Could activation of SSAA improve texture quality?

[demalion]

By your definition, the first is more detailed, and I think that is not a useful definition of it, because it is quite clearly dropping detail. The same with your black and white striped mountain example.

You maintain that detail is only the number of distinct colors and the contrast between them, not the picture being represented by those colors. I don't see how that is valid.

tr.v. de·tailed, de·tail·ing, de·tails (d-tl)
To report or relate minutely or in particulars.

It isn't my definition of the word You are talking about an accurate representation of the image, not detailed.

Please note the absence of the criteria you provided of "contrast" and "distinct color count" being specified as the only meaning of "particulars" that is acceptable. Please then note that my concept of information is then not precluded, nor is the applicability of accuracy to the word "detailed". The point of rendering the textures is to represent something accurately: the color contrast is not the only criteria for achieving that, and the position of the colors is part of the information that make up its "details".

By your usage, it seems a 16-bit texture with a gradient is more detailed than a 32-bit texture trying to achieve the same thing, because the contrast is greater and the number of distinct colors you can count is higher.

I think that you are proposing that one color cannot contain more information than another color, thinking of a general case where that is true and disregarding the cases where it is not, and I think that completely ignores the context in which the color occurs.

If an image has less contrast, it has less contrast...that does not mean it therefore has less detail than another image that happens to only have black and white pixels.

You seem to be ignoring the idea of what the texture is trying to portray...a texture isn't just a bunch of colors and contrast, it is the position of those colors and contrasts as well. Even when supersampling isn't introducing new colors in screen rendering (does your analysis recognize that part?), it is possible that it will still introduce "positional data' (the term I'm making up, don't know of a better name yet).

That is why your definition of sharpness needs to decide between "detail" and "aliasing" (as you propose is "contrast"), it can't apply to both in the way you are proposing. Or, atleast, this definition did nothing to clarify how it can.

Are you comparing SSAA to a higher resolution, rather than to the image of the same resolution without SSAA applied?

I think you need to consider my examples and the concept of "information" I proposed more thoroughly.

 

[Pete]

Forget about AF. I know what is it, how it work and what it will give me. I dont care. I thought that if i apply SSAA it will somewhat blur my textures(MSAA dont affect textures at all).
Ok. We have 2 screenshots.
http://ixbt.com/video2/images/nv35/nv35-ss-aa8x-part.jpg
http://ixbt.com/video2/images/nv35/nv35-ss-aa4x.jpg
On first screen texture is much sharper and ditailed. Could it be result of applying SSAA? Just in theory.

Yes. See here for how GF3's multisample AA (Anand incorrectly labels it as supersample, for some reason) compares to the 8500's supersample AA--the 8500's SSAA clearly affects textures, whereas the GF3's MSAA doesn't. As Humus and others pointed out and Anand shows, though, AF is a better way of achieving the same result.

As nV's 8x FSAA is a combo of MS & SSAA, it will have slightly better-looking textures than 4x's pure MSAA. However, the blurriness you see in ixbt's screenshots may be due more to the NV35's apparent problems with blurriness at 4x AA (as Hexus.net showed in their 5900U preview).

 

[madshi]

That's true, Quincunx is an even larger blur introduction then traditional SSAA.

What is "blur"? In my book "blur" is artifically mixing neighbour colors to make an image appear softer. Quincunx does that, but SSAA doesn't do that. So IMHO your statement (quoted above) is flat out wrong in implying that SSAA would introduce blurring. Maybe SSAA lowers the contrast. But that's different than blurring. A non-SSAAed screenshot has such a high contrast only because of pixel "errors", because of bad texture aliasing. Reducing the contrast by using a blur filter is an option. But that's not what SSAA does. SSAA reduces the contrast by correcting the pixel errors. By reducing the texture aliasing. It adds more image information. Maybe "sharp" is not the right description for that. But "sharp" is less wrong than "blur" is.

For everybody who is reading. Here are two fractals, one without SSAA and one with 16x SSAA. Be warned though, uncompressed = large!

http://madshi.net/fractal.bmp
http://madshi.net/fractal16x.bmp

Check out the top right of the bitmaps. You'll see that the SSAAed bitmap has lots of more details there.

Enjoy!

 

[arjan de lumens]

SSAA can indeed blur textures. Consider the case where a texture is lined up so that each texel is mapped to exactly 1 pixel on the screen, with the texel midpoint matching the pixel midpoint. In this case, SSAA will sample a number of locations within the pixel, outside of the pixel midpoint, and sum them together. For each such sample, it will blend in the texel at the midpoint AND a few neighboring texels, the result being introduction of blurring that wouldn't be present without AA.

