NVIDIA DLSS 4 Transformer

[TheinsanegamerN]

The problem with MSAA is it only works on polygons, not texture or shader aliasing. It also doesn’t work with VRS, or anything that’s not update

Would you rather have no AA on textures or shading, or TAA producing visual fidelity on par with beer goggles?

Well then whole current rendering pipelines need to be changed to the old ways.

Also I would love to see how that MSAA would look on those Stalker 2 screenshots with all the alpha vegetation. But tbh some people would still claim it looks better.

You can make the next mona lisa, but when you rub vaseline all over it and call it TAA, it makes it rather difficult to appreciate the art underneath.

I mean how many videos are out there comparing modern TAA games to older games without TAA, showing a remarkable regression in visual quality?

 

[Visible Noise]

The hybrid of MSAA + SMAA works well. MSAA for geometry edges, post-process AA for textures.
TAA reduces image quality, and relying on upscaling under these conditions is pointless, as it cannot restore lost details—especially since upscaling relies on the temporal filter anyway

To me SMAA is a smear fest. I’m glad it’s obsolete, along with its post process brethren FXAA and the like.

Would you rather have no AA on textures or shading, or TAA producing visual fidelity on par with beer goggles?

You can make the next mona lisa, but when you rub vaseline all over it and call it TAA, it makes it rather difficult to appreciate the art underneath.

I mean how many videos are out there comparing modern TAA games to older games without TAA, showing a remarkable regression in visual quality?

Bad implementation of TAA doesn’t mean all TAA is bad. Epic’s is quite good in recent UE, and can get pretty darn close to DLSS.

 

[remekra]

Would you rather have no AA on textures or shading, or TAA producing visual fidelity on par with beer goggles?

You can make the next mona lisa, but when you rub vaseline all over it and call it TAA, it makes it rather difficult to appreciate the art underneath.

I mean how many videos are out there comparing modern TAA games to older games without TAA, showing a remarkable regression in visual quality?

Well it would be hard to enjoy Mona Lisa if you would take a knife and make sharp cuts around every edge or line. And those sharp cuts get worse when you try to look from a different angle.

You don't like temporal solutions that's fair I guess. Problem is it doesn't bother most people, some like it for edge stability when in motion and overall softer look.
DLSS/FSR have added bonus of boosting performance when upscaling.
And no studio will go and implement MSAA with SMAA or any other post process effect, which would have worse performance and quality of it could be questioned.

 

[Luminescent]

In the meantime, some new Chinese AI model ( deepseek ) that is also open source and free, beat or equaled chatgpt, Nvidia stock and others involved are going down and investors are wondering why they payed billions when the Chinese allegedly did it with millions of dollars, also, open source, anyone can take it.
So, that 2000$ gpu might get a discount soon.

 

[john_]

So, that 2000$ gpu might get a discount soon.

No, it will not. Nvidia dictates pricing and as long as they have the upper hand in the market and the most important of all, favorable coverage, describing their techs as impressive, until the next major version of course, not to mention every other equivalent competing tech describing it over and over again close to garbage, people will still be driven to Nvidia's graphics cards, no matter the prices they put on those graphics cards.

 

[Heiro78]

works for me. could you screenshot where you are looking?

You'll have to excuse my hallucinations..... Kidding, but I dunno how I missed it. I used this one instead.

Was the benchmark done while standing still somewhere? How does the 4090 lose 16% performance and the all-powerful 3060 lose 4%?

View attachment 381945

It also looks strange for TPU to reuse third-party videos.

Cyberpunk 2077 RT Overdrive - Ray Tracing vs Path Tracing On vs Off Comparison | RTX 4080 4K DLSS

What third party video did TPU use? The one in the article is their own

Until yesterday DLSS 3.8 was impressive. Today we learn that DLSS 4.0 is impressive while DLSS 3.8 is not so impressive after all.

Waiting for DLSS 5.0 review in a couple of years vs DLSS 4.x, to read the same, that DLSS 4.x isn't as much impressive as we will be reading for the next couple of years, but DLSS 5.0 is really impressive.

My biggest takeaway from the sentence is the praise towards bring these software features to older cards.

 

[Botan626]

Yes, and you can select the comparison image by clicking on the description. It’s quite well done.

