• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

Fake frames: An uncomfortable truth about DLSS 5

LukeTbk said:
yes 30 fps lived much longer on consoles (and still a thing, GTA 6 will that mode and tens of millions will find it playable), controller and sitting distance/screen also could help, you watched 24fps the other day in the same setup.

But would we play a full year with 1% low at 45 fps on a monitor that has good VRR at that frame rate, we would probably find it acceptable again like we would have in the 90s/earlys 00s, it is being used to much higher that make trying at lower feel strange at first, that would be my guess.


if you try it you will see why it is not out yet, under very high fps it has issues at least at the state the dlls leaked were back then (and why they are going for the under 15 ms high fps crowd for this):


View: https://www.youtube.com/watch?v=tENebDPWgO4
it was on github, i really do not have the I mind latency to judge the result when I did try it (same for previous attempt in the past), at high fps it work well enough, at 30 issh you see artifact.

One way that could make it work would be to render a bit of an higher resolution than what we see so when you move it has something to build the estimated image, but obviously you rapidly come in the may has have higher native frame rate instead
Click to expand...


Overall I agree with this post, except the following bit which seemingly fails to understand the concept:

LukeTbk said:
...you watched 24fps the other day in the same setup.
Click to expand...

"watching" is not the issue.

If I look over someones shoulder and they are gaming at 30fps, it looks fine to me. Sure, it could be smoother, but I'm not really bothered.

It's the moment you grab the mouse in your own hand and use it to look around the scene that the input lag instantly becomes notable and gross, and I want nothing to do with it.

I think controllers tend to already be so bad that the input lag isn't as noticeable, which is why consoles get away with less. The PC is the victim of its own very precise inputs that make us used to very precise and accurate controls, and make it very noticeable and off-putting when they are even slightly lagged.

Input lag is of course not just a matter of frame rate, but also a stack up of everything along the signal chain from your inputs, through the USB bus, through the CPU and game engine, and the render pipeline, and even the monitor being used.

At some point the total of all of those delays exceeds a certain threshold where it becomes very noticeable and undesirable.

I don't know what that total number is, but in most setups I have used inputs feel natural and mostly instantaneous when the framerate portion of the input lag is at roughtly 60fps or higher. So, since frame time is the inverse of frame rate, I guess we are talking 1/60s or ~16.67ms.

But as I was saying, this is really just a portion of the total input lag:

Mouse Lag + USB Bus Lag + CPU/Game Engine lag + rendering lag + frame-rate cut-off induced lag (16.67ms) + monitor lag = total lag. The total lag is presumably the figure we should be interested in.

Some systems will feel more instantaneous than others, even at the same frame rate, presumably because something about them lowers the lag at other places in this chain.

I have - in the past - noticed that some titles feel more responsive than others at the same frame rate, even on the same hardware. I presume this has been due to post processing of rendered frames, increasing the render pipeline lag, but I am not entirely sure.

So, there are a lot of complex variables, meaning that the 60fps thing is by no means a universal cutoff for me, but I find that on most systems that's about where things start feeling good. On an otherwise good system with low lag (good monitor, fast CPU, good input hardware, well designed USB controller, etc.) I have occasionally been happy with 50fps (20ms)

I notably ran many of the newest titles at 50hz with vsync on back in the days before I had GSync/FreeSync/VRR. For instance, back when I had the Samsung JS9000 4k TV as a monitor, I discovered that it had all of the refresh rates for both US (24/30/60hz) and Europe (24/25/50hz) so I was able to set it to 50hz and vsync it, because if I tried 60hz, I just couldn't get the hardware to successfully render that fast, and vsync would drop things down to a flat 30fps, which was intolerable.

I suspect the reason I found 50fps to be acceptable was because the Samsung JS9000 had a really good low latency gaming mode, at least for the time. That might have helped make up for the additional 3.33ms loss in latency due to the frame rate.

Modern TV's and monitors tend to have even better input lag (at least the good ones) meaning that maybe you can still lose some framerate and the cumulative input lag won't feel as bad as it otherwise would, but I still doubt I'd be happy much lower than 50fps. even on a very low latency system. The 50fps feeling acceptable was a clear exception for me, but it was a compromise, and a compromise I was willing to make to get certain really tough to drive 2015-2016 titles to run on my hardware.

So, maybe I could tolerate lower framerate today on monitors that have much lower input lag than in the past. Lowering the input lag on the monitor may just open up some flexibility to allow input lag resultant from frame cutoffs to rise without exceeding the total input lag figure that causes things to feel really problematic.

