Any details on AMD Leo demo?

Status
Not open for further replies.
well i just tried 8x supersampling 480 upscaled 2x 1280x960 and it looks good
but 1650x1080 4xmsaa looks just as good as far as I can tell

ps: 640x480 to 1920x1080 is a much bigger scale it also isnt a perfect multiple

1920 is 1080p and in the case of the dvd disc upscaled in real time by a bluray player it looks quite nice on my tv.
from what distance ? you are aware dvd uses lossy compression games are very different

That's not a multiple of 640x480 unlike 1080p, upscaling to odd resolutions is more troublesome and may introduce additional artifacts.
Where did you learn maths ? 1080/480 = 2.25

And the absence of fine detail at 1080p will be compensated by PERFECT prerendered-CG like IMAGE quality.
PERFECT prerendered-CG like IMAGE quality. with the absence of fine detail ? That makes it imperfect...

4x is good, but since we're dealing with 480p, higher levels of supersampling should be feasible in realtime..
So you want to render at 5120x3840 downsample it to 640x480 and scale it up to 1920x1080. have you met 2010 IQ is unaceptable
 
Do you think that would be better or worse than 2.1875?

?
Where did you learn maths ? 1080/480 = 2.25

1920x1080=2073600

640x480=307,200

x3= 921,600.

1280x720=921,600

I apparently did make a mistake on my calculation earlier and used 720p, instead of 1080p:oops:

Though it is true that the aspect ratio changes, iirc, a more optimal lower res that preserves aspect ratio should allow for even better upscaling.
PERFECT prerendered-CG like IMAGE quality. with the absence of fine detail ? That makes it imperfect...

The question is could a reasonable level of supersampling eliminate all iq problems at a low resolution?

Absence of fine detail? We'll I didn't say absence, fine detail is diminished but it isn't completely eliminated from the image, again look at the provided link showing what's possible. Quite a lot of detail there.;)

The point is assume we finally do something like avatar-like graphics in realtime, if it's all a shimmery aliased mess in some areas, it won't be pretty. While if it looks like an upscaled Avatar dvd, with no shimmer or aliasing anywhere in sight and plenty of detail, it will look superb.

For example look at the following image upscaled from 480. PLENTY of Detail
IMAGE LINK

that image is zoomed in also, iirc, merely to be able to tell the difference.
 
640x480=307,200

x3= 921,600.

1280x720=921,600
Just a small note that 640/480 = 1.3(3) and 1280/720 = 1.7(7). Sure, they have a nice amount of pixel difference but scaling with different aspect ratio won't be easy.

Also I still don't get it why do you want to downsample and upscale when you could just apply any of the modern post-process AA or smart blur filters as I like to call them that provide even better quality at considerable smaller loss of details.
 
steampowered just because upscaling works ok for films does not mean its ok for games
if you want to provide examples use a game, you will see upscaling hurts iq
 
Just a small note that 640/480 = 1.3(3) and 1280/720 = 1.7(7). Sure, they have a nice amount of pixel difference but scaling with different aspect ratio won't be easy.

Also I still don't get it why do you want to downsample and upscale when you could just apply any of the modern post-process AA or smart blur filters as I like to call them that provide even better quality at considerable smaller loss of details.

Nonetheless my dvds look pretty much perfect at 720 and 1080 upscaling. But it is true aspect ratio presents some issues, which is why I mentioned the aspect ratio issue in my last post and that a different resolution could be used for better results.

The point was that modern AA methods seem to leave plenty to be desired in terms of some aspects of image quality. If there are AA methods that will deliver pixar-like aliasing and shimmering free image at reasonable computational costs at 1080p then by all means.

