Adaptive sampling: What am I doing wrong?

[xxiii]

Made a simple scene with sky and ground and camera pointed at horizon, slightly out of focus. adaptive-sampling checked, mitchell/blue noise

Using perspective camera (or real-lens, same problem occurs), If I set antialiasing to 32, I get a relatively blurry (but not quite perfect) horizon, with about 6 minutes render time.

If I set antialiasing to 5, I get about one minute of rendering time, and no matter what I set adaptive threshold to, or oversampling to, it doesn't seem to be interested in making any more than 1 or 2 extra passes, and I basically get a band of ground dots mixed with sky dots, instead of a nice blur effect.

I have tried setting oversampling to 0.0 and 3.0, and values in between. I've tried setting threshold to .075 and .0001 and values in between, none of these seem to make much, if any, difference. The most change i've seen is that sometimes it will only make one extra adaptive pass, and sometimes it will make two. I've also tried setting antialiasing to 2 instead of 5 (render times go down to 25 seconds, I think this is when it makes two extra passes, instead of one).

Its like the threshold thing isn't working somehow. The only way I seem to be able to get decent results is to crank antialiasing to 32 (looks good) or 64 (looks better), but then the render times skyrocket, and this of course defeats the whole purpose of adaptive sampling; but I don't know what i'm doing wrong with adaptive, and I've expermented a lot. I'm assuming I'm missing something extremely obvious (or it wouldn't be this hard to find), but what is it?

[PhantomPhish]

Do you have any geometry or is it just a sky/ground backdrop??

[xxiii]

Yes; I'm actually noticing the problem in various projects, this is just the simplest that one that shows it. My scenes basically look like they are shot on very grainy high-ISO film, and I tracked it down to a combination of radiosity settings (fixed), and this adaptive sampling issue. (The example scene I described above does not have radiosity enabled).

[xxiii]

Attached are some examples taken from masked renders from my current "lets learn hypervoxels" project. the mask is 60 pixels above the horizon to 60 pixels below the horizon, and is 1280 pixels wide.

oversampling was set to 0.25, threshold was set to 0.001.
reconstruction is mitchell, pattern is blue noise.
depth-of-field is enabled, with an f-stop of 8, and a focus distance of 1 meter. The horizon is 1 million meters away. The attempt at water droplets are around .5 ot 1.5 meters away (approx).

As indicated by the filenames, AA was set to 1, 5, 32, and 64.
rendering times were 3m15s, 7m2s, 37m37s, and 75m42s.
extra aaa passes were 4, 1, 0, and unknown (I presume it was 0, I litterally went to dinner during this one and didn't see it). Furthermore, based on the appearance of the progress bar, it appears to have made the same number of passes it was planning on to start with in each case (it appears it already knew it was going to make 4,1,0 and 0 extra passes).

[Exception]

Interesting. It seems the AS process is detecting contrast differences in the backdrop differently from geometry.

Can you send this to lw-bugs@newtek.com including your test scene?

[xxiii]

I forgot to mention that the sky is a half-sphere with an (exr hdr) image map, and the ground is the "floor" of that half-sphere. (all faces have of course been flipped so they are "inside" the sphere). There is also an unseen by camera,etc inner sphere for radiosity purposes (not currently being used). However, I've seen this problem without the inner sphere present at all, and in another scene that is in a completely enclosed room, approximately 5 meters by 10 meters, by 3 meters in size.

[xxiii]

A more controlled example:

In addition to the center of the image, note the white borders in the upper corners. at AA128, it finally starts to look like someone smeared something on your monitor.

At AA5, turning on adaptive made a slight -- but insufficient -- improvement; after that, AAA makes no difference. I also note that one can't set threshold lower than 0.0001 (or at least its displayed as 0.0, in any case, it didn't make a difference when I tried it).

An observation: The filesize of each image can be used as a crude measure of noise (the noisier versions compress less well, I used the same jpeg settings for each one).

[Costanel]

Have you tried setting AA to Fixed of Classic, if that fixes the problem?

Could be that Blue Noise doesn't respond to this kind of AA.

[xxiii]

Just tried Mitchell/fixed, and classic/classic. Also tried using image viewer instead of image viewer FP; same problem. Also, the documenation actually says blue noise will be used instead of fixed if for any samples beyond 64 (when using fixed). Thanks for the idea though.

Incidentially, I also filed a bug report, using the tri-color scene. I suspect there is some problem with the threshold setting and calculating how many extra passes its supposed to make. Debating whether I want to install 9.0 or 9.2 and see if it was working back then...

Ah, HA!, just tried turning OFF "use global" and NOW its working! Interesting, since there is no global setting for threshold (and none of those settings get ghosted out) and it still applies the other camera specific settings in this area, or at least the anti-aliasing one. (Only "adaptive sampling" seems to appear in the global settings, and it mirrors the camera setting (you click it on in global, then look at your camera it will be on, you click it off, and it goes off in your camera as well, and vs versa)).

Just looked in the scene file as well, and there is only one threshold definition. It appears if "use global" is checked threshold is ignored. Adding this info to bug report...
----
I'll get a life when its been demonstrated to be better than what I have now.