Still, when I tested S.T.A.L.K.E.R 2 at launch, it didn't feel that way. At 4k Ultra settings without upscaling or frame gen, in the opening scenes I was GPU limited on my RTX 4090 at ~51fps, and it felt really bad. If I enabled DLSS upscaling (the level didn't matter) I was now no longer GPU limited, but rather CPU limited (on my Threadripper 3960x) at ~53fps.

For shits and giggles I tried enabling frame gen (DLSS3 frame gen at the time, this was right before DLSS4 frame gen became available). When I did, the reported frame rate was about 100fps, and things looked pretty smooth and fluent, but mouse input still didn't feel right, and felt very laggy.

I will concede that it felt almost playable to me, but for a title that I had waited for for 15 years and was excited about, I wanted my first playthrough to be perfect, so almost was not going to cut it. I decided to wait.
 
Zarathustra[H] said:
"watching" is not the issue.
Click to expand...
like you say, framerate and latency while one helping the other are quite different.

You could have good latency at 30 fps, 16.66ms average the theorical best, which is much better than most games has, it is motion smoothness on any significant camera change speed that will never be good at those frame rate.

that why frame gen + reflex is sometime similar or even better than native without both.
 
DukenukemX said:
Only if the frames are real. The whole point of a higher frame rate is to reduce input lag. Fake frames make real input lag.
Click to expand...

This is literally why they recommend a minimum

Once you're past a certain threshold it's gigantically smoother looking and you can't feel shit difference input wise.
 
LukeTbk said:
like you say, framerate and latency while one helping the other are quite different.

You could have good latency at 30 fps, 16.66ms average the theorical best, which is much better than most games has, it is motion smoothness on any significant camera change speed that will never be good at those frame rate.

that why frame gen + reflex is sometime similar or even better than native without both.
Click to expand...
33.33ms :) for 30fps. 60fps is 16.66ms apart.
 
Zarathustra[H] said:
I would argue that 35fps native, and 60fps with DLSS based off of 35FPS native are equally unplayable.
Click to expand...
I guess people weren't playing any games back then?
In hindsight, I was on a CRT, so that may have helped. I played C&C Renegade hardcore for years and it ran at a fairly consistent ~40fps. Before that you had Quake 2 CTF, Duke3D, Doom, etc. that were often slideshows, depending on hardware. As someone said, this was all early 2000's, 90's.
WilyKit said:
Evolution. Back then, the human eye couldn't see past 60fps
Click to expand...
Lol! That's what my friend used to say (brain can't process more than 60-90fps). Maybe my brain was interpolating extra frames.

Why 40fps vs 30fps makes a big difference:

View: https://youtu.be/5QEEpc2ejmY?si=mq-De_LpDbqn6ftE&t=179
 
Last edited:
sleepeeg3 said:
In hindsight, I was on a CRT, so that may have helped. I played C&C Renegade hardcore for years and it ran at a fairly consistent 35-40fps. Before that you had Quake 2 CTF, Duke3D, Doom, etc. that were often slideshows, depending on hardware. As someone said,
Click to expand...
Some if it may have been the CRT and its greater motion clarity for low FPS but a lot of it is just what you get used to. Once you get used to something nicer, it is hard to go back. For years I gamed and worked on 60Hz LCDs and I was happy. However I've gotten spoiled with higher framerates. At work, my mouse cursor is noticeably rough because I just have a 60Hz monitor there, compared to the 240Hz I'm used to on my laptop and desktop at home. I never would have said 60fps was insufficient... until I got to experience better.

Heck, resolution is similar. I was pleased as hell when I got a 1920x1200 monitor and got to game at full HD. These days though, I've really gotten accustomed to 4k and going back to 1080 just looks low rez.

However, as I've said before including in this thread, it is diminishing returns. I notice 60 to 120 a lot more than 120 to 240.
 
Try playing the game with fake frames.. The input latency sucks... Just looking wont tell you shit.
 
LukeTbk said:
you watched 24fps the other day in the same setup
Click to expand...
Watching movies is quite different, I assume that's what you're referencing with 24 fps. Movies work and look smooth at 24 fps because of motion blur, not the fake one in games, but the actual phenomenon. The picture is recorded with a shutter that is open for the entire duration of the frame so all the motion information is present in the final 24 fps film. If you pause a frame it looks blurry but while watching it at normal speed your eyes don't care. Now if the movie is recorded with a quick shutter speed like 1ms, it would be just as jarring as playing a game at 24 fps. But when pausing, the image would look crisp. Some older camcorders even had a "sports mode" which only had the shutter open for a very small time for each frame, not for the entire 1/24s.
 