The only reason why I'm mentioning sub720p images is because of the fact that most game images attain CG-like or Movie like image quality in such pictures, and if similar can be accomplished in motion, then a slight loss in detail would be an acceptable compromise.

steampowered just because upscaling works ok for films does not mean its ok for games
if you want to provide examples use a game, you will see upscaling hurts iq

Assuming there are no image quality artifacts in the original 480p, what precisely would make the upscaling any different from upscaling a prerendered cg 480p movie, in terms of image quality? I presume we can all agree that while not ideal the result of upscaling prerendered cg with good image quality from 480p to 1080p is still of acceptable quality.
 
Nonetheless my dvds look pretty much perfect at 720 and 1080 upscaling.

Well, yes you take an already compressed source and upscale it with a decent upscaler and you can get something that looks like the original source.

That's not the same as taking a 1080p source downsampling to a 480p target and then upscaling it.

Lets take your DVD for example. Assuming there's also a BRD available of that movie. If you upscale that DVD source it will never look as good as source material from a BRD at 720p or 1080p. Granted if you sit far enough away from the TV any differences might not be perceivable but they are there. 480p = 480p upscale < 720p = 720p upscale < 1080p.

Likewise if you take a 1080p source and downsample it to 480p (I have friends that do this so they can play those movies on their portable players) you'll generally have a better looking movie than the same movie from a DVD. Assuming you have a good downsampler.

However, once you upscale that 480p movie from above on a 1080p set it's quite noticeable that it isn't the same quality as the 1080p source it originated from.

Now lets go one step further and assume we're doing that with a computer generated image like you would get from a game or from the Leo Demo or whatever.

Assuming a good downsampler and a very good upscaler, you might be able to reduce aliasing. But in the process you're also going to lose a lot of detail. At best you'd end up with an upscaled image that is blurrier and lacks the detail of the original source that may or may not have most of the aliasing removed (due to the blurring of the entire scene).

So long story short.

[1] - A 480p source upscale will never look better than the original 480p source. Upscalers aren't magic and can only attempt to faithfully reproduce the details available in a 480p source.

[2] - A 1080p source contains more information and is thus more detailed than a 480p source. Hence it is also going to be more detailed than an upscaled 480p source.

[3] - Downsampling a 1080p source to 480p will always result in a loss of detail. Goto [1]

Regards,
SB
 
While it's true that an upscaled image will not look as a good as one evaluated in the higher resolution, it's not an interesting question.

The interesting question is this: given a fixed performance budget, is it better-spent on rendering in some high resolution with fewer samples and FLOPS/pixel or in a lower resolution with a potentially better sampling pattern and more FLOPS, then upsampling the result with a good filter?

Obviously both extremes of this question are not interesting, but the question is at what pixel density (relative to viewing distance) does it cross over to the point where it's a poor use of resources to shade every pixel? The other interesting variable is that for AA purposing scaling with a uniform grid (i.e. high resolution) does not look as good as the same number of samples on a non-uniform grid (MSAA, or similar), which will also cross over at some point.

Thus I propose that shading at 300dpi (to pick a number) and normal viewing distances is a poor use of rendering resources, and instead you'll end up with a better image by using well-placed samples and better shading at a lower resolution and then upscaling it. Where the crossover happens in terms of density is hard to say and probably content-dependent though.
 
steampowered rather than asking, why not run a game at 1080 with 4xssaa benchmark it take a screenshot
run the game at 640x480 with 4xssaa make sure its running fullscreen on your 1080 monitor
benchmark it see what gain in fps you have and if the increase it would allow for 8xssaa again take a screenshot (this may be difficult as the screenshot may be 480)
ps: keep in mind as you increase ssaa you run the risk of running out of vram this will make running at higher ssaa not practical
 
While it's true that an upscaled image will not look as a good as one evaluated in the higher resolution, it's not an interesting question.

The interesting question is this: given a fixed performance budget, is it better-spent on rendering in some high resolution with fewer samples and FLOPS/pixel or in a lower resolution with a potentially better sampling pattern and more FLOPS, then upsampling the result with a good filter?