 

[Joe DeFuria]

SSAA can indeed blur textures. Consider the case where a texture is lined up so that each texel is mapped to exactly 1 pixel on the screen, with the texel midpoint matching the pixel midpoint. In this case, SSAA will sample a number of locations within the pixel, outside of the pixel midpoint, and sum them together. For each such sample, it will blend in the texel at the midpoint AND a few neighboring texels, the result being introduction of blurring that wouldn't be present without AA.

I don't beleive that's correct.

Assuming the SuperSampled "image" has the correct higher resolution texture, all sampled texels will be of the same color, and the resulting AA'd image will be exatly the same as the original image.

 

[RussSchultz]

If the texel size is the size of the (resultant) pixel, then it won't affect anything. Every sub sample you take will sample the same texel.

No?

 

[BoddoZerg]

Saying that SSAA blurs textures because it reduces contrast reminds me of the old saying:

A clock that loses one second a day is accurate only once every hundred years... but a clock that's stopped is right twice a day!

 

[arjan de lumens]

If the texel size is the size of the (resultant) pixel, then it won't affect anything. Every sub sample you take will sample the same texel.

No. Especially if you enable bi/trilinear interpolation. Then, for each sample point, you will get that since the sample point doesn't match the center of the texel, it picks the 4 texels closest to the sample point and blends them together. With sample points spread around within the pixel, the different samples will even pick different groups of 4 texels for each subsample.

 

[BenSkywalker]

Demalion-

Do you consider trilinear to offer more detail then bilinear? If you don't, you need a significantly different line of reasoning. The only way to come close to truly increasing the accuracy of the textures utilized is by using AF, you need to have the z value factored in to sample it properly. SSAA does not do this, it utilizes isotropic sampling patterns which is the same inaccurate sampling.

Madshi

In my book "blur" is artifically mixing neighbour colors to make an image appear softer.

That is a blur filter. The type of blur that SSAA introduces is not as bad, but it still noticeable. Haloing is probably the best example. You have a 3D viz app? Create a cluster of spheres that interlock each other by a reasonable amount and checkerboard them bright red and green. Then run it through 16x SSAA. You end up with brown between the spheres, a color that is not present in the example. The lower the resolution or sampling the smaller the Halos get, but they are still there. This also works if you simply create a grid texture and apply it to a plane and do the same thing.

In this aspect, rotated grid tends to be worse as the sampling proximity is so far off from ideal(in terms of introducing blur, although that is also the best way to reduce aliasing for SSAA).

For your pictures, one of them is 8bit while the other is 24bit, did you upload the wrong image for the non AA shot?

 

[Xmas]

No. Especially if you enable bi/trilinear interpolation. Then, for each sample point, you will get that since the sample point doesn't match the center of the texel, it picks the 4 texels closest to the sample point and blends them together. With sample points spread around within the pixel, the different samples will even pick different groups of 4 texels for each subsample.

If you use 2x2 OGSS and box-filtered mip maps, the result will be identical. Because the sample positions will each be exactly one of the four texels that are used to generate the lower-level mip map used in the noAA situation.


Some people say point sampling gives sharp and detailed textures. I'd say it's neither sharp nor detailed image - it's error. Supersampling certainly gives a more detailed image when done right.

 

[madshi]

The only way to come close to truly increasing the accuracy of the textures utilized is by using AF, you need to have the z value factored in to sample it properly. SSAA does not do this, it utilizes isotropic sampling patterns which is the same inaccurate sampling.

Of course AF is perspective correct and SSAA is not. But that doesn't mean, that SSAA doesn't increase the texture accuracy. It just means that AF is even better for textures than SSAA is.

Haloing is probably the best example. You have a 3D viz app?

No, I'm sorry.

Create a cluster of spheres that interlock each other by a reasonable amount and checkerboard them bright red and green. Then run it through 16x SSAA. You end up with brown between the spheres, a color that is not present in the example. The lower the resolution or sampling the smaller the Halos get, but they are still there. This also works if you simply create a grid texture and apply it to a plane and do the same thing.

Can you provide me with some sample screenshots, please?

