[B3D Article] "Ripping off the veil: The mysterious PS3 hardware scaler exposed"

I made a post about it here. It most definately can be lower quality.

Take an extreme example. Which looks worse: 640x480 or 320x1024? The latter has more pixels, but will look butt ugly.
 
I made a post about it here. It most definately can be lower quality.

Take an extreme example. Which looks worse: 640x480 or 320x1024? The latter has more pixels, but will look butt ugly.

I thought as long as you stay within the target resolution you would still be adding detail to the image when increasing the number of pixels, isn't that the case?

I can see from your extreme example that a scaling filter would have a hard time to reconstruct the detail in horzontal resolution, but isn't the result related to how simple or complex the filter is?

You say "It most definately can be lower quality", but the real question is: Will the 960x1080 image scaled to 1080p look worse than the 1280x720 scaled to 1080p on the PS3?

Can you say that for sure?
 
Mintmaster said:
I made a post about it here. It most definately can be lower quality.
The end result depends on many external factors though, as well as internals of the scaler operation which noone is privy to in public yet.

Last gen 512x512 was perceptually equivalent(even superior in cases) to standard SDTV for majority of consumer base, and it actually has less horizontal and total pixels.
 
You say "It most definately can be lower quality", but the real question is: Will the 960x1080 image scaled to 1080p look worse than the 1280x720 scaled to 1080p on the PS3?

Can you say that for sure?
Okay, you have a bit of a point, because matching resolution to fixed panel displays does help. But pixels with a 2:1 aspect ratio is a bit extreme, IMO.

What I'm trying to say is that just because you have more pixels doesn't mean you'll get better quality. We saw a similar thing with ATI's 6x and NVidia's 8xS AA back in the R3xx/NV3x days. The 8xS had slightly better looking lines in one direction, but looked far worse in the other. Overall, 8xS looked worse because the aliasing artifacts on lines in the bad direction stuck out like a sore thumb, even though it had more samples.

Anyway, you can do some experiments in Photoshop to compare.
 
Overall, 8xS looked worse because the aliasing artifacts on lines in the bad direction stuck out like a sore thumb, even though it had more samples.

Lack of gamma corrected AA was the main culprit me thinks. Secondary was the rotated grid vs ATIs sparse (pseudo random) grid sampling.

Do modern scalers take gamma into effect when they scale ?

A simple doubling of pixels going from 960x1080 -> 1920x1080 would look significantly worse than a bicubic+gamma rescaling.

Cheers
 
I made a post about it here. It most definately can be lower quality.

Take an extreme example. Which looks worse: 640x480 or 320x1024? The latter has more pixels, but will look butt ugly.

This just came to mind as I was reading your post. Regarding IGN's impressions of F.E.A.R.:

While Day 1 has done a great job with the feel of the game, it unfortunately hasn't done so well with the visuals. All of the effects from the PC and Xbox 360 releases are here, like blurring when you enter slow-mo, particle effects aplenty, chips in the scenery from bullet holes and such, but it just doesn't look very sharp. To put it simply, the whole game is quite blurry. We're not just talking texture detail, which is indeed not very good, but the video looks like it's been rendered at a low resolution and then upscaled to HD. The framerate mostly holds steady, though there are a few blips or loading pauses here and there, but the main problem really is a lack of sharpness.
http://ps3.ign.com/articles/760/760663p1.html

Mind you, it's still in development...and it's IGN. Looking at the screenshots, it looks pretty weird, and maybe they're just rendering at 480p. That looks a bit too extreme for this horizontal scaling to be the culprit...but I don't know what would happen *shrug*.

