PDA

View Full Version : PLD antialiasing... what's the dif?



fabmedia
09-30-2005, 10:44 AM
What is the difference between the different AA methods? I've had LW 8.2/8.3 for some time now, but I really can't see much of a difference between the various options (Box, Gausian, Mitchel and Lancoz). Is there just a speed difference with reconstruction?

I couldn't find anything on these options on the threads.

Mylenium
09-30-2005, 01:50 PM
What is the difference between the different AA methods? I've had LW 8.2/8.3 for some time now, but I really can't see much of a difference between the various options (Box, Gausian, Mitchel and Lancoz). Is there just a speed difference with reconstruction?

I couldn't find anything on these options on the threads.

Reconstruction will try to re-introduce information that has gotten lost e.g. due to too much oversampling on a pixel. So for instance a pixel could become too transparent with motion blur. The algorithm would then compare it to other pixels/ ray intersections or even the pixel as it was before oversampling and then apply some corrections. This may sound like a purely image based approach, but it's not. It's more like a weighting on how certain rendering routines do their stuff based on the full precision buffer data. This effect is more pronounced with some algorithms and less with others. In general you will only see an effect with the "usual suspects" - fast motion, and hard contrast. On most setups the difference is not noticeable, though it could still be verified technically. I'm sorry that I can't explain better, I'm not that much of a tec guy.

Mylenium

fabmedia
09-30-2005, 02:06 PM
Are you saying that the PLD is not full precision and is a totally different rendering scheme?

Mylenium
10-01-2005, 03:04 AM
Are you saying that the PLD is not full precision and is a totally different rendering scheme?

No. Using Reconstruction filters does not mean that an AA method is inferior to another. Actually it's quite common in other programs and Lightwave just has not had it for some time. LW's classic AA is a (pseudo) ordered method whereas PLD will divide your image into a grid and analyse pixels based on the intersections/ knots of this grid. The grid itself is displaced meaning that ideally each of the intersection points is at a different position each time it is sampled. This is more random and is supposed to avoid certain problems such as Moiré patterns etc..

Mylenium

Lynx3d
10-01-2005, 07:52 AM
The whole rendering process is always a reconstruction of the scene.
You need a function to position the samples, and obviously a function to reconstruct a pixel from a number of samples. The trivial ways can still be seen as reconstruction too.

PLD places the samples at different positions than the classic method, and the new filters use a different function as the classic filter to combine them to a pixel.
The box filter is the most trivial one, it just averages the samples within a certain area. "Classic" seems to be just a box filter to me, i don't see any difference.
The others weight the samples depending on their distance to the pixel position. Lanczos is closest to the "ideal" reconstruction function, the sinc function which would perfectly reconstruct the image if the sampling theorem would be fullfilled. But because that is basically never the case for geometry, it tends to amplify artifacts on sharp borders, causing "ringing". But with properly filtered textures for example it should reproduce more detail without artifacts.
Mitchell tries to find a compromise between reproducing detail and avoiding "post-aliasing" such as ringing. Gaussian tends to blur things, no surprise as we all know the function gets used for "gaussian blur".

I think the normal, sharp and soft variant only vary in the filter size.

toby
10-01-2005, 09:48 AM
BUT... can anyone show me a PLD aa'ed image that looks better than a classic one? :hey:

fabmedia
10-01-2005, 09:56 AM
Hmmmm… lately I've been using the Gausian method for all of my work and find it very useful. Mind you maybe it's just me, but I think it works the best. I was doing some tests last night until 1:30 in the morning, and have noticed that when you hit 9 passess with Gaussian, it is better than Enhanced Medium (both with dithered blur), and Lanczos comes in a very close second.

Why I'm asking all of these questions is that I am looking at producing an animated short approximately 4-6 minutes long at 720p resolution, and thinking that it's going to take a **** of a long time to render out the scenes (even with a small renderfarm of a dual 867 G4, a 400 G4, and a 333 G3), and I've been tossing the idea of using a third party filter to do the motion blurring (ReelSmart Motion Blur). Hence my question.

Lynx3d
10-01-2005, 11:51 AM
Neither PLD nor the new filters do anything substantial to improve LWs poor motion blur...because it does not influence temporal sampling.
With a proper QMC sampling things would probably look different (same for DOF)

fabmedia
10-01-2005, 12:12 PM
I find that you can get away with 2 passes less with Gaussian AA.

toby
10-02-2005, 02:23 AM
Well, considering the power of your render farm, I can understand your desire to speed things up - I forgot some of us haven't upgraded to G5 yet :D
And with 9000 frames to render....

But you are definitely taking a quality hit. Keep in mind that good anti-aliasing is even more important with animation than with stills. Without it you'll see the jagged edges in motion, which is much more eye-catching.

I'd sooner give up all raytracing than good aa, actually there's almost no raytracing in all the Pixar and Star Wars movies, but the aa is always perfect.

fabmedia
10-02-2005, 02:12 PM
By raytracing, you're talking GI? I don't use GI unless necessary. I've found with Fprime that my lighting is getting better all the time (as long as I have immediate feedback). 9000 frames is a lot, but I'm hoping that I can get some $$$ through a couple of funding areas that are offered by Film Canada, Bravo, Global, and YTV. If I can pull that, then I can boost my renderfarm with 3 or 4 few Mac Mini's and a used G5. But I'm not counting on that. I'm just hoping that I can get through everything. I've been looking at ways to cut back on my rendering times by baking lighting solutions on to all of my objects where I can as to alleviate the render hits, but I don't want to mess with what I'm familiar with unless I can guarantee my results. BUT I'm starting to love multipass renders. I just have to learn the full power of compostiing in Combustion with multiple layers. I haven't figured that out yet. This is one of the reasons why I was looking at a third party motion blur solution. It would really cut back the time for rendering. Anyway, I'll post some samples later.

