• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

Microsoft Auto Super Resolution tested: a new AI upscaler has arrived

Topher

Identifies as young
The gaming highlight of the new Windows on ARM Copilot+ laptops.


The less said about gaming on the new wave of ARM-based Windows laptops, the better. Qualcomm's Snapdragon X Elite has severe driver issues resulting in generally poor performance and while the PRISM CPU translation layer allowing x86 code to run on ARM looks impressive, no support for AVX instructions means many games won't even load. Let's be generous and say that gaming support on 'Windows on ARM' hardware is a work-in-progress - but there's an element I feel is well worth examining. Microsoft's Auto Super Resolution upscaling technology has its limitations, but it works.

In robustly embracing the AI era, Microsoft has mandated that new, AI-focused Copilot+ PCs should contain an integrated neural processor rated for at least 40 TOPs (tensor operations). New processors like the Snapdragon X Elite and AMD's Strix Point line contain the requisite hardware and while we can expect gaming not to touch the NPU for a good while yet, Microsoft's Auto SR can be invoked at will on practically any game. Right now, there's a curated 'white list' of games that support Auto SR, but as far as I can tell, there's nothing stopping you from simply adding the feature yourself within the Windows control panel. Right now, Microsoft is only talking about Snapdragon support for Auto SR, but one would hope that the feature will roll out for all Copilot+ PCs in due course.

While both DLSS and Auto SR aim to take lower base resolution imagery and upscale them to higher resolutions, the way this is achieved is very, very different. DLSS and rivals like FSR2/3 and XeSS have a wealth of data to work from: the base image, motion vectors, colour and depth information and more. They're also integrated into the game engine pipeline, meaning that some aspects of the image - such as post-processing or HUD elements - can be processed at native resolution. Auto SR seemingly only has access to the base image, effectively making it a screen-space upscaler.

Microsoft puts it like this: "Auto SR utilises a large AI-based super resolution model running on the Snapdragon Series X Neural Processing Unit (NPU) to restore the visual detail lost by rendering the game at a lower resolution to increase frame- rate.

"The model is a convolutional neural network (CNN) that provides spatial upscaling between two fixed resolutions. It has been trained to add detail and perform anti-aliasing while upscaling. The high resolution output from the model is set to optimise image quality for the built-in displays of Copilot+ PCs equipped with the Snapdragon Series X processor, and dictates the time required to execute the model. Applying Auto SR adds latency to the game equivalent to the time it takes to execute the model, which is currently 12ms."

By shuttling the upscaling work to the NPU, there is no perceivable impact to the game's frame-rate - however, by adding an additional step with the upscaling, latency is added. If you're gaming with v-sync on a standard 60Hz display, you're essentially looking at a single frame of latency added - which I think is perfectly fine.

More noticeable is the fact that Auto SR only has the lower resolution image to work from. HUD elements can exhibit some artefacting while upscaled text in general (on splash screens and the like) looks OK, but clearly not quite right. The other key limitation Auto SR has which is that as it is effectively a post-process, it has no sense of 'history'. It doesn't seem to refer to previous frames for its anti-aliasing in the way that TAA does. That being the case, it's important to 'feed' Auto SR with the most temporally stable image that you can. The idea of running a 720p image with TAA isn't generally a good idea as it introduces a lot of blur. However, Auto SR does a much better job in increasing detail than it does in dealing with the shimmering artefacts - so using TAA makes a lot more sense. For the same reason, intrusive post-process effects like film grain and chromatic aberration should be disabled (motion blur, however, seemed fine to me).

In looking at Auto SR's upscaling prowess, I went back to Microsoft's blog that introduced the technology and sought to replicate its results on Borderlands 3, which were based on 720p imagery upscaled to 1440p with clearly noticeable quality improvements. Initially, I captured the benchmark run at native 720p, then with an Auto SR upscaling pass captured at 1080p, 1440p and 2160p. It turns out that Auto SR only upscales to one resolution, so any differences between the captured resolutions will be entirely down to the GPU scaler. From that point onwards, I moved to 2160p Auto SR captures - as 4K is the final output resolution for the YouTube video.