What you describe sounds bad, but I don't see it that bad. Think about this: If you look at a newspaper black and white photograph from a distance it looks like it would have several gray tones, but it's really only black and white (at least when looking at cheap newspapers). The eye "calculates" gray tones, because the eye's resolution is not that good from a distance. If you look at it very closely the eye will see that it's really black and white. Now SSAA does exactly the same thing. If you print your greed/red checkerboard and hang it on a wall and go some steps away, you'll also see it brown. That's the best possible way to reflect the real image information, IMHO.

For your pictures, one of them is 8bit while the other is 24bit, did you upload the wrong image for the non AA shot?

The SSAAed picture in the original resolution (2560x1920) was 8bit, too. Just see it as an 8bit texture. Of course the SSAAed output must be in true color.

 

[KimB]

Of course AF is perspective correct and SSAA is not. But that doesn't mean, that SSAA doesn't increase the texture accuracy. It just means that AF is even better for textures than SSAA is.

Not necessarily. Supersampling and AF do different things for textures.

Anisotropic filtering sort of works as "directed supersampling" where a different number of bilinear samples is taken from each pixel.

Supersampling just decreases aliasing across the entire image, allowing for an overall texture LOD increase. In particular, it will decrease texture aliasing for the base texture, and will help out bump mapping significantly (as well as any other form of texture operation that doesn't filter well).

That is, supersampling is sort of a catch-all. It is good for everything, but is very slow. More directed techniques will always be faster. I still expect games to start applying custom filtering techniques in shaders before long (such as for bump mapping), so that supersampling is never needed across the whole scene.

 

[demalion]

Demalion-

Do you consider trilinear to offer more detail then bilinear? If you don't, you need a significantly different line of reasoning.

No I don't need a different line of reasoning, since all supersampling is not trilinear filtering. That is like asking me if Quincunx blurs textures to prove that all multisampling blurs textures.

Why can't you simply address my examples?

Anyways, now that we are starting a new discussion about the specific case of trilinear filtering, I'd have to say it depends on sample selection and LOD. Mip map levels are additional detail, but if the sampling and mip map selection is introducing error, by sampling them more, you might be sampling more error instead of more detail.

The only way to come close to truly increasing the accuracy of the textures utilized is by using AF, you need to have the z value factored in to sample it properly.

No, AF is the better way to increase the accuracy and detail of textures, not the only way. Are you now saying supersampling doesn't increase accuracy?

Why are you making this "trilinear OR AF", and "SSAA OR AF"?

SSAA does not do this, it utilizes isotropic sampling patterns which is the same inaccurate sampling.

Isotropic sampling patterns work because the representation is isotropic...the screen. When the representation is of something 3 dimensional, they are inferior to anisotropic sampling (of sufficient degree). However, they are still better than no sampling of extra detail at all.

For detail introduction, Anisotropic > super sampling for textures in a texture oriented away from the plane of the screen. However, Anisotropic + no sampling at all < Anisotropic + super sampling. Why? Because the supersampling is sampling extra anisotropic sampled data.

Please note that this was said already.

 

[BoddoZerg]

That is a blur filter. The type of blur that SSAA introduces is not as bad, but it still noticeable. Haloing is probably the best example. You have a 3D viz app? Create a cluster of spheres that interlock each other by a reasonable amount and checkerboard them bright red and green. Then run it through 16x SSAA. You end up with brown between the spheres, a color that is not present in the example. The lower the resolution or sampling the smaller the Halos get, but they are still there. This also works if you simply create a grid texture and apply it to a plane and do the same thing.

You're getting sampling confused with blur. You seem to have the idea that, if you are looking at a very detailed texture with only red and green texels, you should only see red and green pixels on your screen. In reality, brown is the correct thing to see when you are limited by resolution.

Take computer monitors and TV screens for example. There is no "white" phosphor on your screen. Any "white pixel" you see on your monitor is composed of a large number of red, green, and blue phosphors very close to one another. The reason that you can see a white pixel instead of separate red, green, and blue phosphors isn't blur - it's because the human eye has a limited number of color-detecting cells (limited resolution), and it is forced to sample from the (quasi-)infinite resolution of the real-world.

Likewise, in computer graphics you have to sample discrete pixels from a mathematically defined construct of poylgons and textures. With simple point sampling and no mipmapping, you get a terrible mess. What you see is not "more detail", it's more error. For example, if you render a very large checkerboard floor with point sampling and no mipmapping, beyond a certain distance you will see a moire pattern. That's not detail you're seeing - the original texture did not contain a moire! What you are seeing is error. Adding mipmapping and bilinear filtering may make the texture contrast lower, and may look "blurry", but it does not decrease the detail of the render. At the point which used to have a moire, now you just see brown... you can not see any detail with or without mipmapping. In fact, even with "perfect" sampling on a finite-resolution display, there will be a point where you cannot see any detail. After all, what is "resolution"? Its the ability to resolve small details. With any finite resolution screen, there will be some size of detail which cannot be properly visualized - the only way to deal with it is to blur it into something else.

