How optimised is my script for batch encoding?

[waxstone]

I'm about to batch filter many Japanese broadcast tapes, there will be slight tweaks with regards to TFF/BFF, and how much each edge will be trimmed.

I'm worried that I may be missing a step. How do my scripts look?

What I'm trying to achieve:
1. Frame rate doubled to 59.94fps for smooth playback (retaining both live action and animation clarity)
2. Slight noise reduction
3. Remove junk noise and empty space through cropping
4. Replace the cropped pixels with black pixels, retaining original dimensions of 720 x 480
5. Center the picture after cropping, for image symmetry
6. Encode at a faster speed (prefetch)
7. Encode at high-lossy with no visible difference to lossless
8. Encode at correct aspect ratio of 4:3

I'm also going to archive the raw captures as FFVI and FLAC. Does this seem sensible, and would it safe to convert these raw captures, and THEN encode, without loss of quality?

I'm also looking for recommendations on techniques for up-scaling correctly to 4K for YouTube, again without losing quality. Not sure how much of a difference it would make to upscale from raw source vs. encode.

My flitering script:
Import("TemporalDegrain-v2.6.6.avsi")

# Load source
video_audio = FFmpegSource2("1_New_Raw_Sample_2.avi", atrack=-1, fpsnum=30000, fpsden=1001).AssumeTFF()

# QTGMC at 59.94 fps
processed = QTGMC(video_audio, preset="Slower")

# Crop after deinterlacing
crop_left = 4
crop_top = 4
crop_right = 12
crop_bottom = 4
cropped = processed.crop(crop_left, crop_top, -crop_right, -crop_bottom)

# Denoise and sharpen in YV12 (this is to avoid unnecessary format switching)
denoised = cropped.TemporalDegrain2(degrainTR=1)
# Reconvert to YV12 due to TempDegrain output
sharpened = denoised.ConvertToYV12().LSFmod(defaults="slow")

# Restore borders
border_left = Round((crop_left + crop_right) / 2)
border_right = (crop_left + crop_right) - border_left
border_top = Round((crop_top + crop_bottom) / 2)
border_bottom = (crop_top + crop_bottom) - border_top
restored = sharpened.AddBorders(border_left, border_top, border_right, border_bottom)

# Output with prefetch
return restored.Prefetch(16)

Encoding command:
ffmpeg -i script.avs -vf "setsar=10/11" -aspect 4:3 -c:v libx264 -preset slow -crf 18 -pix_fmt yuv420p -profile:v high -level 4.1 -rc-lookahead 40 -refs 4 -c:a aac -b:a 192k -movflags +faststart 2_New_Encoded_Sample.mp4

See sample video: https://drive.google.com/drive/folders/1MWTfERgIlVtErIHZY22--jZ-dFPt62aW?usp=sharing

[VideoMilk78]

I recommend looking into TemporalDegrain2_Fast by Arx1meD: https://forum.doom9.org/showthread.php?p=1982594#post1982594

[waxstone]

I recommend looking into TemporalDegrain2_Fast by Arx1meD: https://forum.doom9.org/showthread.php?p=1982594#post1982594

I've just tried it out and wow, it's encoding at 73fps rather than 30. That's a huge leap!

The only thing is that now the scripting is quite different (Degrain2Fast doesnt understand degrainTR=1), so I've just had to go with the default filtering setting.

I can see that the filtering is too heavy on standard settings on TDegrain2_Fast, notice the rounding of edges on the Playstation logo: https://imgsli.com/Mzc3NzQ4

Any filter settings you would recommend?

[StainlessS]

The only thing is that now the scripting is quite different (Degrain2Fast doesnt understand degrainTR=1), so I've just had to go with the default filtering setting.

degrainTR=1, looks like "Temporal Radius", for TemporalDegrain2_fast, similar is RadT, also with default of 1.

function TemporalDegrain2_fast (clip input, int "Strength", int "Y", int "U", int "V", int "RadT", int "BlkSz", int "Olap", \
float "Sharp", bool "PostDeHalo", float "PostMix")
{
Str = Default(Strength, 3) # Noise/grain suppression strength.
# Strength for depth noise (< 2), low (3 ... 5), medium (6 ... 9), high (10 ... 14), veryhigh (> 15)
Y = Default(Y, 3) # Luma plane to process. Value: 2 - copy from input, 3 - process
U = Default(U, Y) # Chroma plane to process. Value: 2 - copy from input, 3 - process
V = Default(V, U) # Chroma plane to process. Value: 2 - copy from input, 3 - process
RadT = Default(RadT, 1) # Temporal Radius of frame analysis. Value 1 or 2
BlkSz = Default(BlkSz, 16) # Block size for motion analysis. Bigger BlkSz quicker. Recommended values: 8, 16, 32
OLap = Default(OLap, BlkSz/2) # The value of overlapping blocks on each other
Sharp = Default(Sharp, 0.4) # Sharpening strength. Range: 0 ... 1
DeHalo = Default(PostDeHalo, false) # Remove halo after sharpening. Value: true or false
PostMix = Default(PostMix, 0) # How much noise/grain will be returned. Range: 0 ... 1


