AMD FidelityFX Super Resolution 2.0 - Deathloop preview

Keep on looking at static rendered objects.

It erases alpha effects, particles even on quality mode. Deathloop was selected because it hardly has any, and the ones that it does are very "cartoony" which scale better. Its great for what it is though.

 

I'm a little confused by "It erases alpha effects, particles even on quality mode". As in removes the effect entirely from the game? Or just looks wrong? Because it's applying to the grass, which has lots of Alpha texture data and it looks great. But then things like explosion effects, lasers and that flickery wall text stuff all look off.

Gah I wish I could bring myself to buy this game so I can just look myself. But I can't justify it given it's price and that it will eventually come to GamePass for free

 

You did cherry picke ONE screenshot , but FSR 2 over sharpening is obvious on every single one. Even without clicking for larger image (or perhaps more so). And it does what (over) sharpening usually does:

• introduces more aliasing (fact)
• without bringing in any more details (imho)

native vs FSR 2

https://www.purepc.pl/test-amd-fsr-...-0-oraz-nvidia-dlss-w-grze-deathloop?page=0,3

 

DLSS is still (occasionally? often?) struggling with PProcessing. Will be interesting to see how FSR 2 fares. Place ur bets

 

It's worth remembering, that Deathlooops default TAA is rather blurry. Where as FSR and DLSS both replace that TAA with there own. Which is going to yield a sharper result with zero sharpening filters no matter what.

So to get this "right" you would need Deathloop screenshots at Native with TAA off, native with TTA on, DLSS Quality and FSR 2 Quality with sharpness deault and put to zero. To really start grasping what's what imho.

... It's 25 bucks @ GreenManGaming... aaaaahhh, must resist lol

 

@Trunks0

Very good point my man.

(hell im not keeping the entire rendering pipeline in my head. thats why we have junior gurus to remind us )

 

Not so much either, "native" without TAA means badly undersampled and textures might even show less detail than with DLSS/FSR 2.0 due to negative texture LOD bias utilized by them.
Only valid ground truth is SSAAed comparison, and hypothetical 64xSSAA ground truth of course would be totally crisp.

 

@aufkrawall2

he did not say 'ground truth'. just mere-mortal 'native'

 

Yes, that's the problem: "native" is undersampled garbage. You need supersampling or the result is broken by definition. It doesn't make sense to compare against something broken.

Particle effects and alpha tests are rendered with FSR 2.0. But the funny thing is: The particles are easier to spot with DLSS, as they show more ghosting trails than with TXAA or FSR 2.0:


When people think the broken result would be the correct one because it's DLSS. Such "pro users"...

 

OK its just one game - lets not die on the very first hill. Like wait for the Alamo or Dien Bien Phu...

The left screenshot with only 3 sparks does remind me of Death Stranding with the old DLSS.
Sparks missing on the carrying case, remember?

 

Its very easy to understand. FSR2 is not up-sampling the effects, and they are being removed. The cost to calculate and resample is too much work, without separate hardware designed specifically for the task.

 

These particles are easier to spot in-game than to capture them in a screenshot. I think the are mostly fine. A bit aliased perhaps, but better than ghosting.
There are other valid issues though:
-the camera sensor rays look a bit too aliased
-there seems to be aliasing caused by some transparent smoke effects
-bloom effect of some trinkets that get dropped by enemies looks broken

I guess nothing irreparable, let's hope next titles won't inherit this.

 

Give this article a read/peek. Because you keep coming back to AI or dedicated hardware being needed to solve the issues. But if you look at the underlying technique at play, which is TAA, you may think otherwise

https://www.elopezr.com/temporal-aa-and-the-quest-for-the-holy-trail/

Nothing we are seeing in FSR 2 can't be fixed in the same way it is with regular TAA. All these techniques (TAAU, DLSS,TSR and FSR 2 etc) are really iterations of the underlying tech of TAA. And when you look at the problems presenting in FSR 2 and DLSS in Deahtloop ... they are the same ones you encounter in TAA. nVidia and Intel are choosing to use AI to try and resolve those issues. But its not the only way to do it and I think that is especially obvious when you look at TAA go from where it started to today and where TAAU has gone from it's early versions to today.

*editied, as it kind of read rude to me and that's not intentional. So sorry about that*

 

No. I keep coming back to additional hardware being required to process the image like DLSS can. There is not a single other solution available yet, imo, that truly comes close. IMO, even in that DF video about FSR2, DLSS was the most impressive thing... by far.