Dion said:
Try playing the game with fake frames.. The input latency sucks... Just looking wont tell you shit.
Click to expand...
the title here is misleading a lot of people, the fake frame here are dlss/fsr upscaled not full interpolated frame generation.

M76 said:
. Movies work and look smooth at 24 fps because of motion blur,
Click to expand...
If camera movement are slow enough, there so much blur can do on fast pan.
 
sharknice said:
IMO the future of frame gen is the game engine making a Quake 3 looking image at an extremely high frame rate and AI image generation making it look photo realistic or stylized in whatever way the developer chooses. The game engine could even feed the AI generation with double the frames it is outputting.
Click to expand...
DLSS 5 seem already a clear first step in that direction and it could go fast. For scene in resident evil that go for photo realistic like regular real urban setting with actual huge training data exist for seem to be quite something (it use a 5090 just to do the dlss pass right now....). For more stylised/non real world standard photography, maybe the game studio would need to provide a training set they made for which they do not need real time rendering, a render farm would do it.
 
Last edited:
LukeTbk said:
DLSS 5 seem already a clear first step in that direction and it could go fast. For scene in resident evil that go for photo realistic like regular real urban setting with actual huge training data exist for seem to be quite something (it use a 5090 just to do the dlss pass right now....). For more stylised/non real world standard photography, maybe the game studio would need to provide a training set they made for which they do not need real time rendering, a render farm would do it.
Click to expand...
it requires 2x 5090s to run according to DF
 
erek said:
opinion?
Click to expand...
if studio can make training set game engine (where like a pixar movie can define a wanted target computer cannot do in real time) combined with as minimal it need for ML to render that vision engine that can run super fast on very simple hardware, this could be really great, displace all teh cost of rendering of millions of people over months to a single event.
 
LukeTbk said:
just one its used for the dlss 5, the other is used to do the regular rendering.
Click to expand...

You can run games from the 2000s on integrated GPUs, or maybe use such a tiny amount of the 5090 that it doesn't matter.

But we don't have enough details on how it works yet. Maybe DLSS5 works better with more detail. There is probably an optimal amount of detail to add and it's probably pretty subjective based on the developer's goals for how they want the game to look.
 
LukeTbk said:
if studio can make training set game engine (where like a pixar movie can define a wanted target computer cannot do in real time) combined with as minimal it need for ML to render that vision engine that can run super fast on very simple hardware, this could be really great, displace all teh cost of rendering of millions of people over months to a single event.
Click to expand...
i don't like the look, seen this 1.5 years ago already in still images,

i've seen this look before in static photos

people would cut and paste screenshots from GTA Online to ChatGPT and get a similar result

it's instantly recognizable and ends up all looking the same, not really artistic naturally
 
erek said:
i don't like the look, seen this 1.5 years ago already in still images,

i've seen this look before in static photos

people would cut and paste screenshots from GTA Online to ChatGPT and get a similar result

it's instantly recognizable and ends up all looking the same, not really artistic naturally
Click to expand...
here's an example from gTA ONLINE, so this tech definitely is unimpressive to me having seen similar looks for almost 2 years

1773691944027.png


1773691982785.png
 
It's WEIRDLY uncanny. Like balls deep in uncanny valley.
 
socK said:
It's WEIRDLY uncanny. Like balls deep in uncanny valley.
Click to expand...
it's gonna all look the same, as in obviously AI

novelty initially at best

i seen the exact same looking faces 1.5+ years ago when people copy and paste screenshots into chatgpt
 
socK said:
Nobody watches the video they're quoting

The final version is intended for one card and is already working on one card internally

The 2 card thing is just for this demo.
Click to expand...

But if they don't misquote incorrect information that's been corrected multiple times then how will they hate on it?