Obviously both extremes of this question are not interesting, but the question is at what pixel density (relative to viewing distance) does it cross over to the point where it's a poor use of resources to shade every pixel? The other interesting variable is that for AA purposing scaling with a uniform grid (i.e. high resolution) does not look as good as the same number of samples on a non-uniform grid (MSAA, or similar), which will also cross over at some point.

Thus I propose that shading at 300dpi (to pick a number) and normal viewing distances is a poor use of rendering resources, and instead you'll end up with a better image by using well-placed samples and better shading at a lower resolution and then upscaling it. Where the crossover happens in terms of density is hard to say and probably content-dependent though.

At typical desktop monitor pixel densities and viewing distances I'm not sure that's a worthwhile tradeoff. Perhaps with potential 4k high DPI monitors that might be quite interesting.

Your proposal becomes far more interesting, however, when you get into the living room with a HTPC or console experience on a HDTV at typical view distances. At that point, I rarely see a difference between, say 720p upscaled and 1080p. Limiting the upscaling to shading resources might even be mostly non-noticeable going from 480p to 1080p.

Regards,
SB
 
Lets take your DVD for example. Assuming there's also a BRD available of that movie. If you upscale that DVD source it will never look as good as source material from a BRD at 720p or 1080p. Granted if you sit far enough away from the TV any differences might not be perceivable but they are there. 480p = 480p upscale < 720p = 720p upscale < 1080p.

I've not argued otherwise, the point is in the very thread I linked to, it is seen that even doing 1080->480->1080, and even zooming in on the image, we observe two things:

First there's plenty of fine detail in the image upscaled from 480p
Second in some scenes it is difficult to tell the difference between upscaled 480p and native 1080p(depends on content in the scene)

In fact the whole linked-thread is to see if the masters used were originally HD masters or not, as it is difficult to tell even that.

That addressed I've already said that 480p might not be the optimal resolution given aspect ratio issues, higher resolutions may be used maybe sub720p or maybe even 720p depending on available resources.

The point is getting perfect image quality. Even the ps2 was able to provide excellent iq in short order through the photomode of gran turismo 4. The gran turismo 5 has an even better photomode.

This is the kind of thing I'm talking about
ps2 photo mode gt4
gt4 photo mode

or even better the new gt5 photo mode
gt5 photo mode
A 480p source upscale will never look better than the original 480p source. Upscalers aren't magic and can only attempt to faithfully reproduce the details available in a 480p source.
I don't expect the upscaler to do magic and add fine detail where it is lost, that said the image can be quite sharp and some areas of the image may very well be virtually indistinguishable from native of a higher res(depends on content). You only have to look at the comparisons, especially on the newer quadhd, where given the additional data of an initial 1080p image the upscaled image will be both extremely sharp and look stupendous(the upscaled quadhd can actually look better than native 1080p). I've seen cases where an upscaled image can actually look substantially better than the original non-upscaled image, being sharper and less blurry....

In fact that's one of the things people complain about when comparing dvds to blu-rays, the upscaled image can look better than the native dvd at times substantially so ;) That's why they often say it is better to compare bluray images to native dvd, not upscaled dvd as the latter can look better and diminishes the magnitude of the difference when comparing.


That said the whole point is whether the downside of lost fine detail is enough to balance the upside of elimination of pretty much all IQ artifacts.

If I can take a 720p image supersample it heavily and I get perfect Image quality, would it be better to do that or to use a higher resolution with artifacts all over and one complaining constantly?


steampowered rather than asking, why not run a game at 1080 with 4xssaa benchmark it take a screenshot
run the game at 640x480 with 4xssaa make sure its running fullscreen on your 1080 monitor

That's not the point, the point is that for the resources that you use to 4xssaa a 1080p image you can apply higher levels of SSAA to the 480p or preferably other resolution that preserves aspect ratio.

