Info HWUB: Big Native Vs DLSS vs FSR comparison

[coercitiv]

If you thought about, why not mention it?

I did, you probably didn't notice because you thought I was here to defend my dissertation on NN Supersampling:

The most interesting aspect of using this tool is not the 1080p assessment by itself, but rather the relative change as we increase input resolution

 

[Heartbreaker]

I did, you probably didn't notice because you thought I was here to defend my dissertation on NN Supersampling:

I mean why didn't you mention that: VMAF would flag cases where DLSS is better than native, as being worse than native.

AKA backwards of intended result.

That's the problem using VMAF to check DLSS/FSR against native.

You can't use it to form any conclusions when it has this kind of behavior.

 

[AtenRa]

FSR 3 + Frame Generation at its finest
Pictures taken from the NVIDIA Video bellow, take a look at the Hammer Swing ......

Native always better no matter what they try to make you believe.

 

[Heartbreaker]

Frame Generation is garbage.

It's completely different from the scaling done in this video from HWUB.

 

[marees]

FSR 3.1 still lags behind AI upscalers in some ways, obvious even in HUB’s 4K/Q tests with 40% of the games more than slightly worse than DLSS. What is special about those bad 40%?

• Lower rendering resolutions: 4K/Quality is most often competitive. 4K/Balanced or 1440p/Quality are usually satisfactory but you can see a bigger gap. 1440/Performance or 1080p/anything have improved a lot from FSR 2.2, but the AI upscalers are still significantly better.
• Failure Modes: FSR has certain input scenarios it doesn’t handle well. The best FSR games are often those that simply avoid those scenarios. In particular, complex alpha particle effects persist a difficult problem for FSR even with high-quality integrations and low scaling factors.
• Integration difficulty: FSR’s analytical system requires more work from game engines for special inputs: the reactive mask and the transparency & composition mask. Many games don’t care to produce those, since no other upscaler needs them. They're both optional, a compromise for adoption; FSR makes do without them but at a cost in quality.

In all areas above, AI-based upscalers often perform better because they “hallucinate” around missing or lower-quality information. It turns out AI hallucination is a problem in the general case but not always. Extra detail that’s fake/wrong compared to ground-truth is still great, if:

1. You don’t have the ground truth! Always true when playing a game.
2. The result looks nice. “Looking nice” does not require being correct.

The PS5 Pro, RDNA 4 & FSR 4.0

Sony’s announcement of the PlayStation 5 Pro rekindled speculation about new gaming tech: a new RDNA 4-based SoC, a new AI accelerator, and…

medium.com

 

[static shock]

RDNA4 for PS5Pro, please!
This rocks!