PS6 Might Not Support PS5 Pro Games Due to Different AI Upscaling Architecture?!

LLMs are a different branch of deep learning and have nothing really to do with the Convolutional Neural Network U-net tensor solution which is what PSSR1.0 is.

DLSS2 - 3.6 and FSR4 are also CNNs, and DLSS4 is a transformer solution all with varying strengths and weaknesses and tradeoffs.
FSR4 is an CNN-Transformer hybrid
LLM are transformer too and it's kinda time aligned transformers went into both language and graphics tech
And models developed prior (dlss 2.0-3.6 and PSSR) stuck with CNN
 
I love how Nvidia marketing always makes people ignorant. The best image recognition models are precisely hybrid, because CNN have some advantages in local features extraction while transformer models exceed on capturing global context. If DLSS4 is better than FSR4 is simply because better training, more input processing, or bigger model, one of those reasons. Being full transformer is not the advantage.
 
A vanilla underpowered console right out of the oven to sell tens of millions of units and to make the pro look worthy for the masses. They expect many ps6 owners to switch to pro.
 
A vanilla underpowered console right out of the oven to sell tens of millions of units and to make the pro look worthy for the masses. They expect many ps6 owners to switch to pro.
What?
Pro will be seriously weaker in number of areas than PS6 and probably discontinued just before PS6 launch.
 
Justin Timberlake What GIF
I know right? Trivial for a trash tier programmer like me, but somehow it's in doubt if a billion dollar company can. It's a mystery...
 
FSR4 is an CNN-Transformer hybrid
LLM are transformer too and it's kinda time aligned transformers went into both language and graphics tech
And models developed prior (dlss 2.0-3.6 and PSSR) stuck with CNN
But LLMs aren't a signal processing problem solution, so it is a stretch IMO - even if they both operate at a holistic level.
 
But LLMs aren't a signal processing problem solution, so it is a stretch IMO - even if they both operate at a holistic level.
I am not saying they are the same
In my opinion LLM research brought transformers to light and graphics ML researchers took notion of it (they for sure look at progress/trends not only in their field but in ML in general) and start evaluating/testing it, leading to modern models.
So they are not the same but linked and rise of one led to progress in another.
 
Even if they use DLL modules for PSSR, the way developers have to patch themselves their games to the latest version of PSSR should mean the DLL is "shipped" with the game, right? So PS6 wouldn't be able to replace PSSR with FSR4 by a system level swap?
I think the ideal solution for PS6, which I assume is in what Sony an AMD are working on, it's to not only just to improve PSSR, but instead to move it to a OS/system level, integrated in how the console renders the visuals, so they can apply it to all games rendered there (including PS5 and PS4 games, including the PS1/PS2/PSP Classics), being able to boost resolution and framerate of any game ran there without needing any per game PSSR implementation in the game's code.

If they achieve this, they could have a portable underpowered hardware compare to base PS5, who would use this to run base PS5 games natively running at the same or lower resolution than the commonly played in PS4 or PS5 (let's say internally works like a PS5 in low power mode limited to 720p-for when in portable mode- or 1080p -for when in docked mode-) and then use this to bump to 1440p -for when in portable mode- or 4K -for when in docked mode- by doubling its width an height, plus pretty likely also bumping fps with an optional 60fps cap.

I assume this would allow them to run any -or almost all- PS4 and PS5 games at 1440p in a portable at 60fps, maybe even the one locked at 30fps, at least in my fantasy. Now it's just a fantasy, but I trust people like Cerny who make stuff that are almost a miracle.

Regarding PS6 games, which I asume all of them will be released for PS5 to run them in the portable and also to take advantage to people who will stay on PS5 years after the PS6 release, would be able to enjoy this without having to spend game resources or dev time on it. Meaning, they may natively do their games with a native 1080p resolution and lower framerate, to focus the horsepower on adding extra detail and more dense worlds, and then use this next-gen, OS level PSSR bump losslessly the resolution from that to 4K or even 8K and also bump their framerate.
 
I don't even understand why PSSR needed to exist. FSR 4 and framegen are more than good enough.

Hell, even the PS5 Pro didn't need to exist, they just did it to set an ongoing precedent of Pro models.
PS5 Pro is better hardware it's just PSSR sucks ass and devs love making stupid decisions which leads to the pro version running worse than the base PS5.
 
PS5 Pro is better hardware it's just PSSR sucks ass and devs love making stupid decisions which leads to the pro version running worse than the base PS5.
The whole point of the PS5 pro was to push PSSR so if PSSR sucks or don't work right the pro basically sucks lol.

