Denoise first, then rescale. But are you sure?

[rgr]

The rule is to denoise first and then rescale.
But the matter is not so clear-cut, I leave it to your judgment.
On the left, upscale from 576p to 1080p, then Neat Video.
On the right, first Neat, then upscaling to 1080p.

https://imgsli.com/MjY2MTY5

(In both cases, first there was:
QTGMC(preset="slower",InputType=1,TR2=3,sharpness= 0)
temporaldegrain2(grainLevel=1,postFFT=3,postSigma= 0) )

[lollo]

QTGMC denoises

TemporalDegrain2 is a denoiser

so there is a strong denoise prior to NeatVideo+rescale and rescale+NeatVideo, and no conclusions are possible on this use case.

Maybe you can compare using just an intraframe deinterlacer (nnedi) or just a lossless bob operation before NeatVideo+rescale and rescale+NeatVideo, checking the kept lossless frames.

[rgr]

Yes, but the source was in poor condition. QTGMC stabilized them well, and TD2 worked well as a pre-filter.

(The first one was converttoyuv444.convertbits(16) to reduce filtering side effects.)

[DTL2023]

Rescale may be upscale and downscale.

When you run mvtools with pel > 1 it also performs upscale to 2x (pel=2) and 4x (pel=4) to make processing at sub-(full)sample level. And mvtools is the main engine for QTGMC (and part of temporaldegrain script).

If you denoise before downscale - the motion search engine may use more real data to search (if it is really present in the source). If you try to MotionEstimate over very soft source like 35mm fim scan to FHD or even UHD - you need to typically use larger blocksize (like 16x16 or even 32x32) to grab more real details in the block of fixed sample-sized (so mapping of real film surface area to block remains the same).

If you denoise after downscale - you can use lower block size like 8x8 at SD and it gets partial denoise at the downscaling process.

At denoise before upscale - it may be less critical. Before upscale pel may be set to larger value like 2 or 4 and if denoise after upscale - pel may be set to lower values like 1.

"On the left, upscale from 576p to 1080p, then Neat Video.
On the right, first Neat, then upscaling to 1080p.
"

With highly non-linear processing the order of operations starts to make difference. So it looks like some non-linear sharpener create more peaking at OSD in one of the 2 ways. NeatVideo damage more fine details - so if you upscale first the any small amount of fine details in the VHS recording become much larger in size (relative to sample size) and NeatVideo less degrades. So it may depends on the denoiser engine used.