Denoising and upscaling animation renders from Blender 3.4 - Neat Video, Topaz Video AI
Vložit
- čas přidán 11. 12. 2022
- The demo scene rendered in this test animation is #3 from this scene pack by Evermotion: blendermarket.com/products/ar... - I created the camera animation but everything else was created by Evermotion.
Rendered on ASUS Strix RTX 3090 GPU at 65% max power level in MSI Afterburner for better power efficiency.
Watch this tutorial to see how to work with LUTs to transform linear EXR data to filmic transform from Blender into Resolve:
• Tutorial: Real-time EX...
Watch the tutorials on NeatClips channel to better understand how to use Neat Video. It will take less than 2 hours of your time but is worth it. Here is one: • Advanced Workflow Tuto...
Download the LUTs here: github.com/sobotka/filmic-res...
I did not go into Neat Video settings in depth nor did I explain the ideas behind a noise profile and how a bad noise profile could skew the effectiveness of the local flicker slider. I also didn't demonstrate the play button and other features in Neat Video. Please watch the tutorials linked below for a better understanding of Neat Video as my video here was more of a general overview of what I do to denoise animation renders: • Tutorials
#B3D - Krátké a kreslené filmy
That was an excellent tutotiral ... Everything was clear and well explained. Just what I was looking for.. Well done.
This is not an in depth video on the usage of Neat Video, I have provided a link in the description showing Neat Video tutorials. This video is an overview tutorial of the general process I use to denoise and upscale my animation renders! Also be aware that in Blender 3.5, the light tree feature means far less samples are required to give acceptable results, reducing render time by almost half for similar results as this. It works on both CPU and GPU rendering :)
Cool....looks FABULOUS 👌 😊👍
Thank you, this is very helpful and valuable information!
Just as a reminder, Blender 3.5 and later has "Light Tree" feature, so you can achieve similar results with about half the render time per frame as what I did here! Could probably be fine using the same techniques, 1024 samples at 1920x1080 :)
Really interesting video with excellent advice, well explained...cheers :)
Thank you! Keep in mind that with Blender 3.5, the "light tree" feature in Cycles will eliminate the need for a lower noise threshold and high max samples count as you saw here. I could use say 512 max samples and slightly lower noise threshold or same and achieve slightly better results with a little over half the render time. These frames averaged 5 minutes on an RTX 3090. I did a test with Blender 3.5 alpha using "light tree" and I got 2m45s renders that looked better. Just so you know. However, most of what I said still applies in general, it's just Blender 3.5 will be even faster at rendering less noisy images :)
@@BackusCreativeImaging I have not played with Blender 3.5 yet and did not know about this new feature...it sounds awesome and I am currently working on a pretty heavy scene so this should help massively...you have a great teaching style and I learned so much just in that short tutorial...thanks
@@nadstunes77 Yep, Blender 3.5 will give you half the render time on a light heavy scene like this. Blender 3.5 should be released by April 2023. Thanks a lot, I'm glad this video helped you :)
Awesome video! I'm using a similar worklow, but sometimes the Resolve Denoiser works better for me, I've tried the Flicker Reduction and it also tends to work for me sometimes. I guess with archviz you don't run into such problems, but I do tend to get motion aritfacts with both Neat and Davinci Denoisers. I often render fast moving objects and they get very smeary or choppy because there is just too much inter-frame motion. I've even tried Optix temporal denoising - it doesn't even seem to work at all. So untile OIDN comes with a temporal solution, I can't seem to find a way to work around this issue.
