I can't really tell if the 16x aniso shot has curved MIP-map boundaries. I also can't tell if the regular trilinear shot's boundaries are spherically curved, or angled like the 8500 (vertex in the middle of the screen, straight lines extending left and right--a "triangle," not a sphere).
Anisotropic Options with 9700
- Thread starter Joe DeFuria
- Start date
It is curved. ATI have told me this and I'm just running something to prove it.
[edit] - here we go, the inorgural showing of Basic's Anisoview.
[edit] - here we go, the inorgural showing of Basic's Anisoview.
Doomtrooper
Veteran
The End.
Hmm, quite a difference between 2x and 4x... and those boundaries are a bit angled, but that's not very obvious.DaveBaumann said:It is curved. ATI have told me this and I'm just running something to prove it.
[edit] - here we go, the inorgural showing of Basic's Anisoview.
Anyway, it's finally trilinear, and fast.
Thanks, Dave. So the tri and bi are curved; tri clearly using a spherical approximation, and bi too low-res to tell if it's a real sphere or an approximation. Why does the red/yellow boundary on the 4x shots look straight, while the further boundaries look curved?
DaveBaumann said:
Excellent, thanks for the pics. This certainly looks much better than the 8500's aniso. I don't think it looks quite as good as the GeForce line's, but I seriously, seriously doubt it will be distinguishable in any situation.
Still, I would really like to see a comparison of that same program on a GeForce4. If you don't have one on hand, you could send me the prog and I could e-mail you the pics...
Chalnoth said:
No, that's not quite accurate, although that sort of effect is captured in many screenshots. The boundary shape is a multi-sided polygon when doing isotropic filtering on the Radeon 8500, and depending on the viewing angle, you can see several straight-lined segments.
An important thing to keep in mind is that these boundaries only indicate where the selected mipmap changes... it doesn't necessarily tell you much about how the texture sampling behavior is changing.
Doomtrooper
Veteran
Chalnoth said:
I don't know how you say that with a straight face, screen shots of Comanche 4 and Serious Sam proves your IQ superiority theory is simply
Dan G said:
Right...the MIP map selection algorithm is only one piece of the overall quality of texture filtering.
Plainly-speaking, the further away a certain MIP map is, the more texture clarity is possible, but also the more fillrate is required to render the shot with minimal AA.
What I'd really like to know is whether or not the higher-precision pipelines of the R9700 and NV30 can eliminate some of the texture aliasing I currently see on my GeForce4 TI 4200. The aliasing that I am talking about is either a result of the linear interpolation, or crops up within the pipeline due to mathematical errors. I can see this aliasing even with an LOD set to, for example, +3 (It's not remotely noticeable in most situations...just a few). If the aliasing can be fixed with higher-precision pipelines, it may require a little bit of high-frequency noise inserted somewhere in the pipeline to completely get rid of if 32-bit (or lower) source art is used.
If the aliasing is a result of the linear interpolation done, then it could be reduced further by switching to a quadratic interpolation scheme (requiring three texture samples in at least two directions, for a total of five samples at a minimum, and quite a bit more math), or otherwise reduced with high-frequency noise (which would be more effective with higher-precision color). I believe another word for this high-frequency noise is dithering.
hmm. Why is 4x so much less curved than 2x or no aniso? Even looking at the furthest-away mip levels (even on 2x aniso!), they are a lot more curvy than 4x... ?!?
(edit)
Actually, now that I think about it, a pixel that is further from the center of the screen should have a higher degree of anisotropy. As your anisotropy level increases, you should be able to flatten the mip-map boundaries until they are close to perfectly linear. This would explain why the 16x Radeon9700 aniso shots have linear mipmap boundaries.
However, this raises an interesting question. Weren't the GeForce4's anisotropic mipmaps curved? Given that the geforce4 has very precise anisotropic filtering, it should be able to apply maximum degree of aniso to pixels near the screen's edge. In that case, the curved mipmaps of geforce4 are erroneous!!!
(/edit)
(edit)
Actually, now that I think about it, a pixel that is further from the center of the screen should have a higher degree of anisotropy. As your anisotropy level increases, you should be able to flatten the mip-map boundaries until they are close to perfectly linear. This would explain why the 16x Radeon9700 aniso shots have linear mipmap boundaries.
However, this raises an interesting question. Weren't the GeForce4's anisotropic mipmaps curved? Given that the geforce4 has very precise anisotropic filtering, it should be able to apply maximum degree of aniso to pixels near the screen's edge. In that case, the curved mipmaps of geforce4 are erroneous!!!
(/edit)
Yes, all MIP map boundaries of a GeForce4 with anisotropic are curved. They look very different from what you're seeing with the Radeon 9700 there.
Depending on the angle, the GeForce4's MIP map boundaries change from a semi-circular pattern that is further toward the center of the screen, and closer at the edges, to a hyperbolic pattern that is closer toward the center of the screen, and further at the edges.
I really do not know why nVidia has chosen this MIP map boundary when anisotropic is enabled. Hopefully it is the correct boundary, but not knowing the exact sampling algorithm used in anisotropic, I have absolutely no way of knowing how to calculate whether or not it is correct.
Additionally, I have no way of telling whether the 9700's boundaries are more or less correct. They may be just as good, if the 9700 uses different anisotropic sampling methodology (aside from the fact that the GeForce line's boundaries are always curved, while the 9700 still uses boundaries that are the sides of a polygon...but given the increase in the number of edges of that polygon over the 8500, I doubt it will be noticeable).