Say those same resources 4xSSAA 1080p can achieve such a high level of SSAA at a lower resolution that the image quality becomes perfect. All of a sudden this changes thing, because you have CG-like image quality, and if the resolution is say 720p or slightly lower the level of detail will be indistinguishable from the native 1080p for pretty much anyone even zoomed in when it comes to most content. As I mentioned the content that might be most noticeable for humans(small text), can be kept at native resolution and merged with the final image(for example HUD elements.).
 
Last edited by a moderator:
I seem to have a problem turning off vysnc in games cat 12.2 preview
I wonder if someone could benchmark 1920 4xssaa and 640 4xssaa
as steampowered doesnt look like he's prepared to do it

ps: gt5 video look at the white lines between the floor tiles they look aliased to me

If I can take a 720p image supersample it heavily and I get perfect Image quality
you keep saying that with no proof

just run at 1080 with ssaa if its too much use adaptive ssaa or temporal aa or msaa with a different filter or mlaa fxaa or one of the many variants or 2xssaa with 2x msaa
 
I've seen cases where an upscaled image can actually look substantially better than the original non-upscaled image, being sharper and less blurry....
You know same kind of sharpening and de-noisyfying (if that's even a word) filters can be applied to images without scaling them. I do it with photos all the time.
 
wonder if someone could benchmark 1920 4xssaa and 640 4xssaa
The expense is not comparable for 480p you'd need higher SSAA to expend as much resources as 1080p 4xSSAA.

Take 480p and set it to say 12xSSAA, don't know how much that takes, but it will likely yield a pretty perfect picture. Ideally you'd use something higher than 480p with proper ratio for new tvs, maybe slightly below 720p or maybe 720p.


ps: gt5 video look at the white lines between the floor tiles they look aliased to me
Watch video in youtube itself directly and set resolution in options to native 1080p. Set to full screen after doing that. Aliasing is virtually nonexistent.

You know same kind of sharpening and de-noisyfying (if that's even a word) filters can be applied to images without scaling them. I do it with photos all the time.

True, but without the added memory of the additional pixels all that extra sharpening will have to be compressed to fit into the existing pixels rather less natural than interpolated pixels.

DVD video is already mastered and processed by professionals to get the final result at 480p... upscaling still further improves the image because there are more pixels to store the results over.
 
Last edited by a moderator:
True, but without the added memory of the additional pixels all that extra sharpening will have to be compressed to fit into the existing pixels rather less natural than interpolated pixels.
... or you could just render at 1080p and apply filters there instead of while upscaling from lower resolution and thus loosing details.
 
... or you could just render at 1080p and apply filters there instead of while upscaling from lower resolution and thus loosing details.

The problem is that it seems that existing AA methods that can be applied cannot eliminate artifacts like aliasing and shimmering in all cases at such high resolution.

Assuming we've hollywood-like approximated assets, would you rather it look free from such artifacts or a slight increase in details at the cost of introducing unacceptable artifacts?

The artifacts break the illusion entirely, completely unacceptable, slight loss in detail by rendering at slightly lower resolution may be a more apt decision.

I will say again, someone playing with perfect iq at 1080p or 720p is in a better position than someone playing at 2500x1600 with shimmer and aliasing.
 
What makes you think they would magically go away when doing upscaling?

Upscaling per-se will not make it go away, it will simply rescale the final image to an acceptable resolution.. We know that if the image to be upscaled has perfect iq, the upscaling process can maintain integrity and in some cases even improve the image.

It is the process of supersampling that will do it, the act of doing higher levels for pretty much the same cost. Just as it takes detail away to bring things to a lower resolution it also takes away artifacts, thus by bringing it down to 720p or slightly lower you take away even more artifacts, at some point virtually all artifacts are gone...

An example:
Toy Story 1 was 24 frames per second at 1536 × 922 pixel resolution.

Despite the low resolution the image quality is better than that of someone gaming at 2500x1600
 
Last edited by a moderator:
We know that if the image to be upscaled has perfect iq, the upscaling process can maintain integrity and in some cases even improve the image.