Also keep in mind that Blender 3.5 will have the "light tree" feature in Cycles, reducing render time by 1/2 roughly for same quality in scenes with complex lighting. I tried it on this scene and I got 2m30s renders with less noise and artifacts. Denoising (whether spatial or temporal) will never be perfect but for Archviz this is a great method. For characters and other types of animation with fast moving objects, I do think the temporal flickering will be more noticeable; I'd say turn down the radius to 1 (instead of default 2 in Neat Video), keeping more noise for less smearing and such. I'm guessing you've already tried things like that. I will be using this method and tweaking it as necessary for my future creature / character animations. It will never be perfect but it's way better than waiting an hour per frame or more in my opinion, when there isn't time for it nor a desire to rent time on a render farm :)
@@BackusCreativeImaging Yeah, good points, I'm also looking forward to OpenPGL on the GPU! I've tried most of that... I'm still always 100% for elaborate denoising, even if it does get smeary a bit sometimes:) There is one nice thing I found, but may require added rendertime - upping the render resolution to UHD and dropping sampes to 50% (quater is too low, maybe 1/3 would do). Then the denoise passes are very sharp and it preserves a lot more texture - it also works nice when downscaling to HD. And there's a second nice one but doesn't work for me at all - temporal reprojection with vector passes... again, very nice but gives me artifacts for fast moving particles - this should be avoidable with a properly working AI temporal denoiser... I hoped for Intel but they keep us waiting.
@@maikeru2137 So I did another test render using Blender 3.5 alpha and the light tree feature. I'm really liking the results, at least for Archviz or scenes where objects don't move. Anyway, I rendered in UHD and then did some denoising. Downscaling to FHD wouldn't be a bad idea from there for a super sampled render for character animations. Blender's spatial denoiser is very good, would definitely like to see a temporal denoiser but I'm pretty satisfied with the light tree feature in Blender 3.5. I used to do CG back in 2010 using Maya and Mental Ray renderer. There was nothing like this. 2m30s per frame in 1080p with path tracing was a dream back then. I had to render a "final gather" pass for indirect illumination, it was not fun when objects moved in the animation. Check out my "Personal Projects" playlist and check out the Havenhurst Animation video. Everything was so awkward with rendering back then, that 35 second animation took 35 days to render, 1 hour per frame. Things have come a long way since 2010. I neglected my 3D hobby since 2013 but in 2020 got back into it with Blender 2.8, completely switched to Blender and I love it. I'll be doing some creature animations so will definitely try to master getting renders with minimal noise. I accept some temporal denoising artifacts, I embrace a little imperfection, it takes the CG asthetic off. If you toss some light film grain on top it kind of evens everything out :)
the Temporal NR works really great in Resolve Studio !
Does over exposing the footage like you would with s-log or something not work for the shadow detail noise? This is such a cool concept to me haha
Keep in mind that exposing to the right simply gains up the noise already there. I understand it's a popular technique with log footage but it's not the best if you ask me, unless maybe in extreme examples like astrophotography. Exposing to the right of middle grey begins to tone map colors and reduce their natural appearance. Exposing (mostly) correctly or at most 1/3rd stop over middle grey and creating a proper noise profile with your footage in Neat Video will give better results than just exposing to the right excessively, that's what I think. With all that said, now speaking about 3D animation software like Blender, the same principles apply. However, when rendering out EXR image sequences in 16 bit float or 32 bit float, you literally can't under or over expose your image. I prefer EXRs. I treat the noise like I would with video footage :)
@@BackusCreativeImaging good to know about the EXRs; one more thing on the stack of things to learn haha. just now diving into 3D (blender) from video. learned a ton, ty.