So maybe something like

# ...
STRENGTH = 2 # Default is 3, reduce a bit
RADT=1 # Default is 1 anyway, cannot reduce, is minimum radius.
POSTMIX=0.0 # Default 0.0, How much noise/grain will be returned. Range: 0 ... 1
# extra stuff
SHARP=0.4 # Default 0.4, maybe reduce to 0.3 if too sharp
POSTDEHALO=false # Default false, make true if dehalo-ing.
# ...
denoised = cropped.TemporalDegrain2_fast (Strength=STRENGTH,RadT=RADT,Sharp=SHARP,PostDehalo=POSTDEHALO,PostMix=POSTMIX)
# ...

[waxstone]

degrainTR=1, looks like "Temporal Radius", for TemporalDegrain2_fast, similar is RadT, also with default of 1.

So maybe something like

# ...
STRENGTH = 2 # Default is 3, reduce a bit
RADT=1 # Default is 1 anyway, cannot reduce, is minimum radius.
POSTMIX=0.0 # Default 0.0, How much noise/grain will be returned. Range: 0 ... 1
# extra stuff
SHARP=0.4 # Default 0.4, maybe reduce to 0.3 if too sharp
POSTDEHALO=false # Default false, make true if dehalo-ing.
# ...
denoised = cropped.TemporalDegrain2_fast (Strength=STRENGTH,RadT=RADT,Sharp=SHARP,PostDehalo=POSTDEHALO,PostMix=POSTMIX)
# ...


I've updated my script and the end result is looking great, plus it's encoding much faster. Thank you so much!

Would you have any other recommendations outside of noise removal?

I'm thinking about upscaling to an HD/4K res as an encode for YouTube viewing too. Is there any benefit to upscaling from the RAW rather, or would there be no difference in quality if upscaled from the SD encode correctly?

[VideoMilk78]

I'm thinking about upscaling to an HD/4K res as an encode for YouTube viewing too.

A sharp SD image with minimal compression can look better than a highly compressed 1080p video. I see no value in it although maybe StainlessS might have a different opinion?

[waxstone]

A sharp SD image with minimal compression can look better than a highly compressed 1080p video. I see no value in it although maybe StainlessS might have a different opinion?

I simply ask because YouTube annoyingly adds more compression to SD uploads than HD/4K uploads, meaning that the best option to preserve detail for viewers is to provide higher res video to be uploaded.

I'm aware that the actual encode wouldn't be more enhanced in HD/4K than SD, and would likely look identical.

[StainlessS]

I've never uploaded anything to YT, hence, I am staying out of this :)

[Arx1meD]

waxstone, if you have a good video card, try the next settings for Temporaldegran2:
TemporalDegrain2(grainLevel=1, limitFFT=2, postFFT=4, degrainPlane=4, postDither=0) # for GPU

Or try my TemporalDegrain2_fast with changes:
function TemporalDegrain2_fast (clip input, int "Strength", int "Y", int "U", int "V", int "RadT", int "BlkSz", int "Olap", \
float "Sharp", bool "PostDeHalo", float "PostMix", bool "GPU")
{
Str = Default(Strength, 3) # Noise/grain suppression strength.
# Strength for depth noise (< 2), low (3 ... 5), medium (6 ... 9), high (10 ... 14), veryhigh (> 15)
Y = Default(Y, 3) # Luma plane to process. Value: 2 - copy from input, 3 - process
U = Default(U, Y) # Chroma plane to process. Value: 2 - copy from input, 3 - process
V = Default(V, U) # Chroma plane to process. Value: 2 - copy from input, 3 - process
RadT = Default(RadT, 1) # Temporal Radius of frame analysis. Value 1 or 2
BlkSz = Default(BlkSz, 16) # Block size for motion analysis. Bigger BlkSz quicker. Recommended values: 8, 16, 32
OLap = Default(OLap, BlkSz/2) # The value of overlapping blocks on each other
Sharp = Default(Sharp, 0.4) # Sharpening strength. Range: 0 ... 1
DeHalo = Default(PostDeHalo, false) # Remove halo after sharpening. Value: true or false
PostMix = Default(PostMix, 0) # How much noise/grain will be returned. Range: 0 ... 1
GPU = Default(GPU, false)