Borderlands 3 may have been chosen as a 'best case scenario' by Microsoft from its supported titles owing to a bright, vibrant, distinct art style, but as you can see from the images on this page, there is no doubt whatsoever that the AI model can deliver substantially improved detail that seemingly did not exist before. It even does a reasonable job on tricky elements like grass, which look significantly cleaner. In my tests, I found that bright, colourful games like Borderlands 3, The Witcher 3 and Baldur's Gate 3 produce the best results.

Control was my test: again, I could see more detail in the upscaled output, but at the same time, as Remedy's TAA isn't particularly impressive, sub-pixel break-up and edge aliasing still remained obvious in the upscaled output. Additionally, film grain needed to be disabled as this could produce some weird results. Even so, while capturing the game at 4K, I also fed the output from the capture card into a 27-inch 4K screen. I can tell that it's obviously not a 4K image but the point is that it still looks noticeably better than standard bilinear or lanczos upscaling from the base 720p.

And while Microsoft is busy validating games that'll automatically turn on Auto SR, its status as a post-process gives it a flexibility that other upscalers do not. Cyberpunk 2077 is not a validated game, but I could enable it easily enough and it worked just fine. Again, it's no competition for native 1440p or 4K, but it presents in a pleasing enough manner to enable it as opposed to not enabling it. Bearing in mind that Copilot+ PCs will almost certainly come with relatively meagre integrated graphics solutions, it's a good value added feature.

The flexibility of Auto SR in running on everything also has some interesting possibilities. There's nothing stopping you using Auto SR upscaling on top of a game's existing upscaling features. First of all, I tried God of War, finding that while Auto SR-upscaled FSR 2 wasn't particularly attractive (effectively upscaling from 480p to 720p then to whatever output Auto SR uses), using the game's inbuilt TAA upscaler wasn't a disaster. Meanwhile on The Witcher 3, I actually found that using Intel XeSS at ultra quality mode upscaling to 720p produced the most temporally consistent image available and from there, Auto SR did the business. Upscaling on upscaling sounds like a disaster in the making, but ultimately, the flexibility is useful when GPU resources are limited.

You may have noticed from the video that while games like Control run well on the Snapdragon X Elite (it's mostly 60fps at low settings at 720p), titles like Cyberpunk don't run well at all. None of this is down to Auto SR, but rather the Snapdragon X Elite GPU, where even the Steam Deck GPU wipes the floor with it in the vast majority of games I tested. However, if the Copilot+ feature set is extended to, say, AMD Strix Point processors (which do have a 'qualifying' NPU), the utility of Auto SR will be much more pronounced - many more playable games means many more reasons to use Auto SR.

Of course, right now, it remains to be seen whether Auto SR will deploy on anything other than Snapdragon processors, but the possibilities are still intriguing - I ran my Auto SR captures on the AyaNeo Kun Windows handheld to get a feel for what it may look like. There's a definite increase in clarity, so may find some use, but the bottom line is that basic 720p still looks good enough - I guess it would come down to the power draw of the NPU, and whether the additional latency is noticeable.

So, Auto SR is no DLSS or XeSS, but it certainly has its uses and when it works well, the quality boost is appreciable. That said, it's safe to say that the quality increase is variable (Doom Eternal didn't seem to do much at all), and its absolutely essential that the quality of the input image is reasonable. Dark Souls 3 is supported out of the box, but its anti-aliasing is very, very poor and Auto SR can't seem to do much with it. Still, it's an intriguing start for the technology and I'll be following its progress with much interest.




Haven't seen this posted yet so...
 

Killer8

Member
It looks like shit. Distant detail resolve is so bad that it turns the art style of everything it touches into a watercolor painting:

JfIVzKr.jpeg
 

mrqs

Member
It doesn't "turn" anything into a watercolor painting. Look at the detail compared to the native 720p image it's upscaling from. it restores a ton of detail, just not as effectively as DLSS does.

Exactly. It has some weirdness to it, but in due time most of the gaming we all do will be on massively upscaled images and with most frames being generated. Glad Microsoft is working on this.
 