Captain Obvious
10-02-2005, 03:07 PM
By raytracing, he is talking about raytracing. Reflections and refractions, most notably. They really take a while to render, and it's very common to not use them in feature films. Pixar rarely uses raytracing, as far as I know. Instead, they render at insane resolutions and use perfect anti-aliasing and more complex lighting, more natural texturing, etc etc etc. That really makes for a better result than a low-resolution rendering that's poorly anti-aliased and has low-resolution textures, but full raytracing.

fabmedia
10-02-2005, 03:17 PM
rarely uses raytracing, as far as I know. Instead, they render at insane resolutions and use perfect anti-aliasing and more complex lighting, more natural texturing, etc etc etc. That really makes for a better result than a low-resolution rendering that's poorly anti-aliased and has low-resolution textures, but full raytracing.

Okay, so how do you create reflections then. For instance, a boat in water? I know they are render intensive, but you can't get away from it.

toby
10-02-2005, 03:52 PM
Mapped reflections instead of raytraced.

It's a lot of work, and you'd have to get creative with Lightwave because it has few mapping options - mostly spherical. For a boat, I'd render the bottom of the boat straight up then... perhaps apply it as a texture on the water, or render the the boat first then flip it... something like that.

But raytraced shadows is the other killer - shadow maps are what's used in the afforementioned movies. Lightwave's ratrace shadows are much faster than Renderman's, so they can be practical, unless you want soft shadows. Doing that with raytracing get very render intensive.

toby
10-02-2005, 04:12 PM
That said, I don't think there's a need to eliminate raytracing in your case, just consider cutting down on the newer features that require today's faster machines to render in a practical timeframe, instead of cutting down on aa. Baking is a great idea, and rendering to passes and compositing is even better. If you render your 'beauty' and shadow passes completely separate, you'll have a lot more flexibility in making good looking shadows, never having to worry about re-rendering the beauty pass because of them. And you don't have to render any separate passes that you don't want to - like breaking out diffuse shading from the rgb pass - it's not neccessary if you don't have a plan for it.

fabmedia
10-03-2005, 01:23 PM
Okay here are some rendered images. The model is not mine but Christopher Dessey's

badllarma
10-04-2005, 01:44 AM
Hmmmm… lately I've been using the Gausian method for all of my work and find it very useful. Mind you maybe it's just me, but I think it works the best. I was doing some tests last night until 1:30 in the morning, and have noticed that when you hit 9 passess with Gaussian, it is better than Enhanced Medium (both with dithered blur), and Lanczos comes in a very close second.

Why I'm asking all of these questions is that I am looking at producing an animated short approximately 4-6 minutes long at 720p resolution, and thinking that it's going to take a **** of a long time to render out the scenes (even with a small renderfarm of a dual 867 G4, a 400 G4, and a 333 G3), and I've been tossing the idea of using a third party filter to do the motion blurring (ReelSmart Motion Blur). Hence my question.

The VERY best way of speeding up this type of work IMO is just to comp EVERYTHING background characters etc etc...

For say a shot at a location i'd render the back ground as one frame than comp everything else over it even if your rendering the back ground as a much larger image (to allow for camera movement in the shot) it would still be much faster then rendering the entire scene in one hit.

Unless you need dynamic interaction of course.

Just my 2 cents anyway I would really start thinking of each shot as fore, middle and back ground and I'm sure you could save a lot of render time.

fabmedia
10-04-2005, 10:51 AM
Yes compositing is the way I'm going to take the project. But I don't think it'll render everything out faster will it?

operation
10-05-2005, 04:24 PM
those modes depends of your needs (I switch between thoses modes often).
I use much these modes. but it is with the experiment that you will choose the good one


If :
you have a lot of motion blur.
you want precise antialising ( wireframe render).
pictures for TV
for print
too much blurred texture
Sharp Alpha_channel
Spline or 2 polygons line render.

if you want to see the difference.
make a sphere.
render it as a wireframe.
test different mode ! (Gaussian, etc.. )

look at the edge .. sometime it's blurred a lot, sometime the spline is perfect, sometime it's sharp.

Box blur is a fast blur but ugly.
Gaussian is more soft ( nice blending color but blurry)
Mitchell will be more sharpen.
Lancozs close to the Mitchell but with a different blurring "curve"

make tests as I told you with wireframe or contrasted pic (black and white : white sphere - selfillum on a black background). you will understand :) and don't forget to check alpha channel too (there are differences)!

toby
10-06-2005, 01:46 AM
Yes compositing is the way I'm going to take the project. But I don't think it'll render everything out faster will it?

It's faster in it's flexibility, if you don't like the results in a hypervoxel pass for instance, you don't need to re-render anything but the HV. Rendering things twice because they're in the same frame is a big time-waster. Things are also easier/faster to adjust and add effects to when you're compositing.

fabmedia
10-06-2005, 10:03 AM
So with the AA, should I be more wary about the final results of what the "audience" is expecting? Or will a small bit of blurring okay?

toby
10-07-2005, 12:03 AM
Well it's really up to you, I can't recommend anything but the highest aa, enough to get rid of all jaggies - it's an easy way to help make your images look great, more realistic and pleasing to the eye. Watching a blurred animation makes you try to focus your eyes, which is fatiguing and uncomfortable. The minimum I'd use is enhanced medium, and only if I had a really short deadline and it didn't have to look great. I don't use PLD at all, fine lines are hard enough to anti-alias, and PLD performs pretty poorly in that category.

If you want to really see the difference, render an animation at low aa and another at extreme. The one done at extreme will look buttery!