Thanks you two! I finally understand how it works, I just realized I actually never did understand how adaptive sampling works, amazing. Here is the equal noise comparison:
The one with Equianguar is faster now! It is also interesting that the slower one gets more diffuse bounce, I guess it is because it needs more samples to get to the same noise level therefore having more time to sample the bounces.
EDIT: Just hindsight, I somehow feels the one without it still feels like it has more noise? But I did see the sample count reaching the threshold and goes up crazily fast during both renders. Strange.
EDIT: Just tried a few more times after reading Alaska’s reply below, it turned out to be much mroe tricky since I now need to eyeball the two renders and adjust the adaptive settings, I just give up trying at this point, it really doesn’t matter I guess, it is for sure that Equiangular takes significantly less amount of time to reach the same quality. The specific number doesn’t really matter I think.
Looking at the images, it does seem like the render with “Distance” sampling (the method used instead of Equiangular) has more noise around the “light shaft”, but that’s more likely to do with the adaptive sampling settings you chose. Adaptive sampling tries to stop sampling the scene when it reaches a certain noise threshold. However, it can be “tricked” into believing it has reached the noise threshold in a portion of the image when in fact, it has not. This problem is more likely to occur in scenarios with materials that rarely sample the main contributor to their look. In this case, the distance sampling used by the volume is sampling the light less often than Equiangular sampling and thus fits the criteria to be one of the things that can result in the adaptive sampler stopping earlier than expected. The only way that I know of to work around it as a user is to increase the “minimum samples” option in the adaptive sampling section. However, this comes at the cost of longer render times.
Hi, the same om WIndows 10/11 and Linux with a RTX 2060 (470 Driver).
I always keep a running version because this happen with Alpha/Beta software sometimes.
I guess the developer will fix this next week. I hope.
Experiencing similar error. I have so far been unable to pinpoint the cause, but it does seem to happen quite consistently within the first few minutes of rendering or in the viewport.
Error is: “Misaligned Address in cuMemcpyDtoHAsync…”
Been playing with time based rendering. It bothers me that I have no way to know how many samples I ended up having. I set the time to x minutes, leave the computer, come back after x minutes. It shows me the render time was x minutes, well, I already know that. The problem is, how many samples exactly is this? I hope the end sample count will be displayed beside the render time for the time based rendering.
Still experiencing slowness while using Optix with two GPUs enabled.
My renders are experiencing this strange pixelating, specifically towards the edge of the image. Attached is an image of the left side of a 1920x1080 render where you can clearly see large squares reminiscent of the old chunk rendering method with drastically different color and noise appearances.
I found the reason for this bug. Turn OFF “Persistent Data” and the bug is gone (as a workaround). After this you will be able to change the render-sample-rate
Version of today (28 Jun) don’t recognize my 3090 as an OptiX GPU anymore (Win 10, 462.29 driver). Have to fall back to Cuda. Had this problem a few versions back (1-2 weeks) then it got fixed. Please find and exterminate this bug.
Hi, there is no bug fixing for this build until it is merged with the master.
I always use the .zip version and keep a working version.
One mention to use at least driver 470, cant remember if it was in this thread or the BA thread.
Getting less green with more samples. Have this for a few weeks now, even with newest 471.41 studio driver. Looks like a denoiser debug sample display setting left on. EDIT: greenish “error” only with “Color + Albedo + Normal” input passes setting.
Adaptive sampling on volumes renders longer with very noisy volumes.Very noisy. No matter how smallish the threshold, even on < 0.00001. On the plus side volume rendering without adaptive render a lot cleaner with relatively low samples since the Equiangular volume sampling update.
Just watched the first few minutes, I am not intelligent enough to understand the content so I stopped. But I do hear the words “spectral” “wavelength” or something a lot in the video, I guess this will be relevant to the Spectral Cycles project? I am not sure, this video is long and full with math equations, my impression might be completely off on what this is about
Is OIDN 1.4.1 already in pre-built libraries (Linux)? “config.h” file mentions 1.4.1 but I’m not sure if this is just a definition or something.
I just want to know to apply Sergey patch on Cycles X branch for testing. https://developer.blender.org/D12043
