Antialiasing with float

[timmatron]

So, now that I'm delving into the physical sky, camera, gamma, etc. I have a question about antialising. If you are supposed to leave sub pixel mapping and clamp output unchecked for proper results, how do you get better antialiasing around the bright edges like this?



I just watched Lele's tutorial so I'm basically trying to work as he suggested.

[joconnell]

Love the range of exposure by the way.

http://joconnell.com/view.php?category=3d&view=28

You're kind of screwed with something like this to be honest unless you use really high levels of sampling. A lot of people use a bit of glow around the edges to soften a tiny bit as this kind of bloom occurs quite often.

[timmatron]

Thanks and thanks.

That image was using the Vray sun at multiplier of 1, ISO 400 and the rgb value set to .25.

Compare that with this one where I lowered the ISO to 50 and the rgb to 60 (no math done, just random pick). The float value for the whites is much lower now which helps with the antialiasing. But the lighting looks like crap. This one was done with Irrad. map whereas the other was qmc, so its not quite apples to apples, but still the difference in exposure is very obvious.



I'll give your article a read soon. Hopefully some glow will help.

[joconnell]

yeah indeed it's a tricky one. The main problem is having a range of 0 to 1 in your viewable range (or 0 to 255 in rgb) but having really bright values in the sky outside. You might have a medium grey pixel with a value of 128 on your window frame but then the sky outside could be ten times higher than pure white or something like 2550! Depending on what anti aliaser you use, it'll try and find a halfway point between these two values and mathematically that's and rgb value of 1211 which is still 4.7 times brighter than pure white - it's a correct solution from vray's point of view but in terms of the look of the final image it's still way overexposed.

To be totally honest, float is very nice and handy but if you're not doing any massive colour correction in post after your renders, it might be a bit more trouble than its worth as you may find it extremely costly to get rid of those edge artifacts. If you're close enough in the straight render then clamp output and sub pixel mapping might be the easier way to go at the expense of flexibility after.

[percydaman]

its easy to add bloom effects in post where overbright areas would occur. Its nearly impossible to remove them in post when they come from the renderer. I think float is a little overrated except in certain situations and when maybe rendering out everything in layers to be composited later.

[timmatron]

Ok, that's a good enough answer for me. We usually don't using floating point for stills anyways so thats fine. Thanks guys.

[instinct]

erm...just for the record...FLOAT is never overrated...having values above 1 is (mind you there are more values between 0 and 1 then between 1 and infinity with floating point images :P)

Anyways there IS a big difference depending on the post process...and i would LOVE to be able to have bigger dynamic ranges wich is impossible due to aliasing trouble. Float helps...but having reflections at the right proportional value to the rest would help TONS in post (mind you we're doing 99% cars)

Kind Regards,
Thorsten

[percydaman]

sorry, your entitled to your opinion, but I disagree. We all did just fine without pure float for the longest time. Like I said above, if he renders in pure float, he gets overbright areas that have issues with antialiasing. Light correctly and there's plenty of pixel info there in the image to play around with in post.

[instinct]

Well i am not referring to values over one as stated above. Float allows for values over one, but it doesnt need them. And even if you clamp your output to 0-1 float is making HELL of a difference in post. Values over 1 (and i mean a LOT over 1) would be an additional help as they allow for some post-tricks that need workarounds without them (proper glinting beeing one of many examples). And hey, i did just fine without a car for 18 years...but hell i dont wanna miss it now

Regards,
Thorsten

[studioDIM]

I think, and Thorsten will correct me if i am wrong, that he is referring to the precision that comes with float formats.
Purely because 0-1 is a decimal space (0.567893 is a mid-gray with a slight bias on white), whereas the 8 bit formats constrain values to 0-255, integer (so a mid-gray is either 127, 128, or 129).
That makes a HECK of a difference when keying, or applying any math to the pixels.
For instance, my first exposimeter (the unmanned freight train) had to keep re-exposing as it just measured values between 0 and 255 in 8 bit.
Anything above 255 could not be taken into account, so the script had to guesstimate the next exposure step (down, in this case) and re-measure.
Going float allowed for pinpoint precision in only one go ('course, also because super-whites have a meaning in float format, but not only), thanks to the greater precision the format allows for.
Another great example is the vray color map.
Sure, colours are always looking the same on screen (device-dependent, nothing that can be done bar, maybe, buying a hdr monitor, yuk!), but in many situations the ability to specify them with float precision helps fine tweaks, to a degree that wasn't achievable without...

Also, float values are everywhere in max: a light multiplier is a float value, keys are described through float values, and so on.
Think what would happen if we had 8bit integers to define a camera position, lol.
Only 16.8 million possible places: i know some chaps that by now would have well exhausted the possibilities

Lele

[joconnell]

Or even just the added benefits of having proper highlights when you defocus something so you get this



instead of this

[/img]

[percydaman]

ya for that and yes having easier done glints and glares, float is nice, but how are you going to resolve those issues of AA when the client neither likes or needs nice DOF highlights/bokeh or could care less for glints and glares. All your left with is blown out portions of your render that wont anti-alias properly. Now where does float get you?

Like I said above, float is great in certain situations, namely when it doesn't cause render quality issues. Or for when rendering out numerous passes where you have better control of render output and how it affects compositing.

[werticus]

if you turn on clamp and subpixel, you'll just get aa and the image will look the same... you just lose the control of exposure later with HDR files.

so if you can get it looking perfect in the viewport you'll get the aa as well with clamp and subpixel. after all your monitor cannot display these values

[joconnell]

Bang on there percy, I'm with you on that one - I still haven't used float on a job full stop but I will be looking at it on a test car commercial I'm doing - If anything I'll be hassling thorsten an awful lot about their workflow on the mercedes stuff

At the moment I only do 3d stuff in here and pass it to flame which can't deal with float at all so I'm still doing standard 8bit images - the anti aliasing issues wouldnt be worth my while dealing with since no one could use the float data anyway. What would be nice though is to have some kind of second pass where you have a non clamped version of the colours over 1.0 so you could still get that data to drive stuff like blurs and glows in the comp - remove the aa issues but still have the higher values for the nice glints and what not.

[winberg]

Bang on there percy, I'm with you on that one - I still haven't used float on a job full stop but I will be looking at it on a test car commercial I'm doing - If anything I'll be hassling thorsten an awful lot about their workflow on the mercedes stuff

At the moment I only do 3d stuff in here and pass it to flame which can't deal with float at all so I'm still doing standard 8bit images - the anti aliasing issues wouldnt be worth my while dealing with since no one could use the float data anyway. What would be nice though is to have some kind of second pass where you have a non clamped version of the colours over 1.0 so you could still get that data to drive stuff like blurs and glows in the comp - remove the aa issues but still have the higher values for the nice glints and what not.

Isnt this what the "unclamped colors" render layer is for?...this have actually been around for a while in the rpl/rla file format..long before all this linear/float "hype"......

-Tom