DLSS4 Super Resolution and Ray Reconstruction (Transformer, TNN)

[Scott_Arm]

We don't have a dedicated thread for the new DLSS4 transformer models for Super Resolution and Ray Reconstruction, so I thought I'd start one.

If you have screens or videos of particular cases where it works well, or fails miserably, share it here.

I found a scene in Alan Wake 2 with a moire pattern on some metal grating that's particularly bad. It has noise at pretty much all settings that just gets worse that further you push your upscaling.

Here's an Alan Wake 2 still at 1440p DLSS Ultra Performance (480p base resolution!)



I also took some screens at 1440p DLSS4 Performance in Remnant 2 that I thought worked very well. There is aliasing in motion, but overall it's very usable. Lots of chainlink and overhead wires, small detail in leaves as well as some moving particles. I can definitely play at this setting.

 

[RobertR1]

@Dictator what's preset K? saw it pop up with 310.2 but no real info on it. Not updated in nvapp either. Needs a manual toggle via dlsstweaks or such.

NVIDIA DLSS DLL 310.2 Download

This download provides various versions of NVIDIA's DLSS DLL for download. In this one file, which is bundled with all games that support NVIDIA's

www.techpowerup.com

 

[Scott_Arm]

@RobertR1 K should be the new default, from what I understand.

 

[DavidGraham]

Interesting insights into DLSS4 ..

-Older RTX 20 and RTX 30 cards drop performance by up to 25% using the TNN model. RTX 40 is often around a drop of 5%.
-The drop is greater when DLSS is combined with Ray Reconstruction or RTX HDR (anything that co-uses the tensor cores).
-The drop in performance is most at DLAA or DLSS Quality settings .. DLSS Ultra Performance doesn't drop as hard.

 

[trinibwoy]

I tried the TNN model in Hellblade 1 and it wasn’t too bad on my 3090. Quality mode cost about 1.5ms extra at 1440p ultrawide.

 

[DavidGraham]

Ray Reconstruction adds a huge amount of details to Spider-Man 2, especially for reflections and shadows.

 

[Scott_Arm]

Interesting insights into DLSS4 ..

-Older RTX 20 and RTX 30 cards drop performance by up to 25% using the TNN model. RTX 40 is often around a drop of 5%.
-The drop is greater when DLSS is combined with Ray Reconstruction or RTX HDR (anything that co-uses the tensor cores).
-The drop in performance is most at DLAA or DLSS Quality settings .. DLSS Ultra Performance doesn't drop as hard.

I think it's important to pay attention to the cards being tested. These are all the x060 cards. I'll take a look at Alan Wake 2 on my pc to see how much difference there is on my 3080, but bottlenecks may hit in different places in cards up the stack.

Edit: In Alan Wake 2 I'm losing about 10% by switching from CNN to TNN. It's about a 0.8ms frame time difference. That's in the scene I'm in. I'm not sure how much DLSS is scene dependent. I would guess not much. That's DLSS performance at 1440p.

 

[homerdog]

Ray Reconstruction adds a huge amount of details to Spider-Man 2, especially for reflections and shadows.

At around the 4:30 mark you can see the biggest different RR makes IMO. The ghosting is out of control without RR and gets completely dealt with when RR is on. Ghosting is one of those artifacts that, depending on severity, makes me not really want to use fancy effects. Very glad NVIDIA is taking it seriously.

 

[trinibwoy]

Impressive stuff. Too bad those RT shadows are stationary though. Hopefully this is just the beginning of transformer based upscaling and denoising. The first attempt works very well.

 

[Granath]

At around the 4:30 mark you can see the biggest different RR makes IMO. The ghosting is out of control without RR and gets completely dealt with when RR is on. Ghosting is one of those artifacts that, depending on severity, makes me not really want to use fancy effects. Very glad NVIDIA is taking it seriously.

What the gosthing has to do with RT?

 

[homerdog]

What the gosthing has to do with RT?

Timestamped at 143 seconds:

RT doesn't inherently cause ghosting, but optimizations used to make RT performant can cause ghosting. Looks like the new denoiser is better about this.

Of course this is not the only source of ghosting in modern games. DLSS4 upscaling is also better about ghosting in general in my limited testing.

 

[Scott_Arm]

GN just released their video on the quality of DLSS4 FG and SR.

 

[Potato Head]

GN just released their video on the quality of DLSS4 FG

I couldn’t even get through 60 seconds. God he is annoying.

 

[trinibwoy]

I couldn’t even get through 60 seconds. God he is annoying.

Not sure what happened but he used to be a relatively serious guy. Now he’s a parody of himself with the memes and stupid thumbnails and feigned outrage at every turn. Maybe someone told him that’s what it takes to make it in the YouTube scene. Super annoying.

There was some good content in the video though. I think that’s the best display of framegen artifacts I’ve seen to date. Definitely a flawed tech. I wonder how much of it is due to games not providing adequate motion vectors.

 

[Shifty Geezer]

Not sure what happened but he used to be a relatively serious guy. Now he’s a parody of himself with the memes and stupid thumbnails and feigned outrage at every turn.

Likely he saw moving in that direction increased engagement and revenue.

Maybe someone told him that’s what it takes to make it in the YouTube scene. Super annoying.

And if he tried it and his viewership went down, he'd have stopped, no?

There was some good content in the video though. I think that’s the best display of framegen artifacts I’ve seen to date. Definitely a flawed tech. I wonder how much of it is due to games not providing adequate motion vectors.

Timestamps are really useful! If you can guide Potato Head around the presentation to a key point, it's easier to include that into the discussion.

 

[trinibwoy]

And if he tried it and his viewership went down, he'd have stopped, no?

For sure. That’s what the people want.

Timestamps are really useful! If you can guide Potato Head around the presentation to a key point, it's easier to include that into the discussion.

The video is well annotated. Frame gen comparisons start around 16:00.

 

[Albuquerque]

Yeah, I do miss "old Steve" versus this new thing he's doing.

The content was indeed good, and agreed with @trinibwoy regarding their excellent demonstration of FG artifacting. I wonder how much of this is training data versus some inherent limitation of the technology?

 

[DavidGraham]

HUB analyzed DLSS4 in great details, they say it's simply amazing! DLSS4 Performance looks like DLSS3 Quality/native.

 

[RobertR1]

As Tim stated, the texture clarity is worth it's weight in gold after years of TAA. I really like the way he broke it down into sub categories which will make future versions and competitive options much easier to compare against.