You can think of the effect of SSAA in the same way. It's not removing detail, it's removing error. Some features that would otherwise get obliterated by mipmapping will show up with SSAA... take the checkerboard moire example again. With mipmapping and without AA, the pattern becomes brown at a point very close to the screen. When you turn on SSAA, your sampling is more accurate, so you can push back the mipmap boundaries much further - at the point where you used to see brown, you will now see somewhat "dirty" shades of green and red. Thus, SSAA can improve texture detail.

 

[BenSkywalker]

Demalion-

You were serious with your prior examples? Ok then.

By your usage, it seems a 16-bit texture with a gradient is more detailed than a 32-bit texture trying to achieve the same thing, because the contrast is greater and the number of distinct colors you can count is higher.

It depends on the particular image that is trying to be displayed. If take an image that is designed to have the high levels of contrast that a 16bit texture would and then up sample it, then downsample it utilizing a 24bit AA downsample then you will lose the contrast it is supposed to have.

If an image has less contrast, it has less contrast...that does not mean it therefore has less detail than another image that happens to only have black and white pixels.

The issue at hand is taking an image the already has a given level of contrast and then altering that. Comparing the image that you are seeing at raster time compared to the native texture.

You seem to be ignoring the idea of what the texture is trying to portray...a texture isn't just a bunch of colors and contrast, it is the position of those colors and contrasts as well. Even when supersampling isn't introducing new colors in screen rendering (does your analysis recognize that part?), it is possible that it will still introduce "positional data' (the term I'm making up, don't know of a better name yet).

My analysis is based on what actually happens. I've cited numerous examples of circumstances that reduce detail. Color positioning is only properly done on an anisotropic basis, other then that you are simply amplifying errors by taking an isotropic sampling pattern and then running it over another isotropic sampling pattern. You are compounding the problem in terms of positional data.

Are you comparing SSAA to a higher resolution, rather than to the image of the same resolution without SSAA applied?

No, there is no need to.

No I don't need a different line of reasoning, since all supersampling is not trilinear filtering.

It is simply a redundant isotropic filter, SSAA just adds yet another redundant inaccuate isotropic filter on top of it.

Mip map levels are additional detail, but if the sampling and mip map selection is introducing error, by sampling them more, you might be sampling more error instead of more detail.

Which outside of staring at a wall, isotropic filtering is introducing error anyway using your positional data standard.

No, AF is the better way to increase the accuracy and detail of textures, not the only way. Are you now saying supersampling doesn't increase accuracy?

In the particular instance of textures no it doesn't. It does in terms of image integrity(pixel popping most notably) and in terms of edges in particular.

Why are you making this "trilinear OR AF", and "SSAA OR AF"?

MSAA and AF are overall by far superior in terms of AA because they are based on what the image is actually supposed to be. SSAA and trilinear(by default) are isotropic and are inaccurate, SSAA in particular is a 2D filter.

Isotropic sampling patterns work because the representation is isotropic...the screen. When the representation is of something 3 dimensional, they are inferior to anisotropic sampling (of sufficient degree). However, they are still better than no sampling of extra detail at all.

Not when you are sampling data in a manner that is contrary to what the image is supposed to portray. If you were trying to take a picture of the Mona Lisa would you do so with the camera 18" off the wall it is hanging on from ten feet away? If you took four of those pictures and blended them together do you think it would look better then one taken from head on standing a few feet back?

For detail introduction, Anisotropic > super sampling for textures in a texture oriented away from the plane of the screen. However, Anisotropic + no sampling at all < Anisotropic + super sampling. Why? Because the supersampling is sampling extra anisotropic sampled data.

For the last case you listed, you are performing an isotropic filter on top of the anisotropic thus reducing the detail of the texture.

 

[BenSkywalker]

Bodo

You're getting sampling confused with blur. You seem to have the idea that, if you are looking at a very detailed texture with only red and green texels, you should only see red and green pixels on your screen. In reality, brown is the correct thing to see when you are limited by resolution.

You are talking about what is tollerated as acceptable error, not what is correct.

Take computer monitors and TV screens for example. There is no "white" phosphor on your screen. Any "white pixel" you see on your monitor is composed of a large number of red, green, and blue phosphors very close to one another.