Killer8

Member
It doesn't "turn" anything into a watercolor painting. Look at the detail compared to the native 720p image it's upscaling from. it restores a ton of detail, just not as effectively as DLSS does.

AI upscalers do not 'restore' detail, they use AI models to best guess what detail might have been there if the image was originally rendered at a higher resolution.

A good upscaler will interpolate a lower resolution image in such a way that the final image that gets resolved will look perceptually close to the ground truth i.e. it looks like what it would've looked like if it was higher res to begin with.

This is an example from a research paper comparing two algorithms, SRGAN and ESRGAN, against the ground truth on the right to judge which algorithm produces the closer result:

trbrotQ.png


It's clear from DF's comparisons that Auto SR does a really shit job at resolving detail, creating a very smudged and artificial image. Describing that artificiality of the image as watercolor looking is nothing new.

It's also really poor of DF to not include that native 1440p / 4K ground truth to compare all this with. How can you judge how well an algorithm is interpolating a lower res image to a higher res image... if we don't have the higher res image to compare it to?

Their other article about XeSS had a native comparison.
 

JimboJones

Member
Auto SR only activates at low resolutions so 720p and 900p I think they said. It's a comparison between a raw low res image output vs the AutoSR pass. For a post process it's producing a pretty decent image. I imagine a lot of people would appreciate it if they were gaming on something lower spec and this was their only option vs a raw 720p output.
 

RagnarokIV

Battlebus imprisoning me \m/ >.< \m/
games like Control run well on the Snapdragon X Elite (it's mostly 60fps at low settings at 720p)

I wish this bullshit would end. I’ve seen it on places like Reddit, Mac rumors etc.
Especially bigging up the Mac with shite like “finally AAA gaming on Mac” then you see Cyberpunk at 35fps medium settings 1080p on M2 Pro.

It’s certainly running, just not well.

Why bother to play games at such shit quality?
 

FireFly

Member
AI upscalers do not 'restore' detail, they use AI models to best guess what detail might have been there if the image was originally rendered at a higher resolution.

A good upscaler will interpolate a lower resolution image in such a way that the final image that gets resolved will look perceptually close to the ground truth i.e. it looks like what it would've looked like if it was higher res to begin with.

This is an example from a research paper comparing two algorithms, SRGAN and ESRGAN, against the ground truth on the right to judge which algorithm produces the closer result:

trbrotQ.png


It's clear from DF's comparisons that Auto SR does a really shit job at resolving detail, creating a very smudged and artificial image. Describing that artificiality of the image as watercolor looking is nothing new.

It's also really poor of DF to not include that native 1440p / 4K ground truth to compare all this with. How can you judge how well an algorithm is interpolating a lower res image to a higher res image... if we don't have the higher res image to compare it to?

Their other article about XeSS had a native comparison.
This feature is designed for mobile devices where you wouldn't necessarily have enough power to render the higher resolution image in the first place. I don't see it as a DLSS/XeSS replacement, where the aim is to claw performance back to allow for higher quality settings.

And if the choice is between a blurry mess at 720p and a slightly painterly image with much more perceptual detail, the latter may be preferrable depending on the game and the person's preferences.
 
Last edited:
The article is speculating based on that one vague comment made during that one pre-recorded podcast, along with the “adorably all digital” documents. Got anything else?

We know. We watched the podcast back then. We’ve read the leaks. This is also from March, which is ancient history in Xbox terms.

A full, stand-alone Xbox console next generation is going to remain a dream.
 

winjer

Member
Considering this is basically an spatial upscaler with AI, the results are good.
But still, very far from what a temporal solution can do. So if a game supports DLSS, XeSS, TSR o FSR3, it's better to just use these solutions.

The worst part is that it ads quite a bit of extra latency. And it's limited to upscaling from 720 to 900p, probably because the current NPUs can't process more pixels than that in real time.
 

nemiroff

Gold Member
Perfect for gamepass aye
That doesn't make any sense. It's ARM only and intended to sort of cater for the low end of graphics processing. It's kinda like Lossless Scaling, but potentially better because titles who support it natively can feed it with more information for better results.
 