Can't blame the devs when they are working with crap like the PSSR. Sony actually mentioned that they will be updating it next year.
 
I think the ideal solution for PS6, which I assume is in what Sony an AMD are working on, it's to not only just to improve PSSR, but instead to move it to a OS/system level, integrated in how the console renders the visuals, so they can apply it to all games rendered there (including PS5 and PS4 games, including the PS1/PS2/PSP Classics), being able to boost resolution and framerate of any game ran there without needing any per game PSSR implementation in the game's code.
You need motion vectors for implementing those algorithms. You can't apply automatically for PS4 or PS5 games.
 
Can we stop with this stupid bullshit hyperbole? When used properly PSSR it's far from shitty.
And TOP for TOP, millisecond for millisecond if it was on an even keeled hardware and able to hide a 4ms-6ms process for better accuracy motion vector calculations - technically not PSSR - and spend longer on the - technically not PSSR part of - motion vector reprojection too, like DLSS and FSR4 do, then it would be on par or superior.

We saw DF put a RTX 5090 with DLAA3.x with fast frame-rates that mitigate reprojection ghosting against PSSR with Rachet and Clank and it still wasn't superior in all situations and it was still outclassed in some aspects by PSSR - particularly detail at extreme distance IIRC - and DLAA claimed wins that related more to effective frame-rate than scaling.

Even FSR3 isn't garbage, and PSSR like DLSS3.x/FSR4 are way above that. It is like claiming MP3 at 120kbps is garbage when comparing to analogue radio.
 
Last edited:
I get PS4 games, but why not PS5 games?
Because the same issue applies, DLSS/FSR/PSSR/etc need motion vectors to be able to generate a coherent image. Although interestingly AFMF 3.0 is rumored to support driver-level motion vector extraction through methods like in-game motion blur and TAA, so that could allow for system-level FSR/PSSR injection, but still only for games that use motion vectors.
 
Because the same issue applies, DLSS/FSR/PSSR/etc need motion vectors to be able to generate a coherent image. Although interestingly AFMF 3.0 is rumored to support driver-level motion vector extraction through methods like in-game motion blur and TAA, so that could allow for system-level FSR/PSSR injection, but still only for games that use motion vectors.
So how does AutoSR work at the system level?
 
Because the same issue applies, DLSS/FSR/PSSR/etc need motion vectors to be able to generate a coherent image. Although interestingly AFMF 3.0 is rumored to support driver-level motion vector extraction through methods like in-game motion blur and TAA, so that could allow for system-level FSR/PSSR injection, but still only for games that use motion vectors.
Unless I'm mistaken, and it is a while since I've read much about the technical setup of UE5, but by default it has a scaler that results in all UE5 games needing to calculate standard MV AFAIK. I believe it was the reason DLSS was so easily patched into the Matrix demo on PC, and I suspect Unity does the same for 3D games, unless TAA wasn't as widespread as I thought from DF coverage of TAA in games following the PS4 Pro release last-gen
 
Last edited:
Unless I'm mistaken, and it is a while since I've read much about the technical setup of UE5, but by default it has a scaler that results in all UE5 games needing to calculate standard MV AFAIK. I believe it was the reason DLSS was so easily patched into the Matrix demo on PC, and I suspect Unity does the same for 3D games, unless TAA wasn't as widespread as I thought from DF coverage of TAA in games following the PS4 Pro release last-gen
No way a big company as Sony who need offer support is going to bypass every engine to inject motion vectors to PSSR 2.0 when they are not doing in their own games.
 
Last edited:
No way a big company as Sony who need offer support is going to bypass every engine to inject motion vectors to PSSR 2.0 when they are not doing in their own games.
The GPU in recent PlayStation gens provides all sort of heuristics AFAIK by default - from a video I watched on here a few years back about HDR tone mapping on PlayStation IIRC - so I'm pretty sure motion vector calculation is pretty trivial on PlayStation console hardware. But which of their own games are not using motion vectors and what is the source of that information to confirm that?

/edit

UE5 produces a Velocity buffer by default that it uses for its default Temporal Super Resolution(TSR) upscaler and is the info I was remembering, so at worst PS6 would just need to re-enable that and replace TSR on the many UE5 games. Checking Unity documentation and it too has default features for creating a motion vector texture and a temporal scaling.
 
Last edited:
This thread confuses me. I do not believe the PS6 will run a PS5 Pro game because they do not exist. I believe it will run a pS5 game enhanced with whatever version of upscaling the PS6 uses. If anything, the PS5 Pro was a way for Sony to get experience with PSSR working with PS5 games. Why would they not be able to take that knowledge and use it on the newer version of PSSR?
 
