David,
Just go to download and demo files tab on Blender.org towards the bottom on the page in the Cycles section you will find all the files I used for the previous benchmarks. See link Demo Files. The BWM scene is the bwm27_gpu.blend. I change the settings for Adaptive Sampling(AS) and output size increase from 50% to 100%. All the other settings where left alone.
Did anyone try rendering animation with AS+denoiser? Is it consistent enough, meaning no extra flickering?
Ok. Yes, I know where to get the BMW scene. I was specific about the āadvancedā settings for the Adaptive Sampling. 
Serge,
Currently the denoiser donāt take into account previous animation frames to avoid some of the smoothing artifacts in animation. It is possible depending on the scene to get okay results with a high samples and using low AS thresholds.
This feature call āTemporal Coherenceā is coming to the Intel denoiser. See link Animation Denoising. I am sure Nvidia Optix denoiser is working on a similar solution.
Thanks for cool inputs, it is indeed great to have an adaptive sampling in Cycles. Finally a good old communist way to help to equalā¦ noise
in some of my interior tests that had difficult lighting (small windows + portals, only lightsource being a HDRI) using scrambled distance + dithered sobol was almost twice as fast as using AS. So is that because Adaptive sampling could not help because the scene needed samplesā¦everywhere?
AS is most effective in scenes where certain spots are otherwise significantly easier to sample (and therefore converge much faster). It is also very effective in scenes that otherwise do not need a huge amount of samples, but you have highlights (such as in glass) that require a ton of samples to define.
If you have a scene that is almost all indirect lighting from tiny light sources, then it would be fairly uniform as far as sampling difficulty goes and limit the impact AS provides. If the noise is from light sources, then light trees, once that gets into Cycles, should help with that.
Hi all
Sorry i come with dumb questions about denoising 
The 1st one is: is the optix denoiser ( as i saw it comes from nvidia ) efficient on blender cycles texture baking ?
The 2nd is where and how to get this denoiser and set it up for beeing used in blender cycles ?
The 3rd question: what do you call AS in all your posts ?
Blender denoiser ( in scene tab ) has no option matching thisā¦
thanks a lot !
regards andā¦
happy blending !
anyone know about noise reduction and texture baking please ?
Well, you can use the denoiser node in the compositor to denoise any image input you want. I tested it on a noisy AO texture from an asset and it seems to work fine, just plug the texture into the image socket and leave the normal and albedo input clear. Then go to the image editor and save the denoised texture.
This is just a simple test, I donāt know how well it works on other cases.
denoiser nodeā¦ 
DENOISER NODE ??? where the hell did you find this !!! 
I still got so much to learn on blenderā¦
@AmyLily <=== becoming my semi-god 8))
Thanks a lot for the clue !!! I definetely have to dig in this !
happy blending :))
Go into the Compositor.
You must activate āDenoising Dataā in the Layer Properties.
Compositor Final
In some cases Denoising Normal and Albedo donāt quite work so well. So I still have active the regular normal pass and the diffuse color, combined with object/material ID and cryptomatte passes.
Binke,
I have done testing interior scenes with e-cycles pattern/scrambled distance compared with adaptive sampling (AS) they can both improve the rendering time, but work completely different. Scrambled distance (SD) at 1.0 gives no performance improvement compare to the daily build and a setting of SD of 0.001 gives maximum performance with biasing and terrible GI artifacts.
In testing a scene start with 1000 samples with adaptive sampling setting of 0.01 and then use an scramble distance (SD) of 0.1. Compare the times and quality after the denoiser. If close I usually prefer using adaptive sampling because is not adding biasing and GI artifacts to the scene. For final high quality still I use AS setting of 0.001. This is subjective and depends in the type of scene and preferences.
In the future I would hope that Cycles could add new algorithms that improve Cycles performance with out increasing GI biasing and artifacts. Like LuxCoreRender has done by adding Env. Visibility Map, Direct Light Sampling Cache and other improvements to their render engine.
Hmm, do you have any rendered examples of artifacts you saw with scrambling distance vs AS?
When i use adaptive sampling in masters vs scrambling distance + dithered sobol in BoneStudios custom build, I cant say i see any big noticable quality differences.
doesnt e-cycles introduce some extra biasing over just scrambling distance as well ? You should try a build with just scrambling distance (and dithered sobol) and compare. Try bonestudios build on graphicall!
Keep in mind that the combination of AS+SD is not offically supported. Adaptive sampling is designed to work with the PMJ sampler, SD is designed to work with the Sobol sampler. Using both together means using a suboptimal sampler for one of the two algorithms, which may in cases result in suboptimal results.
Iād also like to point out, that scrambled distance does not introduce bias, where adaptive sampling does introduce bias. Correlation artifacts are not a proof of a biased algorithm, nor does the presence of noise mean an algorithm is unbiased.
Stefan thanks for clarifying the correlation artifacts and bias.
Binke, I build my on Blender with scrambling distance and donāt use dithered sobol it has issues and its main benefit is only in low sample renders. I compare result with the Bone and e-cycles and its exactly the same. Of course scrambling distance produces noticeable correction artifacts if scrambling distance value is low enough. This artifacts donāt denoise well.
Using SD of 0.001 on the BMW scene took 45sec. with denoiser. Notice specular highlight splotches on the floor caused by scrambling distance.
Using SD of 0.01 on Corona test interior with 2000 samples from @rawalanche took 1:21sec. with denoiser. Notice the GI splotches in the ceiling, floor and walls.
Compare this Corona test interior using AS of 0.01 and 2000 samples took 25sec. with denoiser.
The scrambling distance artifacts become much less noticeable with higher values, but you loose performance improvements. In a scene like the BMW I can still notice those specular correction artifacts at higher values.
The question I was trying to answer when to use adaptive sampling instead of scrambling distance/e-cycles pattern. In a scene like BMW using AS with denoiser it will be least 4x faster with the same perceive image quality. In a scene like Corona test interior their is less of an advantage for using AS. If using scramble distance should be careful of those correction artifacts.
I see, yes i rarely use scrambling distance below 0.4, so i have never encountered any noticeable artifacts in my renders, but yes if you go too low it will happen apparently.
But also - are you sure those artifacts are not just from the denoiser? I think to really compare, you need to look at it without the denoiser.
And yes, the benefits of AS differ greatly depending on the scene.
This is amazing! Any crashed course with this? 
I really love it!
is this visual representation also possible with cycles?
