fps on 3060 6gb is 20% slower then the desktop(50+20%=60fps on a desktop version).

@TheKJ

I have City Sample 5.1 running near 1080p(windows bar), at 53-55fps on a mobile 6gb vram 3060(Mind you is 20 percent slower than a desktop 3060, 50+20%=60fps) in 5.1
No postprocess, FXAA, SWRT Lumen and reflections(scaled down but still looks photorealistic) with no flickering and scaled lumen settings scaled down a lot.

Probably they can’t do miracles to gain more ms in those kind of cards. The only thing may be scaling methods.

You’re just enabling the UE5 devs to think current performance is acceptable. It’s not. We need a good 70fps from the Nanite, Lumen, VSMs in FN 5.1 first from current gen to resolution before we can start adding motion blur and important post processing.

That kind of un-innovative, pessimistic talk is ruining games.
That is why I mentioned Intel. They are the only ones thinking differently(beyond hardware).

If the 3060 is a mobile version, then it’s a super weak GPU, not only weak.

It’s not that weak for what it can do with UE5 as I explained. Also Intel is set on bringing path tracing to integrated graphics by micro optimizing actual code. That’s the kind of innovation the UE5 devs are lacking atm(and it’s because no one gives an eff but I do and gamers do). The UE5 devs are just slapping on TSR instead.

[Yaeko]

just an addition to your graph:

SOurce: TechPowerUp

using 1080p for such a comparison is not fair to the high end cards, since they will often be bottlenecked by the CPU, which in some games can even happen at 4k, but that would be unfair for the 3060 since it heavily suffers there.

The 4090 is on average 3.27x faster than a 3060(4060)… and then consider the 4090 not even being the fastest card nvidia could have made. (Now imagine what happens, if we get a (low) performance improvement of 30% for RTX 5000… or a high one of 50-60%, or… to say the unthinkable: 80% or so …

3060 and 4060 will age like old milk, and it already started.

So, its actually worse than you showed

[TheKJ]

@Yaeko

3060 and 4060 will age like old milk, and it already started.

Not if you stick to 1080p. People buy the corresponding GPU for what resolution they can afford to play at. People who buy a 3060 aren’t trying to play at 4k.

There’s a thing said about pc players that they already paying more than console so they must have more money. But a gaming PC has many other uses worth investing in, even a a standard 1080 rig.

People will buy a 3080 or 4070(for 3080 performance but half the price and 2GB more VRAM) to play games at 4k.

Producers or extreme ultra modding enthusiast will buy RTX 90s iterations.

[Yaeko]

you seem to missunderstand, that game-developers dont care what people buy. (They dont care that most people hafve 8GB cards, or less - they just yeet it and now those cards are running into issues unless you turn the settings down.)

Afaik, even nvidia themselves admitted, that they screwed up.

Not if you stick to 1080p. // People who buy a 3060 aren’t trying to play at 4k.

That card already has issues at 1080p in half the new games, and in some even with lowered settings (if you consider 60 fps the target.)

And then nvidia made a 2nd 3060, called it 4060… that just as slow as a 3060

EDIT: Did you look at some of the upcoming games, and how they literally scream “I am going to melt your GPU”? The outdoor parts of “Star Wars Outlaws” for example… I doubt even my 6900XT can still get 60 fps there, all maxed out - and that thing is ~2x a 3060. (And at that point I didnt even take into account, that - if AC Valhalla is any baseline - the game will performa a lot better on AMD than nvidia…)

You may disagree with me… but you will remember my words in 2024/2025.

[Arkiras]

You guys should fight this out in PMs, this is not useful nor does it belong here