As far as I'm concerned, the video card that can have its MIP map boundaries the furthest from the viewpoint in a given scene, while producing the same amount of aliasing and using the same degree of anisotropic, has the superior filtering algorithm/MIP map selection algorithm.
What such a measurement of texture filtering requires is a comprehensive, non-subjective method of analyzing texture aliasing.
Depending on the angle, the GeForce4's MIP map boundaries change from a semi-circular pattern that is further toward the center of the screen, and closer at the edges, to a hyperbolic pattern that is closer toward the center of the screen, and further at the edges.
I really do not know why nVidia has chosen this MIP map boundary when anisotropic is enabled. Hopefully it is the correct boundary, but not knowing the exact sampling algorithm used in anisotropic, I have absolutely no way of knowing how to calculate whether or not it is correct.
Additionally, I have no way of telling whether the 9700's boundaries are more or less correct. They may be just as good, if the 9700 uses different anisotropic sampling methodology (aside from the fact that the GeForce line's boundaries are always curved, while the 9700 still uses boundaries that are the sides of a polygon...but given the increase in the number of edges of that polygon over the 8500, I doubt it will be noticeable).
As far as I'm concerned, the video card that can have its MIP map boundaries the furthest from the viewpoint in a given scene, while producing the same amount of aliasing and using the same degree of anisotropic, has the superior filtering algorithm/MIP map selection algorithm.
What such a measurement of texture filtering requires is a comprehensive, non-subjective method of analyzing texture aliasing.
Joe DeFuria
Legend
Good luck.
Seeing as aliasing appears to be pretty much "subjective" (depending on the person, you are more or less sensitive to it), I'm not so sure a "non-subjective" method is realistic.
You might be able to create a measurement tool that can spit out an "aliasing score". However, I would wager that the "score" at which one person declares that aliasing is evident is different from the next person.
In isotropic filtering the 8500s mip boundries are squareish but the square deforms depending on the viewing angle.
ie If you turn the view while watching the 'vertex' of the mip level it will move as you turn, the sides of the mip level will also move as the view is turned.
In anisotropic filtering the mip lines are always straight and have no 'vertices' at any angle. They also seem to be at only quarter resolution accuracy. The mip levels aren't the only visible lines though, theres also the lines caused by what I assume is the aniso filter used changing shape or increasing in samples and the shape of these is just like the 8500s istropic mip levels.
ie If you turn the view while watching the 'vertex' of the mip level it will move as you turn, the sides of the mip level will also move as the view is turned.
In anisotropic filtering the mip lines are always straight and have no 'vertices' at any angle. They also seem to be at only quarter resolution accuracy. The mip levels aren't the only visible lines though, theres also the lines caused by what I assume is the aniso filter used changing shape or increasing in samples and the shape of these is just like the 8500s istropic mip levels.
Joe DeFuria said:Good luck.
Seeing as aliasing appears to be pretty much "subjective" (depending on the person, you are more or less sensitive to it), I'm not so sure a "non-subjective" method is realistic.
You might be able to create a measurement tool that can spit out an "aliasing score". However, I would wager that the "score" at which one person declares that aliasing is evident is different from the next person.
But what I'm talking about is independent of user preference on what is deemed acceptable aliasing (I don't personally think that any aliasing whatsoever is acceptable...but there's no hardware out today that can offer that...). What I'm talking about is evening the playing field for measuring visual quality by evening out texture aliasing. After all, if you would rather have lots of aliasing, just play around with the MIP LOD.
I would wager to guess that no matter what level you set the aliasing to on the various cards, the cards will not switch positions. The only differences may be with regards to "average aliasing" or "maximum aliasing". The non-curved MIP map boundaries would obviously put the Radeons at a disadvantage when you're looking at maximum aliasing (which is what I would consider the best measurement...since this is a rendering error, it needn't be applied to the whole scene to be distracting...just like edge AA).
How about you just use your card. If you like it, great. Pretty sad when you have to start creating programs to look for problems that aren't seen in actual use. Especially when you want to do this only to try and say "my card is better than your card." This is completely pathetic.
There is undoubted value in saying that one card has better image quality than another, if it can be proved. The problem with mipmapping, filtering etc. is that it is very subjective and there's no obvious 'right/wrong' answer. e.g. some people prefer a slight sharpness at the expense of some aliasing; others like it a touch blurry but very smooth.
The only way that is really 'fair' is to do blind tests of full-moving images rendered out of a game on a large group of people. Screenshots rarely tell the whole story. Even then there's so many variables on top of just the filtering and mapping; frame rate, resolution, monitor sharpness, output image quality, etc. etc.
Frankly I'd be surprised if in a large double-blind test of Ti4600 vs. R9700 there was a big group one way or the other, and I'd be astonished if more than about a few people could say which is which. Both are 'very good'. Personally I think R9700 is better, but of course I would be expected to
The only way that is really 'fair' is to do blind tests of full-moving images rendered out of a game on a large group of people. Screenshots rarely tell the whole story. Even then there's so many variables on top of just the filtering and mapping; frame rate, resolution, monitor sharpness, output image quality, etc. etc.
Frankly I'd be surprised if in a large double-blind test of Ti4600 vs. R9700 there was a big group one way or the other, and I'd be astonished if more than about a few people could say which is which. Both are 'very good'. Personally I think R9700 is better, but of course I would be expected to
Similar threads