And there is no 'white light'. It is a collection of red green and blue.

The reason that you can see a white pixel instead of separate red, green, and blue phosphors isn't blur - it's because the human eye has a limited number of color-detecting cells (limited resolution), and it is forced to sample from the (quasi-)infinite resolution of the real-world.

No, it is the properties of light.

For example, if you render a very large checkerboard floor with point sampling and no mipmapping, beyond a certain distance you will see a moire pattern. That's not detail you're seeing - the original texture did not contain a moire! What you are seeing is error. Adding mipmapping and bilinear filtering may make the texture contrast lower, and may look "blurry", but it does not decrease the detail of the render.

Mip maps are downsampled blurs of the base texture, how can you say that isn't decreasing detail? Yes the moire pattern is an error, on a pixel by pixel basis it is more accurate then mip mapping it out and having 100% error rates on the pixels. That difference is one of what is less bothersome to people. Even when witnessing the horrendous moire that such an image would offer, at the very least a portion of the pixels inside what appears as moire will be correct. If you use mip mapping and change all the pixels to brown you assure 100% error rates.

After all, what is "resolution"? Its the ability to resolve small details. With any finite resolution screen, there will be some size of detail which cannot be properly visualized - the only way to deal with it is to blur it into something else.

The accuracy involved is what is ar question. When does the blur itself produce a less accurate image then what could be maintained without it? If we could manage 2048x AF do you think there would be a need for mip maps any longer? Instead what people are supporting is just throw a blur on it ASAP to reduce noticeable aliasing issues.

You can think of the effect of SSAA in the same way. It's not removing detail, it's removing error.

Why don't we all just run Quincunx all the time then? Seriously, let's just blur everything as soon as we can...?

When you turn on SSAA, your sampling is more accurate, so you can push back the mipmap boundaries much further - at the point where you used to see brown, you will now see somewhat "dirty" shades of green and red. Thus, SSAA can improve texture detail.

Put it at a 37% angle and try it

 

[Joe DeFuria]

Mip maps are downsampled blurs of the base texture, how can you say that isn't decreasing detail?

Did he say a smaller mip-map isn't decreasing detail relative to the larger mip-map/base texture? He's saying (from what I can tell) a smaller, downsampled mip-map isn't losing detail relative to a smaller mip-map produced via point sampling.

And he's correct.

Yes the moire pattern is an error, on a pixel by pixel basis it is more accurate then mip mapping it out and having 100% error rates on the pixels. That difference is one of what is less bothersome to people.

Not correct.

If 100% of the pixels in a downsampled checkerboard are the "blended" color, that means that 0% of the pixels should be 100% of either original color. So no matter whether the error is "blended" or apparent in sparkling or moire, there is still "100% error rate per pixel" (using your terms).

So a "point sampled" moire is NOT more accurate than a properly downsampled image.

Even when witnessing the horrendous moire that such an image would offer, at the very least a portion of the pixels inside what appears as moire will be correct. If you use mip mapping and change all the pixels to brown you assure 100% error rates.

Again. Wrong.

In this situation where we have a checkerboard base texture and in which the resultant porperly downsampled mip-map is such that 100% of the mip-map pixels would be some "blended" color (i.e. none of the mip-map pixels would actually be one of the colors in the base texture), then by definition NO INDIVIDUAL PIXELS should be 100% of any original base map color. So you are also assured of "100% error rates" with point sampling by your definition.

The accuracy involved is what is ar question. When does the blur itself produce a less accurate image then what could be maintained without it?

Never, if by blur you are talking about a correctly implemented AA technique, such as supersampling.

You can think of the effect of SSAA in the same way. It's not removing detail, it's removing error.

Why don't we all just run Quincunx all the time then? Seriously, let's just blur everything as soon as we can...?

Because (the blur filter aspect of) Quincunx doesn't remove error...Quincunx can in fact INCREASE ERROR because it takes colors from pixels that aren't necessarily sub-samples.

 

[BenSkywalker]

Joe-

Did he say a smaller mip-map isn't decreasing detail relative to the larger mip-map/base texture? He's saying (from what I can tell) a smaller, downsampled mip-map isn't losing detail relative to a smaller mip-map produced via point sampling.

His quote-

Adding mipmapping and bilinear filtering may make the texture contrast lower, and may look "blurry", but it does not decrease the detail of the render.

Adding mip mapping+bilinear does not decrease detail.

Not correct.

