What’s your go-to-way for Noise Removal?
Is it Nyx or something else outside of TVAI?
DVD sources: Proteus 4 manual ‘Fix compression’. Usually 20 to 45 does the job.
Blu-ray sources: Nyx 3 relative to auto ‘Reduce noise’ usually set to -20 or so. All other sliders must be at -100 and 0 for Anti-alias/deblur. I’ll move it up into the positive numbers if the noise is very heavy.
I have tried QTGMC in the past with its EZDenoise option. That works pretty good when combined with Proteus 3, but Proteus 4 keeps more details with the noisy input.
Depends on the type of noise. Noise from analog digitization has very different characteristics to CCD digital noise, monte carlo noise, Gaussian noise, “blender” digital noise, low-light capture noise, analog dot-crawl, rainbows, luma vs chroma noise etc. There is not one “fix all” automagic noise filter.
The industry does a poor job of providing tools for objective noise measurement and visualizing and distinguishing the different types of noise. In order to reduce noise you have to be able to distinguish and measure different types of noise and apply an appropriate filter. Most tools out there rely on subjective “feel”, often less-is-more and an editor must be careful to avoid the placebo effect while applying non-granular sliders in a darkened room - a trait that we are all guilty of. Noise Reduction is fundamentally destructive and should be avoided unless targeted to specific, quantifiable artifacts.
For temporal noise, FFmpeg’s hqdn3d filter is really good. For spatial noise, nlmeans is much better (although currently limited to 8-bit). Best practice is usually temporal NR, then spatial NR.
As in go-to? Neat Video is the yardstick commercial product for noise reduction, giving independent control over temporal vs spatial, with presets, previews and pixel-peep. Neat Video’s sweetspot is per-scene noise reduction during edit based on a 20-frame sample, rather than as a blanket-setting-in-post or per-file. The Neat Video website & UI looks like cancerous junkware, but it is used throughout Hollywood. Neat Video relies on an NLE-plugin architecture, which is a burden. I don’t want to install Resolve just to run Neat Video. If Topaz and Neat Video were to merge or Neat to provide a plugin for TVAI, life would be better.
I’ve also heard good things about Cinnafilm/Pixelstrings but have not used their pro-tools in anger.
Most NR filters tend to operate in the YUV domain. Most AI upscalers tend to operate in the RGB domain. It is usually worth applying your NR early in the chain, prior to YUV>RGB conversion and especially prior to any upscale.
As for TVAI models or tools, they don’t distinguish between the different types of noise. I tend to lean on FFmpeg or Neat Video prior to import into TVAI (FFV1 FTW). Sure, the TVAI filters are ok, but they don’t offer objective measurement, differentiation of temporal vs spatial, nor visualization of luma vs chroma noise. In the future, this may become moot with diffusion-based synthesis, but for the meantime TVAI doesn’t provide enough granular control or visualization to compete with Neat Video or a series of hyper-targeted FFmpeg filters.
Enjoy the rabbit-hole of NR.
I think noise reduktion is important topic but the problem will get smaller as we already can see if you use heavy reconstruction models like Starlight that swirl and recompose pixels new, a lot of “noise” is gone, gets “content integrated”.