Edit: hm... I guess looking at some image scaling on my own... it does look bad with horizontal scaling, but what you're seeing in the F.E.A.R. screenshots below is quite extreme. Of course, it would also depend on the resampling filter too...

screenshots:
http://media.ps3.ign.com/media/848/848829/img_4298800.html
http://media.ps3.ign.com/media/848/848829/img_4298801.html
http://media.ps3.ign.com/media/848/848829/img_4298802.html
http://media.ps3.ign.com/media/848/848829/img_4298806.html
http://media.ps3.ign.com/media/848/848829/img_4298809.html
 
Last edited by a moderator:
Mind you, it's still in development...and it's IGN. Looking at the screenshots, it looks pretty weird, and maybe they're just rendering at 480p. That looks a bit too extreme for this horizontal scaling to be the culprit...but I don't know what would happen *shrug*.
Isn't the scaler in the B3D article about scaling to 1080p?
 
Isn't the scaler in the B3D article about scaling to 1080p?


ah...sorry, I think I was thinking of 640x1080. :oops: Well, something weird is going on with the screenshots they have (even if 720p) and it could be horizontal scaling.


Having brain farts all over today.
 
Have you ever seen a 1024x1024 16:9 Plasma HDTV? I don't think that they look bad IMO.
Remember that the image going to a 1024x1024 plasma TV is completely antialiased because it's captured by a video camera. I don't think most devs will be doing 4xAA in their games at 960x1080. Finally, 1080i sources usually don't have even close to 1920 pixels across of actual information.

I'm make a picture later on today to show you guys what I mean.
 
Those pics look like all the game assets were downsampled by 2. They can't ship it looking like that, their is zero reason to butcher the textures. My PC can play FEAR at 1368x768 with 4X AA with 256MB 7900GT, so I doubt it's a memory constraint at 720P.
 
Anyway, you can do some experiments in Photoshop to compare.
Master and Servant. ;)

1. original 1080p image
originalqx0.png


2. 1080p image scaled to 960x1080 and back to 1080p
scaledto1080960andbacktha8.png


3. 1080p image scaled to 1280x720 and back to 1080p
scaledto1280720andbacktqb0.png


I used the standard bicubical filter in Photoshop when scaling both ways.

There is not a big difference between picture 2 and 3, but no 3 is significantly brighter than the original and I think the water in picture 2 looks significantly better than picture 3. Concerning the water, there maybe some connection to the horizontal lines in the water, but the colour of the water looks better as well in my opinion.

Maybe there is some connection to the bicubical filter and the fact that the picture (no 3) that is scaled in both horizontal and vertical direction get brighter, perhaps someone have more insight in the beahviour of this filter. Does it slightly favour bright pixels?
Picture 2 might also be slightly brighter than the original.
 
These are crops from a render without any AA (scaling is B-spline filter):

1920x1080
1080p.jpg


1280x720 scaled
720sc.jpg


960x1080 2.0 pixel aspect scaled
960.jpg
 
These are crops from a render without any AA (scaling is B-spline filter):

...

960x1080 2.0 pixel aspect scaled
960.jpg

The scaling of your 960x1080 image is wrong since you're losing vertical resolution as well as horizontal. You shouldn't be losing any vertical resolution.
 
How should I do it then ? No problem to do another one.
Maybe, Should I do field rendering and then composite the image - Is that what a hardware scaler does ?
 
Last edited by a moderator:
Crossbar, you've probably figured out that lack of AA is where the biggest concern is so photos aren't very representative of reality, and furthermore your original picture is very blurry to begin with.

macabre, that's exactly what I was talking about. Horizontal-ish lines like the helmet near the eyes or the top of the shoulders look better with 960x1080, but the vertical-ish lines look so much worse that it doesn't matter. Just imagine how that would look in motion.

Just a question: Did you generate the 1280x720 and 960x1080 pics using point sampling on a 1080p pic, and then upscale them using B-spline? The 1280x720 pic could be at a slight disadvantage this way, though there isn't really any other way of doing it without getting a native 720p screenshot.
 
Crossbar, you've probably figured out that lack of AA is where the biggest concern is so photos aren't very representative of reality, and furthermore your original picture is very blurry to begin with.
Hmm... :)

Yeah, it seems pretty obvious that AA is essential for stuff to be scaled. The pictures also showed that scaling is by no means a substitute for the real thing.
 
Back
Top