What are we even talking about? Just so you know, I am advocating this hardware as unavoidable in the future. I am being critical of AMD for, yet again, opting not to include hardware that has been making waves in the industry for 2 solid years now. Are you sure you want to argue against that? Its clearly a feature destined to be driven by hardware!

 

Nvidia and Intel use AI to train as many NN samples as possible to render an upscaled image with the most accuracy, especially with objects in motion. TSR and FSR 2.0 rely on less information and is based on the previous frame which can cause anomolies, (ie, the radio antenna movement causing the distortion in the wheat field). I believe FSR 2.0 in these situations is supposed to blur when the previous frame lacks enough information.

Currently FSR 2.0 and TSR are both good options in their current implementation though some rendering issues will be difficult to resolve due to having less samples to work with. Rendering discrepancies will become more obvious once more games and different art styles start to use FSR 2.0.

 

Ah, that makes more sense. AMD didn't focus in on the AI sector like nVidia did and hadn't really even attempted to really address it till around this generation. So I wouldn't expect tensor style acceleration for abit. GPU designs are planned/taped out years in advance. So you can't just adapt like that. Not that they couldn't consider another method to accelerate an up-scale method if it's not going to use AI anyway.

And if you don't think FSR 2.0 is close... I'm flummoxed. As they closer than they are far apart. And the differences that we are seeing, are fixable without an AI if you look at how temporal tech works(The Disocclusion and Color Clamping need refinement for instance. But you can do that without training an AI) . We will know allot more though once the source code for FSR 2 is released and I'm excited to see people dig into all the underlying tech and how it's doing things.

They are working with the same data set. The algo used by DLSS though is a pre-trained AI model with years of refinement at this point. It looks great.

But, looking at the history of TAA's own progression in quality, it should be obvious the same can be done with temporal up-scaling given how related they are. Without using an AI algo. It does take longer though, I think that's kinda clear too given the reality of where we are. But looking at FSR 2 and TSR, I can't help but feel we are starting to see the non-AI methods of temporal up-scaling catch up.

 

I wonder how FSR 2.0 compares to UE5's TAAU.

It seems like the currently hierarchy at the moment is: [DLSS 2.3.X (latest) > FSR 2.0 (or TAAU?) > Spatial FSR 1 > Traditional single frame upscalers (bilinear / lanczos / bicubic)].

I'm as happy as the next guy that FSR 2 is here and that we have options that don't require the Tensor cores, but having tried the options available in Deathloop recently, at least with a 1440p output, DLSS does look noticeably better to my eye than FSR 2 (FSR 2 is still much better than FSR 1 of course, great that we have it as an option, I would use it for sure if I needed the performance and lacked a modern Nvidia GPU).

Part of why I wonder if people are saying FSR 2 is "sharper" than DLSS is because Deathloop defaulted the post process sharpen (that goes on top of FSR 2) to the max (at least it was max when I checked the settings). This did make the image look very sharp, but sorta also made my eyes bleed at times so for me I scaled that back a lot personally.

 

I didn't realize it wasn't scaling/reconstructing those elements, that's a pretty big deal seems to me. I wonder if it will be feasible for them to include those in their reconstruction down the line in future versions of the tech?

I did notice one of the issues Digital Foundry pointed out where in Deathloop with FSR 2 sometimes the AI will move and their animation breaks apart/pixelates. I couldn't see that with DLSS though sometimes there is visible ghosting.

 

I would say that DLSS and FSR deal with the same information, basically color buffers, motion vector buffers and depth buffers.
The difference is how each solution deals with the information. FSR is a generic "dumb" way of dealing with the information these buffers provide.
DLSS has the ability to add weights to each input, allowing to choose what data is more relevant for each scene.
FSR 2.0 gets close to 90% of what DLSS 2.x does. But it's that last operation that is making a good deal of difference.
DLSS 2.0 had similar issues as FSR 2.0, when it was released. But as the AI was trained, it managed to get more accurate weights to then be multiplied by the input.

The issue here is that DLSS 2.x is a motion target, increasingly improving it's quality.
AMD might be able to improve FSR 2.0 a bit, if done manually. But unless they start using a deep learning algorithm, like nVidia and Intel, the gap will increase further.
AMD does have Matrix Core Units in their CDNA2 arch. So they can bring those cores into RDNA3.

One thing that I think DLSS is differently, is using a longer history of frame data.
This means it could access more temporal data to parse. This might help maintaining more detail, especially in motion.
But it also means it gets a higher probability of ghosting artifacts. nVidia is probably betting on it's AI training to be able to better weight what information to use, from each previous frame.

 

Yet DLSS still has its motion blur issues and its own nasty kind of ghosting. FSR 2.0 isn't getting enough praise for that.