So the framegen is just interpolation, right? LS can't do "predictive" framegen just like it can't do FSR2+ given that it all it has to work with is the final rendered image?

[kalirion]

Just making sure I understand what this "framegen" is.

Given that LS doesn't have access to the rendering process of the game and acts only on the final rendered image, this "framegen" is LS making the monitor lag a frame behind so that it can insert an interpolated frame in between two frames which were rendered the normal way by the game/gpu? Sort of like Double/Tripple Buffering on steroids?

[(ノ°□°)ノ]

Okay. Interpolation is the same as generation (Nvidia's marketing name which became popular and is now used more often than interpolation). All known technologies for use in games - DLSS 3, FSR 3, AFMF, as well as LSFG do interpolation (generation), that is, they create a new intermediate frame between two existing ones. All of these add latency. Accessing game data only affects the quality of this new frame.
What you're calling 'predictive' is called extrapolation. No one is using it yet.
Double/tripple buffering has nothing to do with this topic.

[kalirion]

My understanding of DLSS 3, and possibly FSR 3, framegen was that they are predictive. That's why in DLSS 3 you can get weird results if you turn around too quickly, for example, as that's not something that the AI could predict.

If you just interpolate two frames to insert a new one between them, why would having access to the vectors and whatnot even help.

Edit: https://www.makeuseof.com/what-is-dlss-3-can-you-use-on-existing-hardware/

So just like DLSS, DLSS 3 uses tensor cores to increase the resolution of the frames, but it also has special optical flow accelerators which help the GPU predict frames. To predict the frames, the optical flow accelerator gets several high-resolution data frames generated by DLSS. The optical flow accelerator then uses this data to generate the optical flow field.

[(ノ°□°)ノ]

Originally posted by kalirion:

My understanding of DLSS 3, and possibly FSR 3, framegen was that they are predictive.

No, they are not.
For FSR 3 you can read about here: https://gpuopen.com/fsr3-in-games-technical-details/
They show it's pipeline and clearly state that this is interpolation and adds latency.
For DLSS you can't, for obvious reasons.

Originally posted by kalirion:

So just like DLSS, DLSS 3 uses tensor cores to increase the resolution of the frames, but it also has special optical flow accelerators which help the GPU predict frames. To predict the frames, the optical flow accelerator gets several high-resolution data frames generated by DLSS. The optical flow accelerator then uses this data to generate the optical flow field.

The fact that they used the word "predict" doesn't mean what you think. The creation of an intermediate frame is also a prediction because it never existed. Then they say it uses an optical flow accelerator to generate an optical flow field from multiple frames. That means it generates it BETWEEN those multiple frames, because you can't generate optical flow from a frame to a unseen frame.

[d0x360]

Nvidia is full of crap in regards to it's actual requirements which is made very clear by the fact that you can easily mod FSR frame gen into a game on Nvidia GPUs and it's identical to Nvidia frame gen in every way as is modding dlss frame gen into a game that only has FSR frame gen. Zero difference in quality regardless of fast moving or not.

Nvidia is using terms to be deceptive in an attempt to get people to upgrade to the 4000 series when they may not actually need to do so based on their needs.

If so was actually used to analyze and generate the frame then it's processing would be done by the tensor core which would in turn break fsr frame gen injection.

[MisterEdd]

Originally posted by d0x360:

Nvidia is full of crap in regards to it's actual requirements which is made very clear by the fact that you can easily mod FSR frame gen into a game on Nvidia GPUs and it's identical to Nvidia frame gen in every way as is modding dlss frame gen into a game that only has FSR frame gen. Zero difference in quality regardless of fast moving or not.

Nvidia is using terms to be deceptive in an attempt to get people to upgrade to the 4000 series when they may not actually need to do so based on their needs.

If so was actually used to analyze and generate the frame then it's processing would be done by the tensor core which would in turn break fsr frame gen injection.

To be fair there was already a mod to use dlss 3 on older hardware. Just like nvidia said the
optical flow acceleration was indeed not fast enough to consistently interpolate the frames. Also lets be honest AMD frame gen isn't even close to the same quality. Nvidia could make a driver based shader interpolation but it wouldn't look nearly as good and game devs wouldn't be as incentivised to make a proper implementation of dlss 3. In the end you can easily mod any dlss 3 game to use fsr framegen so who really cares. Nvidia doesn't need to restrict dlss 3 to rake in money. They literally cannot produce gpu's fast enough anyways. Restrictive VRAM on all but the flagship cards are the real issue.