Noise Threshold is BURNING Your PC in Blender
VloĆŸit
- Äas pĆidĂĄn 9. 06. 2024
- This video dives deep into optimizing Blender's noise threshold and adaptive sampling settings. Learn how to balance render time and quality by understanding key parameters like noise threshold, maximum, and minimum samples. Whether you're using a top-tier graphics card like a 4090 or working with limited resources, these tips will help you streamline your workflow. Don't forget to subscribe for more Blender optimization tips!
Videos used in this video to explain concepts and their creators (subscribe to their channels đ):
âȘ@DerekElliott⏠: âą Brighter Shadow Causti...
âȘ@cgvertex⏠: âą Blender cycles continu...
âȘ@asokerov⏠: âą masking the reflection...
âȘ@laviestudio⏠: âą Screen Space Global Il...
âȘ@Olav3D⏠: âą [2.93] Blender Tutoria...
âȘ@lanewallace.⏠: âą The Best Volumetric Fo...
#blender #3d - ZĂĄbava
Based on my underestandings , noise threshold is the aspect ration between Signal and noise , for example a value of ( 1 ) noise threshold means the aspect ration between signal and noise would be 50% , 50% , so if u decrease the noise threshold , you will increase the aspect ration of the signal compared to the noise , so u get higher quality image
The noise threshold is the maximum variance accepted for each pixel, it's not a ratio
@Lucas72928 i dont quite underestand what you mean
That is stil aspect ratio aint it?
@@aliosanlou4425 aspect ratio is a measure of how stretched a rectangle is. "Ratio" is just a division. The variance is related to probability and it's not a division
adaptive sampling (aka noise threshold in blender) stores half the samples in a separate buffer (a separate image if you will), meaning there's now two versions of your render in memory: one with all the samples (cleanest version) and one with only half the samples (noisier). the pixels of the two images are compared, and when the difference is lower than the value you set in the "noise threshold", that pixel stops being sampled
âââââ@@Lucas72928 it is a ratio, if noise is 1500 and noise treshold is 100, you doing 100/1500 to decide how many samples you should reduce in this chunk of image. in this example this is 0.06 aka 6% of samples.
if it is 100/100 then you have 1, that means 100% so render in this region is ended.
it is really bad for caustic, it creating lumps of caustic pixels because caustic noise is spatially bigger than other noise types.
They even renamesd Samples to Max Samples to make it clear it's not the same setting as pre 3.0 Cycles. And yet the "reduce samples" mantra has not changed on Reddit, it's the ancient "wisdom" that just won't die.
when using denoise oftentimes you'd want to disable adaptive sampling outright, as non uniform samples will confuse the denoiser and make it effectively hallucinate detail and create strange artifacts
This 100%. Usually in darker scenes with out of focus details. Usually you can resolve this by adding a minimum sample count though, so it samples uniformly at the start, the adaptive cleans up noise.
Had this happen on a cloth/fluid sim
it also frequently throw up volume scenes
0:28 You should have said holy grain
This is a really well thought out and informed vid. May you have many many views! Thanks for sharing and taking the time to put this together so well. Your efforts to make it understandable have truly paid off, I can promise you.
What a coincidence, just two days ago I watched your sampling video and was waiting for the promised noise threshold. Thank you!
I just render in the lowest sample rate with "good enough" noise, apply denoiser and use a scene grease pencil to keep the sharpness.
I am just lazy so I use render artifacts as my style. It made my job a hundred times easier that I may return to CZcams at some point idk.
Very helpful information thank you !!!
I find it super helpful to increase the minimum number of samples particularly in scenes that have an overall darker color scheme, and lots of shadow. I had an indoor scene where my max samples were only about 128 (for animation), but because the minimum samples were set to 0, Blender kept undersampling the darkest areas, leading to some really weird denoising. You can set your noise threshold actually fairly high if you up the minimum samples, because then even darkened areas still get enough samples that the denoiser doesn't have to guess too much on them, leading to a smoother color gradient, instead of weirdly mushy shadow interiors.
Through a lot of testing I found that noise threshold very often produces inferior results than just an equivalent render time without it - and that's especially true in scenes that are inherently very noisy across the entire image (like volumetric heavy scenes).
That being said noise threshold is incredibly useful if there are parts of the scene that require a lot of samples and other parts that require barely any at all (for example a sky, hdri or just transparency).
It's possible that this is because I never touched the minimum sample value and perhaps Blender is setting it too high as you said in the video - knowing that information now I will be trying to set the minimum samples manually to something very low like 32 and see if that helps.
Still it's weird when I have a scene with a noise threshold set to some value and then the renderer proceeds to leave some areas pretty noisy and then spends more time on areas that are already clean. It feels like it's working pretty much backwards, but again, that might just be me.
It probably has to do with how Blender detects noise in the image. I remember seeing a paper on adaptive sampling and from what I gathered, it uses a special pass for detecting noise rather than analyzing the final combined render. This pass doesn't take into account variables in your scene that increase noise like small light sources or subsurface scattering. I can tell for a fact that SSS noise isn't being taken into account because it's consistently always noisier than almost everything else.
I don't recommend using denoise for animations, especially when using volumes in scenes as clouds or mist.
So does this mean it's now better to render each frame as a whole frame. Because back in the day you were supposed to break it up into smaller chnunks. But it seems like adaptive sampling would be less effective with that technique
I am a little less confused now. Thank you
But one problem thou, i personally usue sample limit at just 128, and if it gets blurry due to denoise i tend to just render the scene at a higher resolution like 4k or even 8k, it tends to clean just as good as havingn the base 4069 base sample count that blender uses but is way faster
Ooh thatâs a really good idea đź
04:30
I could be wrong but I think if the min_samples is set to zero this is how it is setting the min_samples:
const int min_samples = (int)ceilf(16.0f / powf(adaptive_sampling.threshold, 0.3f));
adaptive_sampling.min_samples = max(4, min_samples);
so unless you set the threshold to zero the min sample number shouldn't change based on the max sample count.
but then the adaptive_sampling itself is doing some quirky bitwise operations (that I don't want to think about now đ). at first glance I assume that probably changes the stopping conditions so that it will process more samples which from the user's pov feels like a higher min_samples.
-idk xD
I need to Try and Error until i see what works best for a particular project
thank u boss
Showing denoise difference with CZcams's quality is not a great idea, mate.
me watching this in a high end laptop
(I dont need it)
3:20 time is the same
I rarely use Cycles as I've noticed quite a few creators, keep recommending cycles, like Pixxo 3D, his videos are great, his graphics card is always compatible and he always when making textures goes with the highest possible textures he can do, with the latest versions of Blender , just Cycles doesn't work to well with me.
Ryan King is a good one, as he will use both Eevee and Cycles, and shows different ways to get textures to work in a similar way to Cycles, and ways to make sure your computer doesn't struggle with rendering.
Thatâs not a Boeing 747! đĄ
I really wanted to subscribe, the editing and content is really good but this god awful ai voice is unbearable.