No we dont know that at all

The expense is not comparable for 480p you'd need higher SSAA to expend as much resources as 1080p 4xSSAA.
it's the start of the process if we have 50fps at 1080 and 40 at 480 we dont need to go further as we can figure out that that result wont allow you to double ssaa and still hit the original framerate. If the difference is much bigger then we can go from there.

It is the process of supersampling that will do it, the act of doing higher levels for pretty much the same cost. Just as it takes detail away to bring things to a lower resolution it also takes away artifacts, thus by bringing it down to 720p or slightly lower you take away even more artifacts, at some point virtually all artifacts are gone...

You've got that backwards higher resolution reduces aliasing not lower

the upscaling process can maintain integrity and in some cases even improve the image.

Again no it cannot
lets imagine you have 1 white pixel and 1 black pixel and you upscale to creat a middle pixel
what colour will that pixel be ?
Answer : yes you guessed it grey
but what if the middle pixel in the original hires frame was red or blue or whatever - and thats why upscaling is bad your asking the hardware to make a guess

you have a hypothesis, prove it, do the work dont start making claims based on a different medium that doesnt even use aa
 
You've got that backwards higher resolution reduces aliasing not lower
Supersampling is one of the ways of solving this problem. Samples are taken at several instances inside the pixel (not just at the center as default) and an average color value is calculated. This is achieved by rendering the image at a much higher resolution than the one being displayed, then downsampling (shrinking) it to the desired size, using the extra pixels for calculation. The result is smoother transitions from one line of pixels to another along the edges of objects. The number of samples determines the quality of the output.-wikipedia article on AA supersampling

Of course rendering at low resolution won't lower aliasing, what will do it is supersampling.
it's the start of the process if we have 50fps at 1080 and 40 at 480 we dont need to go further as we can figure out that that result wont allow you to double ssaa and still hit the original framerate. If the difference is much bigger then we can go from there.
You should be able to apply even more supersampling to a supersampled 1080 image by using it to create a lower resolution image with even less artifacts.

you have a hypothesis, prove it, do the work dont start making claims based on a different medium that doesnt even use aa
It does use AA when it comes to computer graphics

Pixar uses something called Stochastic Antialiasing.
How PhotoRealistic RenderMan
Works

Images used in motion picture special effects need to be photorealistic—that is, appear of such high quality that the audience would believe they were filmed with a real camera. All of the image artifacts that computer graphics researchers had had to live with up to this point were unacceptable if the images were going to fool the critical eyes of the movie-going masses.

The resulting design brought together existing work on curved surface primitives and scanline algorithms with revolutionary new work in flexible shading and stochastic antialiasing to create a renderer that could produce images that truly looked like photographs.

Other studios also use AA
Many other movie effects use Mental Ray for rendering, like Posseidon's water or some Neo vs. Smith scenes in the Matrix sequels. MR uses adaptive AA with 1, 4, 16, 64... jittered samples per pixel, but decides the actual number based on contrast. Shading and geometry are sampled together as far as I know.laa-yosh

Regards to upscaling the new pixels are in all likelyhood not generated by rules that are too simple as to cause confusing or erroneous output in general, it is likely based and constrained to perform reasonable and limited generalization based on results obtained with test images.


While googling came upon these nice stats
Machines used to render Toy Story :
87 dual-processor and 30 quad-processor 100-MHz SPARCstation 20s
Total number of processors = 294

SPARCstation 20 (single processor) had SunOS 5.4 installed and used a HyperSPARC @100 MHz with 27.5066 MFLOPS

Theoretical maximum performance of the setup used by PIXAR

294 * 27.5066 = 8086.94 MFLOPS

Movie was rendered at 1526x922 pixels using Stochastic Anti-Aliasing
Scan-line rendering used, shadow mapping for shadows ( no ray tracing )

Polygon per frame of Toy Story = 5-6 Millions


-Anil_Mahmud
 
Last edited by a moderator:
Status
Not open for further replies.
Back
Top