@@jvsonyt I used to use paid software like Maya, Mari, Nuke, Zbrush etc. All great software but expensive. Blender isn't better than that combo of specialist software but it can do pretty much everything to a good degree and is free. I bought courses from CG Boost website to help me transition to Blender. Check out their free training course too. Also, I actually use Blender to edit 360 degree footage shot on GoPro MAX. It's the best way to do it, total creative freedom: czcams.com/video/Q7W2TSwAYBA/video.html
You'd have to get comfortable with Blender first though, but I assure you that it's useful to know 3D alongside video. I was present as DOP on a 4 day short film shoot. The film is "Treasure", should be out this year. I'll make a video when I finish it (working on audio editing in Resolve now). I did all on set shooting (including pulling focus manually on my GH5M2 camera) and all post work except music creation. I also did the VFX, camera tracking and animating realistic fairies, all in Blender. It's extremely rare that someone knows cameras and 3D. Expand your knowledge and have fun :)
@@jvsonyt I also bought this course: czcams.com/video/K6lYYgLKH04/video.html along with several others. Haven't gone through that one yet but it's excellent. Definitely upgrading my skills for my future work :)
@@BackusCreativeImaging haha im excited to hear that you even know what a GH5 is lol. I'm getting pretty comfortable with blender to the point where i can watch this video and not have to scratch my head to figure out what you're doing. you have shown me that there is a rabbit hole that i for sure want to fall down haha. I'll for sure keep an eye out for your movie. Good stuff
Why don't you just use topaz video enhancer on the raw noisy footage? Don't use the blender denoiser and neat video at all, just topaz and it will be the same end result, just way faster and easier workflow
I tried just using Topaz Video AI on the original noisy sequence (no denoising in Blender) exported from Resolve as a DNxHR 4:4:4 10 bit file to test this. The results were pretty good, flicker was gone but there were still problem areas of high speckled noise patterns in some areas. I tried 1080p (source scale) along with 2x upscale to 4K. Not bad at all, Topaz Video AI is very good. However, using Neat Video alongside Blender's own denoiser gives better results overall. Why? Blender's spatial denoiser takes into account the albedo and surface normals, which seems to preserve more details. Neat Video allows you to fine tune your noise profile (which I really didn't demonstrate in this video), and you can truly remove all the noise (even with Reduce Noise set to 95, I still saw some of what I call noise splotches in Topaz Video AI when using only it to denoise). However, with that said, using Topaz Video AI for "everything" turns out to not be bad at all ... it is actually really good, but the results were not exactly the same as with my method; results were slightly sharper since Topaz was doing everything, but there were more problem areas in the end like noise splotches and dancing details which weren't super obvious but they were there. I do think using Topaz only could be fine in many cases though. I will still use my method because the results are more refined, even though it takes longer. What you said is a valid method though, it does produce good results which are better than I expected, proving just how good Topaz Video AI is. I still prefer using each software for their strengths though, Blender for spatial denoising, Neat Video for temporal denoising and local flicker removal, and Topaz for upscaling (even though it does have exceptional denoising). I also re-rendered some frames in Blender 3.5 alpha using the light tree feature for much better results. Instead of 5 minutes per frame, one can now go to 3 minutes per frame for equal to slightly better results :)
But Topaz is MORE EXPENSIVE...
How long did each frame take to render?
Approximately 5 minutes per frame on RTX 3090 only, in Blender 3.4. I reduced render time to half (so 2m30s a frame) with far less samples for same (if not slightly better) results using Blender 3.5 alpha and the new "light tree" feature in Cycles, but I didn't mention that in this video as I was using Blender 3.4 here.
@@BackusCreativeImaging I use a 3090 at work and a 4k 100-300 samples render for a far lighter scene took about 2-3 min. How could be possible to you to pull out a 16k samples render in 5min? I'm sorry but I can't believe it.
@@gabrielgonach6330 It's 16K samples with 0.03 noise threshold. 16K samples does not mean all pixels get 16K samples, sampling is determined by the noise threshold. That's why the setting is called "Max Samples", it doesn't mean a pixel will get 16K samples. The low noise threshold is why the render was faster than you'd expect at 16K samples. Production preset is 0.01 noise threshold which means that 16K max samples on a scene like this even on a 3090 would likely take a few hours per frame. The 0.03 threshold that I used was reached quite fast for most of the scene (meaning most areas probably got 200-500 or so samples) yet for the lights which had the most variation in noise, they took longer because they actually did get the full 16K samples or very close to it. So yes, 16K max samples with an average frame time of 5 minutes per frame on a single 3090, in Full HD 1920x1080 then upscaled to 4K in Topaz Video AI as mentioned in the video. This was kind of like "path guiding", where I focused more time of the render in the worst areas within the 5 minutes of a frame render. It is not true path guiding of course. Blender 3.5 will have the "light tree" feature which produces even better results and means I would not need 16K samples to smooth out high noise variation in the lights as the light tree does it. I re-rendered some frames of this scene in Blender 3.5 alpha a week after making this tutorial with far less max samples (I think 1024 or something like that, using 0.02 noise threshold) and got around 2m30s for most frames and similar quality as here if not slightly better, so the light tree feature in Blender 3.5 is a big deal :)