Last edited:

BlackTron

Member
I wish this bullshit would end. I’ve seen it on places like Reddit, Mac rumors etc.
Especially bigging up the Mac with shite like “finally AAA gaming on Mac” then you see Cyberpunk at 35fps medium settings 1080p on M2 Pro.

It’s certainly running, just not well.

Why bother to play games at such shit quality?

I guess if you own a Mac it's better than nothing.

ydbbrkU.jpeg
 

RoboFu

One of the green rats
I wish this bullshit would end. I’ve seen it on places like Reddit, Mac rumors etc.
Especially bigging up the Mac with shite like “finally AAA gaming on Mac” then you see Cyberpunk at 35fps medium settings 1080p on M2 Pro.

It’s certainly running, just not well.

Why bother to play games at such shit quality?

That particular game is running through a translation layer which is why it is impressive.
 
Last edited:

analog_future

Resident Crybaby
AI upscalers do not 'restore' detail, they use AI models to best guess what detail might have been there if the image was originally rendered at a higher resolution.

A good upscaler will interpolate a lower resolution image in such a way that the final image that gets resolved will look perceptually close to the ground truth i.e. it looks like what it would've looked like if it was higher res to begin with.

This is an example from a research paper comparing two algorithms, SRGAN and ESRGAN, against the ground truth on the right to judge which algorithm produces the closer result:

trbrotQ.png


It's clear from DF's comparisons that Auto SR does a really shit job at resolving detail, creating a very smudged and artificial image. Describing that artificiality of the image as watercolor looking is nothing new.

It's also really poor of DF to not include that native 1440p / 4K ground truth to compare all this with. How can you judge how well an algorithm is interpolating a lower res image to a higher res image... if we don't have the higher res image to compare it to?

Their other article about XeSS had a native comparison.

I know how AI upscaling works. For all intents and purposes to the end user it restores detail, hence my phrasing.

Again, it's not as effective as DLSS but it clearly adds a ton of perceived detail compared to an actual native 720p image. To suggest it's making that 720p image worse is asinine.
 
Last edited:
I wish this bullshit would end. I’ve seen it on places like Reddit, Mac rumors etc.
Especially bigging up the Mac with shite like “finally AAA gaming on Mac” then you see Cyberpunk at 35fps medium settings 1080p on M2 Pro.

It’s certainly running, just not well.

Why bother to play games at such shit quality?
That's not even shitty quality, 1080 medium with a lock 30? Thats good experience.
 

JackMcGunns

Member
The article is speculating based on that one vague comment made during that one pre-recorded podcast, along with the “adorably all digital” documents. Got anything else?

We know. We watched the podcast back then. We’ve read the leaks. This is also from March, which is ancient history in Xbox terms.

A full, stand-alone Xbox console next generation is going to remain a dream.


And you're not speculating based on...?

Explain.
 

mystech

Member
Worth noting that we're likely seeing an early version of what Microsoft is intending for their next-gen Xbox hardware. Discrete NPUs may be the next big change in gaming hardware.
I was thinking the same thing! Trying to run games natively at high resolutions has been a big challenge this generation and in many ways held this generation back. It seems like the next gen will be all about running at lower resolution internally to give them extra headroom to push things further. In other words, next gen will be all about working smarter not harder. My bet is that the PS6 launch unit will be smaller than PS5 “Slim”.

PS5 Pro will give us a little taste of Sonys new tech. Nintendo has a slight edge by teaming up with Nvidia.
 

analog_future

Resident Crybaby
I was thinking the same thing! Trying to run games natively at high resolutions has been a big challenge this generation and in many ways held this generation back. It seems like the next gen will be all about running at lower resolution internally to give them extra headroom to push things further. In other words, next gen will be all about working smarter not harder. My bet is that the PS6 launch unit will be smaller than PS5 “Slim”.

PS5 Pro will give us a little taste of Sonys new tech. Nintendo has a slight edge by teaming up with Nvidia.

This generation is already about running at a lower resolution internally. Almost zero AAA games this gen are actually running native resolutions on console.
 
Top Bottom