Yes what? I asked if it's left or right, when vertical line is on the right edge, and you're telling me yes.

 

[john_]

My biggest takeaway from the sentence is the praise towards bring these software features to older cards.

Then again AMD's FSR that is available to a number of GPUs that suppose to not support DLSS and AMD's AFMF 2 that offers the option of upscaling and frame generation to almost every game out there, is constantly described as unusable and garbage. I don't see praises there. Maybe I missed them? While FSR might be (way) inferior, wording makes all the difference. When one is garbage and the other is impressive (until it's next version), we all know what will happen in the market.

 

[Heiro78]

Then again AMD's FSR that is available to a number of GPUs that suppose to not support DLSS and AMD's AFMF 2 that offers the option of upscaling and frame generation to almost every game out there, is constantly described as unusable and garbage. I don't see praises there. Maybe I missed them? While FSR might be (way) inferior, wording makes all the difference. When one is garbage and the other is impressive (until it's next version), we all know what will happen in the market.

I agree and applaud AMDs efforts to make FSR usable on any video card. I'm also a native slob and will continue to advocate for that. I also just see the writing on the wall that these DLSS, FSR, XESS offerings are going to be the norm

 

[Steevo]

I wonder what the difference in games that had Super Sampling AA is compared to all the post processing effects. Maybe I will spend some time tonight trying FSR and other things VS old school SSAA which is what DLSS was supposed to be, the next iteration of filters.

 

[InVasMani]

Defiantly improved image quality relative to CNN. The performance is roughly the same a hinter lower, but a worth while improvement in quality. A step in the right direction.

 

[truerock]

Sometimes I see an improvement - sometimes I don't see an improvement. Seems kind of random

The Nvidia GeForce RTX 4090 can process 84 billion triangles per second - so I'm not sure why there is a need for upscaling. Well, obviously the games are not generating high-resolution output? They could - but, they don't?

Is it to save disk space? Is the limitation on I/O between primary memory and the SSD drive? Is the bandwidth between primary memory and GDDR memory the bottle neck?

It is not the ability of Nvidia graphics cards to process triangles. This compression and/or upscaling seems unnecessary to me.

 

[Visible Noise]

Yes what? I asked if it's left or right, when vertical line is on the right edge, and you're telling me yes.

Sorry, I misunderstood your question. The descriptions are over the section they are describing. So if the line is all the way to the right, you are looking at the entire left side.

 

[Calmmo]

Hard to tell in static images, far more obvious in movement.

 

[Apart]

what is up with that weird black shadow on the palm trees?

It is a known bug with RT Lighting and PT. It's being reported by the Cyberpunk 2077 community since 1.6, but they haven't bothered to fix it.

 

[AusWolf]

What's CNN? What's transformer? Can somebody shed light on some details? I feel like I've just read some marketing material, not a review, to be honest.

 

[Raffles]

Sometimes I see an improvement - sometimes I don't see an improvement. Seems kind of random

The Nvidia GeForce RTX 4090 can process 84 billion triangles per second - so I'm not sure why there is a need for upscaling. Well, obviously the games are not generating high-resolution output? They could - but, they don't?

Is it to save disk space? Is the limitation on I/O between primary memory and the SSD drive? Is the bandwidth between primary memory and GDDR memory the bottle neck?

It is not the ability of Nvidia graphics cards to process triangles. This compression and/or upscaling seems unnecessary to me.

DLSS works on pixels, not triangles. The triangle count is the same obviously, regardless of resolution,

In modern graphics workflows with complex shaders, stacks of post processing, and now raytracing - pixels are far more expensive than triangles. That's where DLSS comes in, to alleviate the high per pixel cost of modern graphics.

 

[TheinsanegamerN]

Well it would be hard to enjoy Mona Lisa if you would take a knife and make sharp cuts around every edge or line. And those sharp cuts get worse when you try to look from a different angle.

Oh, you mean like how TAA makes something like those power lines render incorrectly or not at all?

You got the point, stop torturing the metaphor.

You don't like temporal solutions that's fair I guess. Problem is it doesn't bother most people, some like it for edge stability when in motion and overall softer look.
DLSS/FSR have added bonus of boosting performance when upscaling.

This is total cap, given the number of people on social media, forums, and youtube COMPLAINING about how bad TAA looks and how image quality has gone backwards.