func_name = "TemporalDegrain2_fast: "
Assert(Y == 2 || Y == 3, func_name+"Luma Y plane must be 2 or 3")
Assert(U == 2 || U == 3, func_name+"Chroma U plane must be 2 or 3")
Assert(V == 2 || V == 3, func_name+"Chroma V plane must be 2 or 3")
Assert(Y == 3 || U == 3 || V == 3, func_name+"One of the planes Y, U, V must be 3")
Assert(RadT == 1 || RadT == 2, func_name+"Temporal Radius of frame analysis must be 1 or 2")
Assert(Sharp >= 0 && Sharp <= 1, func_name+"Sharpening strength must be between 0 and 1.0")
Assert(PostMix >= 0 && PostMix <= 1, func_name+"The noise return value must be between 0 and 1.0")

dPlane = Y==3 && (U==3 || V==3) ? 4
\ : Y==3 && U==2 && V==2 ? 0
\ : Y==2 && U==3 && V==2 ? 1
\ : Y==2 && U==2 && V==3 ? 2
\ : Y==2 && U==3 && V==3 ? 3 : 4
pad = Max(Blksz, 8)
pel = 1
chr = U==3 || V==3 ? true : false

# denoising 1st way
dgLimit = GPU ? input.FFT3DGPU(sigma=Str, sigma2=Str*0.625, sigma3=Str*0.375, sigma4=Str*0.250, \
bt=RadT==1?3:5, plane=0, bw=BlkSz*2, bh=BlkSz*2, ow=OLap, oh=OLap)
\ : input.neo_fft3d(sigma=Str, sigma2=Str*0.625, sigma3=Str*0.375, sigma4=Str*0.250, \
bt=RadT==1?3:5, Y=Y, U=U, V=V, bw=BlkSz*2, bh=BlkSz*2, ow=OLap, oh=OLap)
dgSpatD = mt_makediff(input, dgLimit, Y=Y, U=U, V=V)

# denoising 2nd way
dgNR1 = GPU ? dgLimit.KNLMeansCL(h=0.1+Str/5.0, d=RadT, a=2, s=4, channels="Y", device_type="GPU")
\ : dgLimit.vsDeGrainMedian(modeY=0, limitY=Str, limitU=U==3?Str+2:0, limitV=V==3?Str+2:0)
dgNR1D = mt_makediff(input, dgNR1, Y=Y, U=U, V=V)

# combine 1st and 2nd ways
dgDD = mt_lutxy(dgSpatD, dgNR1D, "x range_half - abs y range_half - abs < x y ?", Y=Y, U=U, V=V, use_expr=2)
dgNR1x = mt_makediff(input, dgDD, Y=Y, U=U, V=V)

# sharpen the edges only
dgNR1x = Sharp > 0 ? mt_merge(dgNR1x, \
dgNR1x.RemoveGrain(17).Sharpen(Sharp), \
dgNR1x.RemoveGrain(12).mt_edge("prewitt", Y=3, U=2, V=2).mt_inpand(chroma="-128").Blur(1.58), \
Y=3, U=2, V=2)
\ : dgNR1x

# denoising 3rd way
sup = dgNR1x.Blur(1.58).Blur(1.58).MSuper(hpad=pad, vpad=pad, pel=pel, chroma=chr)
MultiVec = MAnalyse(sup, multi=true, delta=RadT, blksize=BlkSz, overlap=OLap, search=5, dct=7, chroma=chr, truemotion=false, global=true)
dgNR1xS = MSuper(dgNR1x, hpad=pad, vpad=pad, pel=pel, levels=1, chroma=chr)
dgNR2 = MDegrainN(dgNR1x, dgNR1xS, MultiVec, RadT, plane=dPlane)

# combine 1st, 2nd and 3rd ways
dgDD2 = mt_lutxy(dgNR1x, dgNR2, "x range_half - abs y range_half - abs < x y ?", Y=Y, U=U, V=V, use_expr=2)

