Aniso quality changes between NV driver relases

[Reverend]

I have read that there appears to be a difference. Apparently the diff progressing from one aniso level to the next is smaller compared to older drivers. I'm not sure if this is specific to GF4 or GF3 or both.

Can someone point me to the last driver version with the supposedly better aniso quality?

PS. Sorry, no time to find the forum threads for this... I'm sure Sharkfood would know about the driver version.

 

[K.I.L.E.R]

On my Geforce 3 that would be 22.80's.

 

[Ailuros]

Reminds me of, has anyone any idea what the "texture sharpening" variable stands for?

 

[Babel-17]

Regarding "texture sharpening": As far as I can tell all it does is bump the anisotropic filtering a notch unless you already have it maxed out.

 

[Ailuros]

That's what EE told me too when I asked him, yet I was hoping for a more detailed explanation than that.

 

[Reverend]

LOD bias.

 

[Ailuros]

Rev,

There's a variable for "texture LOD bias" and one for "texture sharpening". It's not hard to guestimate that both address mipmap adjustments. What's the difference between them?

 

[noko]

Where is this "texture sharpening" variable? How can I activate it? Is it only in the Register?

What values do you add to change the setting in the Register?

 

[Reverend]

One allows a range of values.

The other (the sharpening one) sets a fixed value (I believe it is no more than -0.25) but only if AA is enabled at the same time.

 

[Sharkfood]

Howdy Rev!

I'm not sure when the exact point of the change occurred, but I think the previous guess of 22.xx's/23.xx's sounds about right.

The *easiest* way to see this is to use a GF3, hopefully a non-Ti if you have available as I don't know about sense/detection in earlier dets. My test sample is a Leadtek Winfast GeForce3 TD.

You also need a game that has a fair amount of texture aliasing. NOLF might be a good test although I haven't tried this game as it's not an "everyday" game on my systems. Instead, I originally spotted it in Need for Speed-Porsche Unleashed.

The test suite involves doing the following:
Load up the GF3 + 12.90 drivers.
1) Set the game to 1024x768x32, LOD Bias "Best Quality" in the display properties (i.e. LOD Bias of "0" using a tweaker) and also set maximum anisotropic filtering level (8x).
2) I use Schwarzwald track as this track has a lot of "flicker" and grainy, pixelated ground and brick patterns with regular moire near the town spot. Race and view. If you want a couple spots to take screenshots, I use the spot by the logging buildings near the first half of the track with the brown, pixelated track.. as well as near the end at the little town where Hansel and Gretel are waving at eachother across the street. I'll slam on the brakes and hold the e-brake still for like 10 seconds when I have found a prime spot. Then save the replay at end of race for taking exact spot shots later on.

On the 12.20's -> 12.90's, the track texture is almost totally free of pixel shimmering and flicker. It does a substantial job at removing the aliasing. In the town spot, moire is greatly reduced although still has some that exists.

On the 28.32->29.42's at 8x AF, a couple new behaviors occur:
1) I see a single pixel line in the guardband area along left and top borders (almost like the old V5's used to do with AA).
2) Performance between 4x->8x is no longer a linear change, but instead a minimal hit in performance.
3) Shimmering on initial logging building area is greatly visible compared to the 12.xx's. The brick road stage by the buildings in town still has traces of moire, but is also overall less sharp, even with same LOD (almost as if a filter or non-linear LOD is applied with anisotropic filtering enabled).

In all, it is nearly impossible to distinguish any change between 4x->8x in the 29.42s, as well as the performance hit has been greatly reduced. In the 12.xx's, the switch to 4x -> 8x did indeed yield improvement, albeit subtle and hard to find in screenshots, it was readily visible in motion, especially on this track. The GeForce4 follows this trend, obviously, although I havent done side-by-side with same monitor to see if there is any extra degree to this yet.

The whole issue should be readily apparent to anyone migrating from a GF3 on older drivers to a GF4 with 28.xx/29.xx as AF between 4x and 8x has been visibly reduced in quality, while also being improved in performance (the hit between the two that is).

On the other thread, it took a custom UT level with infinite clipping plane (full distance Z) on a texture consisting of a grid pattern to see ANY difference between 4x and 8x at all. Obviously, if it requires a Z value that is well above and beyond any current 3D engines clipping plane, it's of little practical use. This was NOT the case previously with AF + NVIDIA cards as 8x was the extra little bit of "ummphh" to clear up problem children titles like NFS-PU and NOLF to eliminate a small sample more of aliasing/shimmering.

