[brodie_geers]

I used Studio to export my physical sky to an HDR and the results are looking quite different.

First image: Physical Sky


Second Image: IBL with all channels disabled except for illumination


Notice how in the HDR version the lighting is brighter, the colors are more washed out, it's noiser (both reached SL 12.0), and there are no caustics shining down from the glass onto the sidewalk.

Any ideas about this? I was hoping to use the HDR to soften the shadows but the lighting quality is too inferior to use that as an option.

-brodie

[Mihai]

Are you using Physical sky for the disabled channels? What happens if you turn on the HDR for all channels, do you get the caustics back? I've noticed the HDR from physical sky can look brighter, but if you lower the ISO in the render a bit you should get back the same lighting. It may not be exactly the same color in the render (less saturated) but I suppose this is because of how the physical sky model works, it's values can't be converted exactly into an HDR.

[brodie_geers]

I ran another, more reproducible, test. Here's the process:
open studio
import plane (scale 10,10,10)
import knot (move up above plane slightly)
apply default material to plane
apply chromium wizard material to knot
create camera, default settings
set target SL 12.0
Production Quality

Test 1
default physical sky settings



Test 2
Save physical sky to .hdr 2000x1000
IBL: disabled background, reflection, refraction (disabled = physical sky); illumination channel uses .hdr created from phys sky



Test 3
IBL: .hdr used in all channels



Test 4
Create new IBL from physical sky 6000x3000
IBL: .hdr used in all channels




Conclusions
Using hdr created from physical sky...
...there is noticeably more noise at SL 12 (converting phys sky to .hdr used to be faster, it seems to be the opposite now - see below)
...it is darker (bumping the iso up by 10-20 got it in the same neighborhood, but i left these tests unadjusted)
...the colors are different (testing a portion of the ground plane in photoshop phyiscal sky HSB = 343,2,44 RGB = 111,109,110; IBL HSB = 292,4,41 RGB = 104,101,104: notice how the Hue shifted from more of a redish into more of a purplish)
...no caustics


Noise Comparison Physical vs. IBL (in that order)


Comparison Gif (may take a second to load, but it should swap back and forth between phys and hdr lighting image - cropped and enlarged)


-Brodie

[3dtrialpractice]

i dont know about the casutics.. but i think the general rule for the ILLUMINATION channel is to have a small HDR resolution (this will also probably give you the softer shadows u were looking for) try setting the hdr res to 128x256-even 64x128 maybe it will be a cleaner render... (obviously if you had a detailed hdr pic not one of just a sky gradient the resolution for the reflection channel would be another HDR file w/ higher res)

[brodie_geers]

i dont know about the casutics.. but i think the general rule for the ILLUMINATION channel is to have a small HDR resolution (this will also probably give you the softer shadows u were looking for) try setting the hdr res to 128x256-even 64x128 maybe it will be a cleaner render... (obviously if you had a detailed hdr pic not one of just a sky gradient the resolution for the reflection channel would be another HDR file w/ higher res)

More tests along these lines only show more odd behavior. I performed a series of tests with all the channels disabled (to phys sky) except illumination and saved out the physical sky to different resolutions.

HDR 582px wide (2:1 Ratio)



HDR 581px wide (2:1 Ratio)



Anything 582px wide and larger turned out pretty much like the first image (it seems to get SLIGHTLY brighter at higher resolutions as can be seen in the 6000 px wide version above). Anything smaller than 582px wide pretty much turned the lighting to mush. Basically everything appears to be in shadow (you'll notice there are no noticeable shadows and the whole ground plane is about the same color as the color of what's in shadow in the other image). The change is quite dramatic and sudden. I had to do a lot of tests to find out exactly where the turning point is. Whether 582px is always the magic number (or WHY there's a 'magic number' I don't know).

Also noise levels are pretty much the same as when I was using higher res .hdr's

-Brodie

[brodie_geers]

Testing further, I tested those 2 resolutions again, this time with the .hdr chosen for all channels, not just illumination, and have found what happened at that magic number.