LOWLRHs42LPUDtal3m.gif
 
I've long been flummoxed by this weird "fake frame" purity test in a space where, uh, everything is fake. Rasterization is fake, texture mapping is fake, lighting is fake.... and yet frame interpolation is somehow a sin. I don't get it.
 
erek said:
it's not real yet, proof will be in the pudding and as many suspect likely dialed back on quality and or performance
Click to expand...

so then again even if, the information you were quoting that is proven incorrect by just 10 seconds later in the video, will likely also even be incorrect by what you're saying now then (if it's dialed back and dialed down, it wont be as difficult to run!)
 
kram182 said:
so then again even if, the information you were quoting that is proven incorrect by just 10 seconds later in the video, will likely also even be incorrect by what you're saying now then.
Click to expand...
how's it incorrect? they're suggesting it'll be improved enough to run with 1x 5090 later

optimization usually means it's dialed and scaled back for production.. demos often overshoot to reel people in and wow them

beyond that, the look of it.. i'm disappointed because it's just like i said... chatgpt processed looking, completely novel and people will won't be impressed for very long at all.. characters having a similar appearance and mundane
 
erek said:
how's it incorrect? they're suggesting it'll be improved enough to run with 1x 5090 later

optimization usually means it's dialed and scaled back for production.. demos often overshoot to reel people in and wow them

beyond that, the look of it.. i'm disappointed because it's just like i said... chatgpt processed looking, completely novel and people will won't be impressed for very long at all.. characters having a similar appearance and mundane
Click to expand...


They said in the video it's running already in the lab on just 1x GPU and that that's their target for release - 1x GPU - they said they only ran 2x GPUs for the demo just as precaution/so everything went smoothly
 
kram182 said:
They said in the video it's running already in the lab on just 1x GPU and that that's their target for release - 1x GPU - they said they only ran 2x GPUs for the demo just as precaution/so everything went smoothly
Click to expand...
dog and pony show

will wait for real independent reviews and launch
 
erek said:
optimization usually means it's dialed and scaled back for production.. demos often overshoot to reel people in and wow them
Click to expand...
but 4-5-6 months in the ML world right now is a good amount of time, the distalled version this fall could be better than the full model now, easily.

and by 1 gpu i doubt they mean 1 5090, it will be way larger than that base of support.

As for characthers having similar appearance, will depends how much the semantic control give you and how you can play with the inputs (and if they inject their own trained model to it), those are Nvidia heavy demo.
 
kram182 said:
Like any of that will stop you when it's actually shown as such

View attachment 791678
Click to expand...
like i said like 3-4 times, the aesthetics are indeed junk.. GTA Online people have been pasting screenshots into ChatGPT for the same result for almost 2 years and to see this Tech producing the same looking characters is of course less than interesting

it's not hate, it's disappointment
 
DukenukemX said:
It's stupid. If we're back to SLI then maybe upscaling was a mistake?
Click to expand...
Ugh why do people keep parroting this as a gotcha. It is an early model, through distillation and optimization they can lower the needed requirements. They literally say by the time it releases it should run on a single GPU
 
sphinx99 said:
I've long been flummoxed by this weird "fake frame" purity test in a space where, uh, everything is fake. Rasterization is fake, texture mapping is fake, lighting is fake.... and yet frame interpolation is somehow a sin. I don't get it.
Click to expand...

That is a strawman and you know it. Frame gen is basically no different from you turning on motion smoothing on your TV, you have the real actual frames and then there are the "fake" frames which the motion smoothing algorithm tries to guess what they look like based on the real frames on hand. Sometimes it succeeds without artifacts, other times it creates a mess, but the end result is still slower real performance and increased input lag because it has to buffer frames for the generation part, that thing takes time and it cannot guess the future, it has to work on old frames and hold them back before showing them. Nvidia is simply peddling an ancient solution as a miracle performance enhancer (which it is not) and that rubs people, who actually know how things work, the wrong way.
 
MaZa said:
That is a strawman and you know it. Frame gen is basically no different from you turning on motion smoothing on your TV, you have the real actual frames and then there are the "fake" frames which the motion smoothing algorithm tries to guess what they look like based on the real frames on hand. Sometimes it succeeds without artifacts, other times it creates a mess, but the end result is still slower real performance and increased input lag because it has to buffer frames for the generation part, that thing takes time and it cannot guess the future, it has to work on old frames and hold them back before showing them. Nvidia is simply peddling an ancient solution as a miracle performance enhancer (which it is not) and that rubs people, who actually know how things work, the wrong way.
Click to expand...

And has been explained before - we've hit a brick wall with pure raster other methods are needed - why isn't AMD (or Intel) just pumping out then pure raster only GPUs to beat and show up Nvidia and surpass them with raster only performance?
 
They could have called smoothness x2, x3, etc... instead of frame x2, x3 (frame kind of infer game engine state snapshot being rendered on the screen), calling fake frame for those extra interpolated generated frame seem perfectly fine and a distinction people get easily, the state of the game is some ground truth in a real frame in a way it is not for those (even thought it is of course about a made up world)
 
You must log in or register to reply here.
Back
Top