If 100% of the pixels in a downsampled checkerboard are the "blended" color, that means that 0% of the pixels should be 100% of either original color. So no matter whether the error is "blended" or apparent in sparkling or moire, there is still "100% error rate per pixel" (using your terms).

So a "point sampled" moire is NOT more accurate than a properly downsampled image.

Point filtering under any conditions I've used only works if it is capable of catching the colors represented by each pixel. I've been advocating anisotropic filtering, not point. I use point sampling and very simple textures for examples as it is the easiest way to illustrate it.

If you are arguing that 100% of the pixels would be a "blend" of color based on proper downsampling, then by definition NO INDIVIDUAL PIXELS should be 100% of either original color. So you are also assured of "100% error rates" with point sampling by your definition.

'Proper downsampling' is the issue. You tell me your choice- keep mip mapping and trilinear filtering along with SSAA to deal with textures or run 512x AF. 'Proper downsampling' is quaint and all, it serves its purpose until we can get some proper sampling without it suffering from down syndrome

Never, if by blur you are talking about a correctly implemented AA technique, such as supersampling.

Z. Why do so many people focus on x, y? A proper AA technique for 3D would be utilizing z values to calculate out its sampling patter and filtering. SSAA does not do this. If you take a texture and rotate it at a 35% angle and then sample it based on x, y you are blurring the texture. It is not being sampled properly.

Because (the blur filter aspect of) Quincunx doesn't remove error...Quincunx can in fact INCREASE ERROR because it takes colors from pixels that aren't necessarily sub-samples.

Exactly. And SSAA takes samples based on screen coordinates when they aren't necessarily the samples that should be taken. Unless you can assure that you have a perfectly level texture displayed on screen at the exact resolution the texture is natively then you are not sampling in the correct fashion. You are weighting different areas of the texture in an unbalanced fashion, extreme in some cases.

 

[Joe DeFuria]

His quote-

Adding mipmapping and bilinear filtering may make the texture contrast lower, and may look "blurry", but it does not decrease the detail of the render.

Adding mip mapping+bilinear does not decrease detail.

Decrease detail relative to what. I'm saying he meant relative to not having mip-mapping and filtering.

Point filtering under any conditions I've used only works if it is capable of catching the colors represented by each pixel.

?? don't understand that.

I've been advocating anisotropic filtering, not point.

Anisotropic filtering and "full scene" AA techniques are complimentary. I'm not really sure at this point what your main argument is.

Is it:

1) You'd rather have aniso vs. supersampling?

or

2) Supersampling can decrease detail / accuracy relative to not using it, all else being equal?

or something else?

I use point sampling and very simple textures for examples as it is the easiest way to illustrate it.

But if the illustration is not correct, what's the point?

'Proper downsampling' is the issue. You tell me your choice- keep mip mapping and trilinear filtering along with SSAA to deal with textures or run 512x AF.

I'd like both, actually. But given the choice, aniso for texture quality.

Z. Why do so many people focus on x, y? A proper AA technique for 3D would be utilizing z values to calculate out its sampling patter and filtering. SSAA does not do this.

Sure it does. Every pixel in a (for example) 4x (2x2) supersampled image (thus every sub-sample of the resultant pixel) is calculated using x,y, and z.

If you take a texture and rotate it at a 35% angle and then sample it based on x, y you are blurring the texture. It is not being sampled properly...And SSAA takes samples based on screen coordinates when they aren't necessarily the samples that should be taken.

SSAA and texture filtering are completely separate operations.

The result of applying SSAA is not any less accurate than not applying SS. (Apply SS to a point sample, bi-linear, tril-linear, or aniso filtered texture, and it's not less accurate than NOT applying SSAA to the texture.) Do you agree with that?

I don't see anyone arguing that Ansio doesn't provide more accurate results that supersampling AA on top of Trilinear. It does appear that you are arguing that applying SSAA on top of filtered textures can cause a lack of accuracy. Please clarify.

 

[MfA]

In the limit case a supersampling pattern should have even density across the entire pixel in image space (treating pixels as little rectangles for a moment). The sampling is only uniform in image space, if there is anisotropy it isnt uniform in texture space. Parts of the texture which contribute more to the pixel will have a greater number of samples.

In our case of small numbers of samples and (anisotropic) filtering the footprint of the texture filter will simply be bigger on average for the samples taken from closer to the viewer (they have a different texture LOD). So texels from the part of the texture closer to the viewer will have a greater contribution to the final colour.

Marco

PS. oops, that should be smaller of course