AMD's FSR 3 upscaling and frame interpolation *spawn

Just tested it. Incredible technology if you have a high refreshrate monitor. Sadly reflex or antilag is not available in Forspoken, so input lag is acceptable but not great. I can imagine with it on it will be even more awesome.
The demo won't even launch for me. I have little interest in the game but I was excited to try FSR frame gen.
 
Bearing in mind I have a sample count of one as I only have the Forspoken demo (which has bad IQ anyhow) but the artifacting with FSR3 is unbelievably bad that I would think people were trolling about it having good IQ if I didn't know the internet better.


Took a quick video and just look at the hair phasing in and out of existence - this is with the new Native AA option so a best case scenario (for 1440p anyhow). I appreciate that this is a tough test for any frame gen but stuff like this is obvious just playing normally. Do the same thing with DLSS3 in something like Spider-Man and you will get noise around the character that you may or may not notice without pixel peeping but a blind man on a galloping horse could surely notice chucks of the character disappearing.

It's hard to speak on how it feels as my base frame rate is already high at 1440p but I'll see if I can get it to around (or below) 60fps later and see how it compares to DLSS3 but IQ wise (IN THIS GAME AT LEAST - OTHERS MAY BE BETTER!) it is laughably bad.

I'm playing Cyberpunk 2077 atm with DLSS3 and noticed fewer frame gen issues in the 20 hours I've been playing Phantom Liberty than I did in the first 20 seconds of messing about in the Forspoken demo.
 
Here’s a rebuttal:
No, in it's current state, it's completely useless.

When standing still and when moving the character with the keyboard, Fluid Motion Frames appears to be working fine. However, the moment we start moving the camera with the mouse, AFMF completely breaks down.



What this basically means is that AMD’s driver-based Fluid Motion Frames solution is completely useless. At least in its current state. I don’t know about you, but I want to move the mouse when I’m playing games like Resident Evil 3, Starfield, Far Cry 6 or Cyberpunk 2077. So, due to this major issue, and until AMD fixes it, we suggest staying away from AFMF!



 
Last edited:
No, in it's current state, it's completely useless.







I’ll wait for the final release. If they solve the mouse issue this will be the last hurdle for them.
 

Took a quick video and just look at the hair phasing in and out of existence - this is with the new Native AA option so a best case scenario (for 1440p anyhow). I appreciate that this is a tough test for any frame gen but stuff like this is obvious just playing normally. Do the same thing with DLSS3 in something like Spider-Man and you will get noise around the character that you may or may not notice without pixel peeping but a blind man on a galloping horse could surely notice chucks of the character disappearing.

It's hard to speak on how it feels as my base frame rate is already high at 1440p but I'll see if I can get it to around (or below) 60fps later and see how it compares to DLSS3 but IQ wise (IN THIS GAME AT LEAST - OTHERS MAY BE BETTER!) it is laughably bad.

I'm playing Cyberpunk 2077 atm with DLSS3 and noticed fewer frame gen issues in the 20 hours I've been playing Phantom Liberty than I did in the first 20 seconds of messing about in the Forspoken demo.
Your post is the best resource regarding AMD frame interpolation so far over the whole internet. :D

Looking at pharmas post just above it seems working well, but there is not much motion shown, so that's expected.
Your's shows plenty of artifacts in a worst case scenario to analyze.

There's especially this artifact catching my interest: We can see a hard cut around the character where parts of the sky have been attempted to be reconstructed or guessed.
The artifact looks exactly like standard methods to do this to achieve screenspace motion blur or DOF. Basically copy pasting nearby sections of the background to deal with missing information. It's quite interesting those methods are good enough in practice although the artifacts in a still image are heavy.
But obviously it does not work well in this extreme case of fast camera rotation.
My first impression is: 'Damn, it's not good enough. So we need those otherwise useless tensor cores indeed, just to make crutches work we ideally should not even need.' : (

But i'm not yet willing to accept this. I think AMD could improve this simply by blurring the mask so the hard cut becomes smooth, and then we would not notice the trickery so easily.
Likely the same applies to other issues as well.
I hope they can improve it further and don't leave it at this seemingly early state.

On the long run, at some point i'll have to give up my resistance against ML for rendering games, i'm afraid. But no - not yet, please.
 
Your post is the best resource regarding AMD frame interpolation so far over the whole internet. :D

Looking at pharmas post just above it seems working well, but there is not much motion shown, so that's expected.
Your's shows plenty of artifacts in a worst case scenario to analyze.

There's especially this artifact catching my interest: We can see a hard cut around the character where parts of the sky have been attempted to be reconstructed or guessed.
The artifact looks exactly like standard methods to do this to achieve screenspace motion blur or DOF. Basically copy pasting nearby sections of the background to deal with missing information. It's quite interesting those methods are good enough in practice although the artifacts in a still image are heavy.
But obviously it does not work well in this extreme case of fast camera rotation.
My first impression is: 'Damn, it's not good enough. So we need those otherwise useless tensor cores indeed, just to make crutches work we ideally should not even need.' : (

But i'm not yet willing to accept this. I think AMD could improve this simply by blurring the mask so the hard cut becomes smooth, and then we would not notice the trickery so easily.
Likely the same applies to other issues as well.
I hope they can improve it further and don't leave it at this seemingly early state.

