Chaos Corona Forum
Chaos Corona for 3ds Max => [Max] General Discussion => Topic started by: Jpjapers on 2024-05-15, 15:23:45
-
Does a higher res image require more passes to reach a certain noise level? Or does each pass just take longer?
For instance will and image of 1000x1000 and an image of 5000x5000 reach a 3% noise level at the same number of passes but each pass takes a longer time on the 5000x5000 image?
Ive never really thought about it before, i usually just let it run until it looks nice.
This help page is really useful but it didnt answer my question unless i missed it.
https://support.chaos.com/hc/en-us/articles/4528236666257-How-many-passes-is-enough
-
Actually, higher resolutions usually need less passes to visually "look as good", though I don't know how that translates into noise level. But of course it still takes longer to get there, due to the extra processing needed for all the extra pixels :)
-
Actually, higher resolutions usually need less passes to visually "look as good", though I don't know how that translates into noise level. But of course it still takes longer to get there, due to the extra processing needed for all the extra pixels :)
Interesting, thanks Tom. Just doing a lot more render farm work recently and trying to wrap my head around the pricing has raised some questions. this being one of them.
-
Less passes to look visually good, but also less passes for same noise, but never checked any formula :- ).
But this is good thing to pay attention, so you don't end up with too little AA done. Since high-res images don't need as many passes for GI/Light, they can for example end up being around 2-3perc. noise (visually clean) with only 25 passes for example, but roughly 100 passes are needed for really clean anti-aliasing from my experience.
-
Less passes to look visually good, but also less passes for same noise, but never checked any formula :- ).
But this is good thing to pay attention, so you don't end up with too little AA done. Since high-res images don't need as many passes for GI/Light, they can for example end up being around 2-3perc. noise (visually clean) with only 25 passes for example, but roughly 100 passes are needed for really clean anti-aliasing from my experience.
This is the sort of thing im trying to work on. Clients asking for higher and higher res images year on year and i need to get better at adjusting the finer render settings to suit the output rather than just leaving it to render for a nominal amount of time on my workstation because theres only so many hours in the day.
-
If you are rendering in hi-res (let's say 4k or larger), you can actually bump up GIvsAA to 32 or more. This way you will get less noise from GI and when rendering in such resolutions you don't "care" about AA that much so you can render a lower number of passes.
Quick thought experiment:
- imagine rendering a tiny image like 50 px - it will probably become "noise free" very quickly (i.e. there will be no large difference between pixel colors between each next pass)
- imagine rendering a huge image like 50k px - it will require a lot of time to reach acceptable quality at 1:1 zoom, but if you zoom out, it will probably look as good as the 50px image :)
-
Hah that's like the exact opposite I do :- D, but I think I just keep my default at 8/2 for almost any work (it's faster for IR/AI-Denoiser work too) and that gives me enough passes for 4-8K images (it also allows for more adaptivity re-comps)
I do find AA to be important always, imperfect AA is big CGI-giveaway to me, as cheap photo will be blurry mess, but it will never have that artificial edge 3D stuff has.
-
Ok, I guess that would require some testing. :)
-
Also I find that if AA is not near perfect, doubling the resolution with, for example, Topaz Gigapixel creates very noticeable artefacts where the AA was jagged. Just something to consider.
-
I wonder if perhaps a future feature request could be some sort of adaptivity region where we fan freeform draw a region around certain areas of a render that are noisy to place more importance. Or perhaps a noise region where a render will continue until a certain area reaches an acceptable noise level. It seems currently that the noise limit feature works on some sort of averaging?
-
The total noise amount is averaged, but adaptive sampling means Corona puts more processing power into areas that are noisy.
-
The total noise amount is averaged, but adaptive sampling means Corona puts more processing power into areas that are noisy.
Are there any settings we can tweak to adjust the adaptive sampler? Im finding that even after 500 passes some shadows are still noticeably noisy pre-denoiser whereas the rest of the image is crisp and clear.
-
The total noise amount is averaged, but adaptive sampling means Corona puts more processing power into areas that are noisy.
Are there any settings we can tweak to adjust the adaptive sampler? Im finding that even after 500 passes some shadows are still noticeably noisy pre-denoiser whereas the rest of the image is crisp and clear.
Does this happens with light mix? Adaptive sampler is calculated using beauty pass.
So if you have multiple hdri setup for example, and your beauty pass looks like moments after nuclear explosion. That could be your problem. If you have a shadowy place in light mix, that's brightly lit up in beauty pass. Then adaptive solver actually diverting recourses from that specific place, cause it sees it as noise free.
-
Yes, it happens with all scenes, but it's LightMix that can make it show up because you can turn off the intense light that is covering a place of the image (and that is noise free in that area when that bright light is on), to reveal that the other lights underneath have received less processing in that area and so it is still noisy there.
-
Yeah probably there's no good way to address this from dew side, without overcomplicating things. But from my testing I noticed, that in scenarios when this happens. It's actually better to disable adaptivity in general. That way at least noise is more uniform, but tested just a couple of times so don't know if there are some extreme drawbacks. Most of the time just disable all the unnecessary lighting setups for final renders.
-
One related issue we've always found with noise levels, adaptivity and AA etc. with resulting issues in rendertimes is when you're using super "busy" stuff in your scene like scatter carpets/rugs, grass or other such very fine and busy looking geometry. It seems to throw the noise/adaptivity calculation off to the point where we basically never use noise level limit in production in most scenes, and instead do a quick manual test where you just render the whole image and draw some quick regions in the areas that are obviously struggling to clean up quickly, typically dark areas or where there's lots of reflected light or something - keeping an eye on the number of passes until it visually looks good enough to finish off with some light denoising. Then we plug that value into the pass limit and call it a day. It's the only fool-proof way we've found to get the correct quality and predictable rendertimes without leaving things to chance. But if the adaptivity side of the engine was a bit more intelligent at figuring out that it's wasting a bunch of time/passes on areas that don't actually need it, that would sure be a nice improvement! :)
-
Well, one solution is to keep all lights at a similar intensity when rendering, so they all get roughly equal importance, and then do the changes in intensity in LightMix. Similarly, keeping all lights white at render time will allow for the greatest flexibility in color changes in LightMix. Does mean your Beauty pass will be useless, but if you are intending to use LightMix, that should not be a concern :) (replying to the post before Alex's)