Hope that makes sense! I can clarify more if needed but it really only takes any knowledge or experience with the GF3/GF4 for any amount of time in the past year or so to see the difference.

Cheers,
-Shark

 

[KimB]

On the other thread, it took a custom UT level with infinite clipping plane (full distance Z) on a texture consisting of a grid pattern to see ANY difference between 4x and 8x at all. Obviously, if it requires a Z value that is well above and beyond any current 3D engines clipping plane, it's of little practical use.

Just because that level showed the difference doesn't mean it's the only way to show the difference. In fact, there was an obvious difference in previous shots, too, with much lower z-distances, if you knew where to look. And if you want to say it's invalid if you, "know where to look," then don't forget that these are the sorts of things that jump out at you when the game is in motion, that distract you when you are used to seeing the superior image. Image quality problems that are visible in, at most, 1% of a given frame and only about 2% of the time can still be very annoying.

To me, increasing visual quality in games is largely about leaving nothing for the game to distract me from the game. If I see texture blurriness, I'm not paying attention to the game. If I see edge AA or texture AA, I'm not paying attention to the game. A noticeable flaw seen only occasionally is enough, for me, to make it hard to get back into the game. Of course, as hardware gets better, I get more and more critical. I currently play at 2x FSAA and 8-degree aniso pretty much all the time. It is now very easy for me to notice if either is turned down. I'm sure with future hardware I'll be expecting 16x FSAA and 32-degree aniso on all the time (or something to that effect).

Update:

If you want to show that there's a difference between these drivers on the GF3, take a GF3 with those drivers you mentioned, and show it with screenshots (that level I made is perfect for the comparison....).

 

[Ailuros]

Thanks Reverend.

 

[Sharkfood]

. In fact, there was an obvious difference in previous shots, too, with much lower z-distances, if you knew where to look.

It has absolutely nothing to do with "knowing where to look"- it has to do with there being little to no difference.

then don't forget that these are the sorts of things that jump out at you when the game is in motion, that distract you when you are used to seeing the superior image.

That's the entire problem. There is no improvement with 8x whatsoever with this very issue where there was previously. I thought this was made very clear.

But once again, if some people are willing to settle with half-hearted or incomplete, or otherwise erroneous implementations- more power to them. I personally prefer to see at least last-year's technology bested (or at least equalled) by the current product line and drivers.

And I'm also not willing to make custom UT levels with infinite Z-clipping distance to try and illustrate that something is indeed happening. When it comes down to this level of extremity to defend a product feature, this should be signal enough that something is drastically wrong.

Cheers,
-Shark

 

[KimB]

But it didn't require that level to show the difference! I noticed the difference in the shots you yourself posted, and I even pointed out some differences in some other shots where the view distance wasn't so extreme.

If, for example, I had modified that level so that the 'eye' was much closer to the ceiling or floor, the difference would have been visible much, much closer to the person. I chose to do it in that fashion because the differences leap out immediately in the screenshot. That does not mean the differences are not there, or are not noticeable, in real games.

The point is that there is a real and visible difference between 4x and 8x anisotropic filtering in later drivers. I have proven this without a shadow of a doubt. Can you show, with more than just words, that there was more of a difference with older drivers?

 

[Sharkfood]

Can you show, with more than just words, that there was more of a difference with older drivers?

I already provided very specific, precise, exact details how this can be proofed/tested in the comfort of your very own home and with one's very own eyes. As you claim to be an expert and dissenter on the factual evidence provided, it's assumed you have already performed said proofing requirements and disagree. Or is this more along the lines like your contradictions to Radeon experience having not ever used one but making wild assumption provided they are in favor of a particular IHV?

But don't take my word for it. It's all over the internet. Straw man tactice to avoid the cusp of an issue suggest simply probing for more and more evidence yet more and more is brought forth. If you wish to dissent with this any further, I'd truly suggest you bring some evidence of the contrary as I've already done my share x10. But here are even more references with distinct, pure, specific changes between 4x and 8x (as some have been labeling as 32-tap or 64-tap at the GeForce3 launch):
At NvNews, some nice distinct shots on the GeForce3:
http://www.nvnews.net/previews/geforce3/images/anisotropic/q3_anis32.shtml
http://www.nvnews.net/previews/geforce3/images/anisotropic/q3_anis64.shtml

At Xbitlabs, with a nice article:
http://www.xbitlabs.com/video/gf3-anis-filtering/
And a nice "delta" shot difference between the higher levels:
http://www.xbitlabs.com/video/gf3-anis-filtering/scr-3.jpg

Reverend had an EXCELLENT descriptive article on this, with lots of great shots and information on his GeForce3, but it appears to be down right now (edited the link, should be okay with new link - Rev):
http://www.3dgamingnews.com/3dpulpit/misc/aniso_gf3/

Yet another site here on the Geforce3 has some good examples, albeit a bit small, it's still very visible even at this lower scaled percentage:
http://www.nvworld.ru/docs/r200_gf36.html

But it didn't require that level to show the difference! I noticed the difference in the shots you yourself posted, and I even pointed out some differences in some other shots where the view distance wasn't so extreme.