Last edited:
Although interestingly AFMF 3.0 is rumored to support driver-level motion vector extraction
This isn't a rumor - AFMF always worked that way.
Extracting motion vectors from history has been done by TV algorithms at least since mid 00s, and if you want a 'driver level' example because TVs don't count for some reason - Oculus Space warp did it in 2016 (albeit they get motion tracking data for free so it's half cheating).

However - motion alone is not enough for resolution reconstruction - you can do motion inter/extra-polation, but for resolution-reconstruction you need additional spatial samples otherwise you're just doing a spatial upscale with made-up data.
All TAA derivatives - ML or otherwise, (all the way back into late 00s where this whole thing started) achieve this by sub-pixel jittering of the camera. And that's data you won't have access to if you're operating only in post/driver - a PSSR/DLSS injection isn't really possible.

TLDR - take the Digital Foundry's favourite test - a static camera - there's (basically)no motion vectors to extract, and all pixels in history are in the same place. Meaning no temporal resolution reconstruction either.
 
Last edited:
Remember the FUD about "only select PS4 games will run on the PS5, it doesn't have BC" and "PS5 doesn't support raytracing XSX does!"?

I guess we're at it again in record time. Just wait 2 years instead of this early concern trolling.

And it'll not play re8 at 1080p and it'll run like shit and catch on fire and so much more. It's fucking ridiculous.
 
Not going to happen but wish it would. Not supporting previous PlayStations would probably be the best thing to happen to a PS6. Think of all the remastering we could justify then!!
 
So there's been some interesting talk from Kepler about the PS6's hardware and it might be bad news for anyone expecting full compatibility with the PS5 Pro





If this is true, it basically means Sony's next-gen console could have a totally different AI upscaling pipeline making the PS5 Pro's custom PSSR tech incompatible
That would be huge since Sony's been heavily marketing PSSR as a key feature for the Playstation brand

PS6 was never going to use the legacy PSSR.
 
I don't even understand why PSSR needed to exist. FSR 4 and framegen are more than good enough.

Hell, even the PS5 Pro didn't need to exist, they just did it to set an ongoing precedent of Pro models.

FSR4 is literally the result from what was learned with PSSR. Sony simply didn't have the time to implement FSR4 for launch of PS5 Pro. It will require some optimization to work properly on the Pro, hence why we won't see it until 2026.
 
Last edited:
This isn't a rumor - AFMF always worked that way.
Extracting motion vectors from history has been done by TV algorithms at least since mid 00s, and if you want a 'driver level' example because TVs don't count for some reason - Oculus Space warp did it in 2016 (albeit they get motion tracking data for free so it's half cheating).

However - motion alone is not enough for resolution reconstruction - you can do motion inter/extra-polation, but for resolution-reconstruction you need additional spatial samples otherwise you're just doing a spatial upscale with made-up data.
All TAA derivatives - ML or otherwise, (all the way back into late 00s where this whole thing started) achieve this by sub-pixel jittering of the camera. And that's data you won't have access to if you're operating only in post/driver - a PSSR/DLSS injection isn't really possible.

TLDR - take the Digital Foundry's favourite test - a static camera - there's (basically)no motion vectors to extract, and all pixels in history are in the same place. Meaning no temporal resolution reconstruction either.
You're confusing motion vectors with optical flow analysis, OFA is used by all forms of motion interpolation from DLSS FG, FSR FG, TVs Motion Smoothing, etc.

The rumor is AFMF 3 can directly access existing motion vector data from a game's TAA/motion blur implementation to improve frame-gen image quality.
 
You're confusing motion vectors with optical flow analysis, OFA is used by all forms of motion interpolation from DLSS FG, FSR FG, TVs Motion Smoothing, etc.
The rumor is AFMF 3 can directly access existing motion vector data from a game's TAA/motion blur implementation to improve frame-gen image quality.
Ah direct access to game motion buffers is a different story of course (and that one I guarantee would be much harder on consoles).
But again - that would work for frame-gen, but still wouldn't help for reconstruction in-driver (for that you'd need pre-resolve color buffers, rely on the fact game already has camera-jitter inserted, and then need info on the jitter offsets to correctly interpret the samples).
 
Last edited:
I thought Alan Wake 2 had a bad implementation of PSSR. Did they fix that?
That 40fps mode full blast gfx is fine.
Pssr got semi fixed.

But this game looks tits 😍
Also forgot to add that callisto game also looks fucking insane on ps5 pro
 
Top Bottom