# sharpening
allD = Sharp > 0 ? mt_makediff(input.Sharpen(1), dgDD2.Blur(1.58).Blur(1.58)) : NOP()
ssD = Sharp > 0 ? mt_makediff(dgDD2, dgDD2.RemoveGrain(20)) : NOP()
ssDD = Sharp > 0 ? mt_lutxy(ssD.Repair(allD, 12), ssD, "x range_half - abs y range_half - abs < x y ?", Y=Y, U=U, V=V, scale_inputs="allf") : NOP()
out = Sharp > 0 ? mt_lutxy(dgDD2, ssDD, "x range_half y - "+String(Sharp)+" * -", Y=3, U=2, V=2, scale_inputs="allf") : dgDD2

# dehaloing
m0 = DeHalo ? dgNR1x.mt_edge("prewitt", Y=3, U=2, V=2).mt_inpand(chroma="-128") : NOP()
m1 = DeHalo ? mt_lutxy(out, m0, "y range_half > x y ?", Y=3, chroma="-128", scale_inputs="allf").mt_binarize(threshold=128).Blur(0.5) : NOP()
out = DeHalo ? mt_merge(out, dgNR1x.Blur(0.1), m1, Y=3, U=2, V=2) : out

PostMix > 0 ? mt_lutxy(out, input, "x x y - "+String(PostMix)+" * -", Y=Y, U=U, V=V, use_expr=2) : out
}

TemporalDegrain2_fast(Strength=3, Y=3, U=3, V=3, RadT=1, BlkSz=16, Sharp=0.4, PostDeHalo=false, PostMix=0, GPU=true)

[StainlessS]

Arx1meD, Your GPU arg is new compared to the script we were looking at.
Does your script have its own thread, if so, then where ? (It really should have its own thread)
I did a D9 search on "TemporalDegrain2_fast", both as thread title only, and also anywhere in any post, and found NOTHING!!!
(not even in this thread, whats wrong with D9 search)

EDIT:
Although we also did Google search on "TemporalDegrain2_fast" site:forum.doom9.org
with results,
https://www.google.co.uk/search?q=%22TemporalDegrain2_fast%22%20site%3Aforum.doom9.org&sclient=gws-wiz


.

[Arx1meD]

StainlessS, TemporalDegrain2_fast does not have its own thread. My script cleans the video from noise worse than TemporalDegrain2. I do not think that he needs his own thread.

[DTL]

After FFT3D and KNLmeansCL executed on GPU - the part of mvtools can also be tested for ME at GPU using that version of mvtools2 - https://forum.doom9.org/showthread.php?t=183517

[tormento]

If you have a nVidia card, unless some personal requirement, I strongly suggest you to use modern filters such as BM3D_CUDA.

If you carefully tune the parameters, it can outperform both in speed and quality MVTools and have less artifacts too.

[waxstone]

If you have a nVidia card, unless some personal requirement, I strongly suggest you to use modern filters such as BM3D_CUDA.

If you carefully tune the parameters, it can outperform both in speed and quality MVTools and have less artifacts too.

I've got a NVIDIA GeForce RTX 4070 Laptop GPU. Reading up on BM3DCUDA (https://forum.doom9.org/showthread.php?t=183066), it could benefit my workflow. Would this be a smooth transition and be implemented into my .avs scripts with easy?

StainlessS, TemporalDegrain2_fast does not have its own thread. My script cleans the video from noise worse than TemporalDegrain2. I do not think that he needs his own thread.

I was also confused as to why this doesn't have its own thread. Although I'm happy with the increased speed of my encodes, when you say it cleans "worse" than the original it doesn't really fill me with confidence. Why this over the above mentioned BM3CUDA method?

[tormento]

Would this be a smooth transition and be implemented into my .avs scripts with easy?
A typical script could be:

[Your source]
z_ConvertFormat(resample_filter="Spline64", pixel_type="yuv444ps")
BM3D_CUDA(sigma=6, radius=3, chroma=true, block_step=6, bm_range=12, ps_range=6)
BM3D_VAggregate(radius=3)
z_ConvertFormat(resample_filter="spline64",dither_type="error_diffusion",pixel_type="YUV420P16")
fmtc_bitdepth (bits=10,dmode=8)
Prefetch(2,6)

But you have to manually tune the parameters according to your necessities and GPU memory. Read the avisynth+ wiki about BM3D_CUDA about it.

[VideoMilk78]

I simply ask because YouTube annoyingly adds more compression to SD uploads than HD/4K uploads, meaning that the best option to preserve detail for viewers is to provide higher res video to be uploaded.

I'm aware that the actual encode wouldn't be more enhanced in HD/4K than SD, and would likely look identical.

If you do that I recommend just resizing...
Another solution would be to upload the sd clip and link a download to the high quality version