Hello all! This is for RTX 4XXX series owners....

Which do you use and find better with less artifacts? Lossless Scaling Frame Generation (LSFG) or the default Frame Generation (FG)?

Edited June 19, 20241 yr by iFlySimX

30 minutes ago, iFlySimX said:

Hello all! This is for RTX 4XXX series owners....

Which do you use and find better with less artifacts? Lossless Scaling Frame Generation (LSFG) or the default Frame Generation (FG)?

Use them both: LSFG combines with FG for best results and with no artefacts on my pc: honest! If you're seeing artefacts then it's because of differences in hardware and/ or settings: LSFG seems to work best when it has a solid "base" FPS from which to interpolate. Fewer "base" frames = less information from which to make good guesses about how to interpolate = higher risk of bad guesses = higher risk of artefacts. It's also possible that the higher number of RTX cores or whatever they are called in a 4090, gives an edge over the 4080/4070.

Edited June 19, 20241 yr by tfm

What is TAA doing in the poll? And why with DX11 only? Makes no sense to me, it is a completely different story compared to AI generated intermediate frames...

2 minutes ago, AnkH said:

What is TAA doing in the poll? And why with DX11 only? Makes no sense to me, it is a completely different story compared to AI generated intermediate frames...

Because Lossless Scaling Frame Generation works with DX11 & TAA, I think...

Edited June 19, 20241 yr by iFlySimX

I’m using both and it’s so far seems to work great. I can’t vote because there is no option for what I’m using. 

10 minutes ago, fakeflyer737 said:

I’m using both and it’s so far seems to work great. I can’t vote because there is no option for what I’m using. 

Added option.

 

How does FG and LS work together? FG is pretty straight forward as it just takes the average of two following frames and makes a guestimate of how it should look.

When you add LS this gets a bit more confussing. What does it use as reference frames, and where are they inserted in the timeline? I guess the FG frame is not available to LS to use as a reference frame. Won't this cause a lot of artifacts/ghosting?

 

1 hour ago, iFlySimX said:

Because Lossless Scaling Frame Generation works with DX11 & TAA, I think...

Sure, but it also works with DX12 & TAA...

4 minutes ago, AnkH said:

Sure, but it also works with DX12 & TAA...

Right...so some people may prefer to use DX11....Not sure where you are trying to get at....

1 minute ago, iFlySimX said:

Right...so some people may prefer to use DX11....Not sure where you are trying to get at....

All I am saying is that the selection of options in the poll makes no sense to have only "DX11 & TAA" listed, but no "DX12 & TAA", and the explanation is that LS works with DX11 & TAA... It does with DX12 & TAA, so this should also be an selectable option. And, as mentioned, because DX11 & TAA does not results in artefacts, because it is simply no frame gen technique, it anyway makes no sense in the poll at all. Just call it "no FG method at all" for those people not using anything for frame generation because they do not like artefacts at all, no matter if they are due to Lossless Scaling FG or native FG. 

1 hour ago, espent said:

How does FG and LS work together? FG is pretty straight forward as it just takes the average of two following frames and makes a guestimate of how it should look.

When you add LS this gets a bit more confussing. What does it use as reference frames, and where are they inserted in the timeline? I guess the FG frame is not available to LS to use as a reference frame. Won't this cause a lot of artifacts/ghosting?

 

All good questions and I don’t think anyone has been able to answer without ‘speculating’. LS is unlikely to have access to the native FG been produced internally by the RTX 40xx GPU tensor cores. So that means it is excluding those frames in its calculations and I cannot see how that can be a good thing when combined with native FG.

15 minutes ago, AnkH said:

All I am saying is that the selection of options in the poll makes no sense to have only "DX11 & TAA" listed, but no "DX12 & TAA", and the explanation is that LS works with DX11 & TAA... It does with DX12 & TAA, so this should also be an selectable option. And, as mentioned, because DX11 & TAA does not results in artefacts, because it is simply no frame gen technique, it anyway makes no sense in the poll at all. Just call it "no FG method at all" for those people not using anything for frame generation because they do not like artefacts at all, no matter if they are due to Lossless Scaling FG or native FG. 

Next time you create the poll 😉

19 minutes ago, RaptyrOne said:

All good questions and I don’t think anyone has been able to answer without ‘speculating’. LS is unlikely to have access to the native FG been produced internally by the RTX 40xx GPU tensor cores.

Lossless Scaling is actually pretty simple: it just screen-captures the selected game window and then uses these new movie-AI-upscaling algorithms to scale the image up and/or generate intermediate frames, then overlays that on top of the window position (which is also the reason it can't run on an exclusive-mode fullscreen game, since it needs to be able to capture the window area). However, it's not operating in real time. It needs at least the most recent frame and the frame before that to generate the upscaled frames, so the minimum delay is at least a whole frame, plus processing time. As you might also gather from this description of operation, it's not exactly "lossless". I'm not sure what they mean by that, since the AI movie-upscaling model is manufacturing its own new pixels and frames, which if you've ever watched AI upscaled or sharpened stuff you know can also have its own funky artifacts. However, it is a clever usage of this technology for sure, and doesn't need any specific GPU series necessarily.

FG frames are real display frames that can be captured (since you can see them), so those will be utilized by LS. FG itself is not subject to the same kind of latency, as it operates on a different paradigm: the game sends a buffer with data containing the motion vector of each pixel to the GPU, since the game has knowledge of every 3D object's and camera's position and motion vectors. Since this is sent with the frame itself, FG uses that data to draw intermediate frames using the pixel motion vectors while it waits for the next frame, as opposed to having to actually wait for the next frame to arrive and then interpolating between the two. As you can imagine, though, this has its own artifacts, such as when the pixel motion vectors from frame to frame change directions sharply and the algorithm makes wrong predictions about where the pixels are going to end up, or for things that by nature have no motion vectors, like animated textures.

There is a big difference between "better" and "fewer artifacts". For instance TAA has fewer artifacts (better) but less fps (worse). And so TAA is both better and worse. With my 4070 running lots of pixels which is a tough task, then TAA overall is worse (terrible fps). But fewer artifacts.