No, you left everyone scratching their heads as you just took pre-labeled shots and proclaimed to the effect of "Aha! I can tell the difference!" with nothing substantiated.

The point is that there is a real and visible difference between 4x and 8x anisotropic filtering in later drivers. I have proven this without a shadow of a doubt.

There is nothing but doubt as you have proven absolutely *nothing* except that you claim to have some special Voodoo black Majik that seems to be able to decipher the difference between the two modes, from which everyone else is left scratching their heads.

You have proclaimed it's "obvious" and that "you just need to know where to look." When using pics that are labeled as 4x and 8x, it's simple (i.e. Strawman) to say "Well, the 4x is *obviously* 4x! And the 8x-shot! Look at that, yep yep.. it's definately 8x! Can't you tell?"

Well, if it's so obvious, plain as day, and you do indeed seem to "know where to look" then these shots should be a trivial task to point out to the rest of us which are at 4x and 8x, hmm? I'd be very interested to see if you can point out so everyone else can possibly figure out *what* it is you are keying on to designate the difference between these levels with the new drivers.

These are a few shots of the two games you stated were "obvious" to decipher, taken in 4x and 8x, but this time not labeled. How "obvious" is the difference and would you care sharing "where to look" and how many times you either had to magnify or otherwise flip between the images to decipher the difference:
http://shark_food.tripod.com/aftest/aftest.html


I would also add that when tactics of using 400-800% zoom and having to inspect the binary of a bitmap is needed to try and decipher a feature level, it kind of defeats the purpose of using said feature.

Lastly, as I'm not a purchaser of the cheap, bottom end, Kmart versions of the mail-order 4200's, I'm more interested in seeing this feature fixed to it's original quality level found in the GeForce3 over a year ago. The test with NFS-PU is a pretty easy one and illustrates the issue quite well. Schwarzwald with shimmer, moire and flicker versus these issues greatly curbed in the generation + drivers past. It really doesnt take a rocket scientist to understand this difference in games.

I have more vested in this hardware purchase and would personally appreciate if product quality would take precedence over IHV preference.

Cheers,
-Shark[/i]

 

[McElvis]

These are a few shots of the two games you stated were "obvious" to decipher, taken in 4x and 8x, but this time not labeled. How "obvious" is the difference and would you care sharing "where to look" and how many times you either had to magnify or otherwise flip between the images to decipher the difference:
http://shark_food.tripod.com/aftest/aftest.html

Well, I can hardly notice any difference between these shots...

I would also add that when tactics of using 400-800% zoom and having to inspect the binary of a bitmap is needed to try and decipher a feature level, it kind of defeats the purpose of using said feature.

And I totally agree with this comment.

 

[KimB]

You didn't post any evidence about what I asked for.

Do you have any evidence that the GeForce4 is showing less of a difference between anisotropic than the GeForce3 did?

 

[Pete]

The point is that there is a real and visible difference between 4x and 8x anisotropic filtering in later drivers. I have proven this without a shadow of a doubt.

It looks like Sharkfood's mystery page o' aniso has brought us past the shadow back to the doubt itself. His pics do indeed show there is no real and easily visible difference between the two aniso modes in his (static) shots. And it seems as if he's using a GF3 for the pics, as indicated in his first post in this thread.

As for the difference between earlier and later drivers, and GF3 and GF4, I'll leave that to him.

 

[noko]

With the 29.42 drivers I can tell the difference between 4x and 8x, it isn't great but it is there.

 

[Mintmaster]

Sharkfood, you should have taken the NFS5U shots from the bumper cam. That's where aniso is needed most, as angles get very oblique

That's where I race from (the car is far more responsive and gives better feedback when the camera is solidly fixed to the car), and its where I realized how utterly important anisotropic filtering really is.

You're driving along, trying to look as far ahead as possible to plan your driving line, so without aniso the textures are so blurry the dashed lane divider disappears, and it can really affect your judgement.