If you look at the images above, you'll see the small specular dot in a couple spots on the knot which is the reflection of the sun itself. At 581px the overall reflection is pretty much identical but that small specular dot is gone. So apparently saving the phys sky out too low of a resolution means that there's not even a full pixel for the sun to sit in, at which point the sun itself is no longer there, just the rest of the illumination. So maybe if you saved it at a higher resolution and downsized it in Photoshop it would be able to account for this?

-Brodie

[brodie_geers]

OK....another test. Exported HDR to 2000x1000, resized in Photoshop to 200x100.



Good news is that the sun remains and does produce shadows, even soft shadows. However noise levels don't really seem improved, the image is darker and bluer than the previous .hdr images (bumping ISO from 100 to 150 in the 200px image got lighting levels close, but the bluish shift is more noticeable). And worst of all you'll notice some white snowflake like artifacts which were introduced.

-Brodie

[Half Life]

These are all things that I can confirm from my tests of thin SSS a little while ago -- the noise does clear (with higher SL) and the brightness change can be compensated for via exposure but the worrisome element to me in all of this is the extreme reduction in the strength of reflective caustics.

Thanks for sharing these Brodie!

BTW, are you using the interpolation option on any of these?

Best,
Jason.

[brodie_geers]

None of those had interpolation checked.

Here's the 200px test with interpolation.


You'll notice the banding (most noticeable in the background) is reduced but otherwise the image (including the artifacts...although they seem to be fewer) are pretty much unchanged.

-brodie

[Half Life]

I will say (from my own tests) that:

1) Downrezing the EXR/HDR is unnecessary in Maxwell and does not result in much if any benefit in any IBL channel.

2) EXR is a larger file size but will give better results.

3) To come to any useful conclusion you need to render to at least SL 18 -- in my experience any late converging elements in the render (due to priority optimization I would suspect) start to show near that point in the process... SL 12 is entirely too low to be of any use other than preview.

Here's a link to my thinSSS tests:

http://www.maxwellrender.com/forum/view ... 97&t=35734


Best,
Jason.

[brodie_geers]

Here's a test all channels at 1000x500 hdr IBL at SL 20. You'll notice that there aren't really any improvements beyond noise reduction.



-Brodie

[Half Life]

Actually the detail is much improved over any of the previous (look for the highlights in the reflections on the metal) -- Exposure-wise the difference is easy to adjust for... the thing that has always concerned me is the lack of reflective caustics.

Best,
Jason.

[brodie_geers]

I'm not seeing what you're talking about. Of the previous tests it's pretty close to Test 2 in it's setup. Comparing those 2 renders I don't see much any difference. Even if you compare it to the 582px render I'm not really seeing anything note worthy.

-Brodie

[Half Life]

The down-sampling made your highlights too chunky/blocky in general (the interpolation helped slightly) -- and all of the others look a bit unfinished (noise or not) in the metallic part because of the lack of reflected highlights on the areas away from direct sunlight (except in the down-sampled versions).

Those subtle details are the difference makers that elevate the finished product -- and if you were not able to get those accurately with IBL vs Physical Sky it would be telling that something was wrong in that area as well.

There is almost always some small subtle touches that come out in the render in the higher sample levels and if one method gets those better (or at all -- as in reflective caustics) it would be preferable.

It is also worth testing to see if you begin to get those reflected caustics at higher sampling levels -- I've seen reflections not fully develop until SL 25 and sometimes higher... but then the question of is that practical comes into play.

The bottom line question for me here is: Are there no reflective caustics or are they optimized to come in at some high sampling level with IBL for some reason... the idea has always been with Maxwell that given enough time it will converge at the correct solution -- I find myself doubtful in this instance.

I'm currently rendering a glass at SL31 and have a couple more sampling levels to go before calling it done or I would give it a whirl

Best,
Jason.

[brodie_geers]

I'm currently rendering a glass at SL31 and have a couple more sampling levels to go

You're silly

I'll probably run a test when I leave for the night and see if any caustics come up. I'm doubtful as well (the caveat to the unbiased render motto, is that it only converges on the proper solution if there aren't any bugs), but even if they do develop, I'm dropping this into the bug bin because there shouldn't be such a broad gap (caustics showing up at SL1 with physical sky and at, say, SL 25 with IBL).

-Brodie