Speaking of improving and stabilizing - there is one thing which bothers me for quite a while.....
Could you send us instructions how to reproduce this? You could contact us about this at https://coronarenderer.freshdesk.com/support/tickets/new
I read your post, but still did not understand fully what was done here, and how to reproduce this. Just two guesses:
1. Maybe it has something to do with the wrong UHD Cache settings in 1.5? Have you tried this in 1.5.2? ("reset settings" in scene setup is required for the UHD Cache to work correct if the scene was previously saved with wrong settings!)
2. Did you drastically increase or decrease light intensities in LightMix? If so, then this is expected. LightMix should be used mainly to change lights' color, or to change their intensity only subtly.
Hi Maru, well, the easiest way to reproduce it would be to download those scenes I attached to that post, there are both default sun and sky value one and the one with lowered ones but every (and I mean every scene not necessarily made by me)) exterior + interior scene with lights setup using the real world values which manufacturers often provide would be good for that test too
1. I would suggest it's not a bug at all + it has been there for ages, it's just that it's easier to see it now because of the lighmix being added. So to make it as clear as I can - that's not some corona version specific bug but rather a way corona's gi currently works, giving some sort of sampling priorities to the brighter light sources so and sun and sky being way brighter than the others (usually 1000times brighter than the rest of the physically plausible setup light source) leads to that all the time, but even if i'm wrong and it is a bug, then it was present in all prior corona versions too, but I've been able to spot it clearly only when the first pre 1.5 lighmix containing builds showed up and of course it is present in 1.5 , 1.5.1 , 1.5.2 and even in that early 1.6 daily.
quote "reset settings" in scene setup is required for the UHD Cache to work correct if the scene was previously saved with wrong settings!" did that too, even if I did that after doing those tests on both 1.5.1 and early 1.6 - no difference there, as this is most probably not UHD cache related thing at all, although after resetting, it sure did a better job on those splotches in the corners of that "room", even way better than setting uhd precision to 10 in the previous builds which I actually had to do when doing those tests, but still, it does not effect that patch tracer produced noise at all.
2. well I had decreased sun and sky's values in lightmix to get some sort of decent sunlinght intensity without lowering exposure (which would kind of ruin the test's purpose) and increased those when doing test with lowered sun&sky values to compensate for that initial intensity decrease and of course set sun's pass to 0 and decreased sky value and added some dark blue color to it in both cases to get those night like renders. If I didn't completely misused the lighmix the whole time )) it is imo the whole point of using it, well, to get those nice night\evening with artificial light and day (without the one) renders out of a single render ) Still decreasing or increasing lighmix channel values does not effect the noise amount at all + I rendered all of them with the beauty pass active if that makes any difference at all + I think there should be any difference either as if I got it right the lighmix is kind of more like mixing different layers in photoshop than anything else right now so there should be no impact on any kind of sampling and only the initial light source intensity does have that impact there which the test proves quite nicely, well of course, if I wasn't completely stupid and did some incredibly stupid mistakes in the whole overthinking of that process )) Well you surely begin to see some "undercooked" layers when you lower the lighmix value for some bright light source which otherwise obscures the rest of the not as bright ones thus masking their noise but that was the whole point of that test - to see how the changes in the initial brightness values of the sun and sky (not the ones set in lighmix but the "real" ones)) effect the gi sampling in general.
So trying to sum it up - decreasing the initial (pre render not lighmix's) intensity values of corona's sun and sky to somewhere around 0.01 can have great benefit for the sampling of interior lights when render is done from the outside of that building model which is especially apparent when there's some glass in windows etc. , thus, greatly increasing the noise coming from those interior light sources cleaning speed in such type of the scenes. It can be seen only when one is doing some lighmix adjustments to the initially daylight+sun lit renders making those renders to represent different lighting conditions afterwards, namely doing some night or the evening type of lighting, if there are no interior lights or if the sun and sky values were initially set to represent the night\evening scene there's no difference at all which is kind of expected.
And finally, to better illustrate the benefit of that trick\needed tweak of corona's PT gi system I post the first results I got using that technique on a production scene I worked on earlier. Those tests with a room full of teapots, what I posted before, were done after these in hopes of clarifying\proving that benefit in a "clean" and more or less controlled scene. So here goes:
1 the render using default values of corona sun and sky which rendered for around 45 minutes
https://www.flickr.com/photos/119850875@N05/31027982325/in/datetaken-public/ - daylight lighmix preset
https://www.flickr.com/photos/119850875@N05/30725862970/in/datetaken-public/ - night lightmix preset - notice the ugly noise inside the interior which didn't go away even after I rendered some region of affected area for additional time almost doubling the pass count for that part
https://www.flickr.com/photos/119850875@N05/30991942826/in/datetaken-public/ - sampling focus pass for that one
2 the same scene with slight adjustments with some additional exterior light fixtures and slightly different view angle (nothing else changed and I doubt that adding those 2 lights would benefit the render speed at all, well, this scene was not intended for testing purposes so you know how it goes) rendered using lowered to 0.01 sun and sky values for around 15-20 minutes (pay attention to that)) when bringing those values up back using only lighmix after the render was stopped.
https://www.flickr.com/photos/119850875@N05/31027971085/in/datetaken-public/ - daylight lightmix preset
https://www.flickr.com/photos/119850875@N05/25391535069/in/datetaken-public/ - night lightmix preset - notice the ugly interior noise gone ))
https://www.flickr.com/photos/119850875@N05/30726015030/in/datetaken-public/ - dusk\evening lighmix preset, just for the sake of it )
https://www.flickr.com/photos/119850875@N05/30992083856/in/datetaken-public/ - sampling focus pass for that one
So you can easily see why am I so agitated by my findings and if those can be proven and reproduced by someone else it could mean what some slight adjustments to corona's gi behavior could make it much (can't calculate how much but A LOT) faster than it currently is when dealing with such a scenes. To make things better, daylight + artificial light interior scenes like this benefit from this trick as well when rendered from inside the interior itself, yet I have no images to prove it cause I hadn't done some specific test for those and the one's I have rendered were done on that same production scene after using that trick, but the interior itself was changed quite a bit since the times I rendered the same views without this trick, sadly have no time to do more tests right now but still it would be good if anyone could reproduce it, well at least to prove what I haven't gone mad completely ))