On the long run, at some point i'll have to give up my resistance against ML for rendering games, i'm afraid. But no - not yet, please.
It's a shame the video is only 60fps then compressed on top of that as it actually looks much worse in person than it does in the video but I wanted it to be as obvious as I could make it as so many of the Youtube videos so far are so badly compressed that it's hard to tell what is actually happening.

Stuff this bad does happen with DLSS 3 too (like frame by frame in a Spider-man video with DLSS 3 and you'll 100% be able to find frames where his arm disappears or something) but it's just 1 or 2 frames out of 140 while with FSR 3 most of the frames appear to have major artifacts which is really obvious even in real time.

I guess if frame gen can't handle fast motion I don't see the point as I'd rather just play at flat 60fps with nice IQ and slower paced games where FSR3 would work fine are less in need of 120fps+.

Here are some screens to show how bad the artifacts are and bear in mind most of the generated frames look like this if you move the camera quickly at least in Forspoken. Again you will 100% be able to get crazy looking artifacts in DLSS3 games but it won't be in the majority of the generated frames which makes it hard to spot in real time.

FORSPOKEN Demo 30_09_2023 11_52_28.png
FORSPOKEN Demo 30_09_2023 11_52_28.png (2560×1440) and 3 more pages - Personal - Microsoft​ Ed...png
FORSPOKEN Demo 30_09_2023 11_52_34.png
FORSPOKEN Demo 30_09_2023 12_07_03.png
FORSPOKEN Demo 30_09_2023 12_07_41.png (2560×1440) and 5 more pages - Personal - Microsoft​ Ed...png
FORSPOKEN Demo 30_09_2023 12_16_19.png
FORSPOKEN Demo 30_09_2023 12_16_21.png
FORSPOKEN Demo 30_09_2023 12_17_29.png
FORSPOKEN Demo 30_09_2023 12_17_34.png

I just spammed the screenshot key so had no way of cherry picking these to look bad, I've not gone through a video frame by frame looking for gotchas or anything it's just that most of the screenshots looked like this so I had an easy time.
 

Attachments

  • FORSPOKEN Demo 30_09_2023 12_07_46.png
    FORSPOKEN Demo 30_09_2023 12_07_46.png
    989 KB · Views: 21
I think FSR 3 is blending over multiple frames in rapid movement to create the illusion of smoothness. Here is an example from Immortals:


You can clearly see multiple previous frames. In movement the latency goes up, too. In this case with a 50 FPS -> 100 FPS latency is over 100ms in movement...
 
I think FSR 3 is blending over multiple frames in rapid movement to create the illusion of smoothness. Here is an example from Immortals:


You can clearly see multiple previous frames. In movement the latency goes up, too. In this case with a 50 FPS -> 100 FPS latency is over 100ms in movement...
that might explain why some of the screengrabs that @Smutz shared show those square artifacts that look like one of my MS Paint "photoshops" when you look close enough -which most people won't when it comes to a regular image, where people aren't going to zoom a 400%-, but it totally reminded me of that when using MS Paint and copying-pasteing very small sections of an image to make certain parts or colors look more uniform.

Example, this imagen shared by Smutz:


forspoken-demo-30_09_2023-12_17_34-png.9708
 
Fluid motion frames will probably never be good.

FSR3 is something I’m very interested in. If I can push 120 to 240, I’ll be happy. At higher frame rates you should have smaller differences between frames which should help reduce artifacts.
 
Using Forspoken demo, and FG works fine with VRR. Even the guy slamming driver-level FG posted on the previous page is seeing it working.


And another confirmation from someone who doesn't really love AMD:


Not sure how good the nvidia overlay latency readings are, it seems to barely add any latency. 140FG vs 120( max without FG ) is same latency at 7-8ms.
That is the latency of the rendering - if a game does not have reflex markers, the Nvidia overlay cannot really read the total system latency.

The way to actually test VRR is to see what happens with a *variable* frame-rate by looking at the output frames (not an internal frame-time graph). The frame-rate cannot be pinned to a number, whether the refresh rate or something else.
 
/edit: Did another test V-Sync off and 100 FPS limit.

Adaptive-Sync doesnt work properly. For me it looks like that there are multiple scenarios in which FRS 3 is falling back to V-Sync and uses buffers to hit the refresh cycle.
A few examples with active Adaptive-Sync and 170Hz:
  1. Without V-Sync FSR 3 is useless. There is no sync within the VRR range.
  2. With V-Sync and no frame limiter my display shows most the time 170hz regardless of the rendertime (i.e. 100 FPS). But FSR 3 is "resyncing" the display ASAP (guess the next normal frame) and the display overlay shows a lower number.
  3. With V-Sync and a frame limiter of 100 FPS Adaptive-Sync works but only under a certain GPU usage (<60%). When the GPU usage goes up, FSR 3 is falling back to 2.
  4. With V-Sync off and a (low) frame limiter of 100 FPS Adaptive-Sync works and latency was only twice as high (average pc latency 25ms -> 55ms). Best experience...
On a nVidia GPU there are a lot of buffers involved with FSR 3. Latency is 3x+ higher than without it.

So the best result is archived by "low" GPU usage, V-Sync off and a frame limiter... Or you could just play without it and having the same frames, a cleaner and sharper image and lower latency. The only positive aspect of FSR 3 is that it looks smoother because of the blurrier picture. Like LCD vs. OLED at 30 FPS.
 
Last edited:
Back
Top