And no studio will go and implement MSAA with SMAA or any other post process effect, which would have worse performance and quality of it could be questioned.

Also total cap, TAA is not universal, plenty of games come out not using TAA. It's not a mandatory feature.

Then again AMD's FSR that is available to a number of GPUs that suppose to not support DLSS and AMD's AFMF 2 that offers the option of upscaling and frame generation to almost every game out there, is constantly described as unusable and garbage. I don't see praises there. Maybe I missed them? While FSR might be (way) inferior, wording makes all the difference. When one is garbage and the other is impressive (until it's next version), we all know what will happen in the market.

AMD's FSR is frequently praised for being open, and criticized for working like dog poo. Both things can be correct.

BTW, the existence of DLSS4 does not mean that 3.8 was total crap. Just like the 4090 didnt meant eh 3090 was unusable anymore. They're these things called improvements. Not everything is a zero sum game where there is a winner and a loser. Sometimes new things can come out that are just....better.

It's great that previous FSRs were open source, and worked on multiple vendors. However, FSR has serious issues with image quality. The same issues affected DLSS 1 and 2, the difference being nvidia has significantly improved DLSS over the years while FSR has seemingly stalled out.

 

[InVasMani]

What's CNN? What's transformer? Can somebody shed light on some details? I feel like I've just read some marketing material, not a review, to be honest.

CNN = Convolutional neural network while a transformer is a way to process sequential data to encode and train it rapidly. It's basically similar to like fragmenting a hard drive in a sense and compressing data and in doing so being able to fit more data in the space saved. Something like that I'm not a expert in this field so someone else experience could probably better describe it in fact just ask the AI if you really want to know the nitty gritty better.

 

[Chrispy_]

I've been running the CP2077 patch with DLSS4 since the weekend and DLSS is still too soft for me, but I can't argue with the improvements, they're great. Honestly, it's the in-motion temporal smear that's been improved the most, and that's huge because temporal smear made prior DLSS a lie - great in static screenshots, godawful in motion and real gameplay.

If I need to upscale then DLLS Balanced or even Performance looks no worse in motion than DLSS Quality used to look with the CNN model. Also, the ghosting behind cars and katana edges is gone, which was easily one of the most distracting and janky artifacts that put me off DLSS before. It was utterly impossible not to see it because it looked so wrong.

Path tracing is usable on lower-end hardware now, because DLSS Balanced gives me enough base framerate to hit >100fps with framegen. It's still brutally expensive, but pathtracing with acceptable image quality at 1080p120 on a 150W GPU is kinda cool.

So yeah, is DLSS4 perfect? No.
Is it good enough that I'll use it more often? Heck yeah.

 

[tfdsaf]

Both look bad, but the new model looks slightly less bad! Why aren't we gaming at 8k resolution in 2025 and even getting rid of AA and upscaling/downscaling? Why are we playing at 720p upscaled at higher resolution with fake frames in 2025? What kind of bullshit is this?

 

[nguyen]

Nvidia chief of AI application said the CNN model just cannot be improved upon anymore, meanwhile the Transformer model is just beginning.

Transformer already exceed the CNN model currently, but there are still some flaws so there definitely are ways to improve the Transformer algorithm further.

Let see Nvidia improve the Ultra Performance mode further with Transformer model, right now it still looks bad at 4K

 

[W1zzard]

You'll have to excuse my hallucinations..... Kidding, but I dunno how I missed it. I used this one instead.

Where did you look initially? Maybe everyone expects some button to be in that place, but I never thought of it.

 

[ViperXTR]

Here's someone with RTX 2060 playing Tokyo Xtreme Racer, left is CNN right is transformer

Even when not in motion, the difference is noticeable, left is CNN right is transformer, you can clearly see pattern on the tail lights

 

[Legacy-ZA]

A famous quote from a TV series "Welcome to the death of the Age of Reason"
How can let's say a 10 megapixels picture be downscaled to 4 megapixels and then upscaled to 10 megapixels and look better than the original ?
Unless, and this sounds crazy, the original doesn't have contrast, sharpness, color and is somewhat gimped so the upscaled image can look better.

I most certainly understand what you are conveying, however... I see it more as a structure that was built, but you can make improvements, not to only make it look better, but easier to navigate in or around.