SVPflow motion interpolation (Avisynth + Vapoursynth, Linux, Mac?)

[Tempter57]

Mounir

SVSuper( "{ gpu:0, pel:2, scale:{up:0} }")

[chainik_svp]

Tempter57
if it was so easy

error: i don't know what "clean1" means

[Tempter57]

Originally Posted by chainik_svp

Tempter57
if it was so easy

Well, preset DVD MDegrain2_SVP mask4 DLS

#RemoveGrainSSE2.dll
#RepairSSE2.dll
#mvtools2SVP.dll
#svpflow1.dll
#svpflow2.dll
#fft3dgpu.dll
#WarpSharp.dll
#mt_masktools-26.dll
#average.dll
#AddGrainC.dll
#GradFun2DB.dll
#VagueDenoiser.dll
#flash3kyuu_deband.dll
#Gradfun2dbmod.avs
#Ylevels.avs
#LimitedSharpenFaster Mod.avs


setmemorymax(768)

# ColorYUV(gain_y=0,cont_y=0,cont_u=0,cont_v=0,gain_v=0,gain_u=-0,off_y=0,off_u=-0,off_v=-0)

source = last
preNR = source.fft3dgpu(wintype=1,degrid=1,bw=16,bh=16,ow=8,oh=8,bt=4,sigma=1.0,sigma2=1.2,sigma3=2.0,sigma4=1.0,plane=4).GradFun2DB(1.01)

setmtmode(2)
sup1 = preNR.SVSuper( "{ gpu:0, pel:2, scale:{up:2, down:2} }")

search_params = "gpu:0, vectors:3, block:{w:16,overlap:3}" 
v21 = sup1.SVAnalyse("{"+search_params+"}")
v22 = sup1.SVAnalyse("{"+search_params+",special:{delta:2}}")

bv22 = v22.SVConvert(isb=true)
bv21 = v21.SVConvert(isb=true)
fv21 = v21.SVConvert(isb=false)
fv22 = v22.SVConvert(isb=false)

mask1 = source.mmask(bv22, ml=100, kind = 1, Ysc = 255, gamma=0.999)
mask2 = source.mmask(bv21, ml=100, kind = 1, Ysc = 255, gamma=0.999)
mask3 = source.mmask(fv21, ml=100, kind = 1, Ysc = 255, gamma=0.999)
mask4 = source.mmask(fv22, ml=100, kind = 1, Ysc = 255, gamma=0.999)
maskb = mt_lutxy(mask1,mask2,"x 255 < y 255 < x y + 2 / x ? y ?")
maskf = mt_lutxy(mask3,mask4,"x 255 < y 255 < x y + 2 / x ? y ?")
tmask = mt_lutxy(maskb,maskf,"x 255 < y 255 < x y + 2 / x ? y ? 32 - 255 * 223 /")# комбинированная маска оценки движения SAD

source2 = mt_merge(source,preNR,tmask,Y=3,U=3,V=3)
sup2 = source.msuper(pel=2, levels=1, sharp=2)
den = source2.MDegrain2(sup2,bv21,fv21,bv22,fv22,thSAD=300,thSCD1=400,thSCD2=102)
Diff1 = mt_makediff(source,den,U=3,V=3)

NR = den.VagueDenoiser(method=4, nsteps=8, wavelet=2, Wiener=true, auxclip=preNR, percent=95, chromaT=0.8, wratio=0.75, threshold=0.4)
Diff2 = mt_makediff(source,NR,U=3,V=3)

lutxy = mt_lutxy(Diff1,Diff2,"x 128 - abs y 128 - abs < x y ?")
smB = source.mt_makediff(lutxy,U=3,V=3)

# MASKING
mE = mt_edge(smB,"prewitt",0,255,0,0,V=1,U=1).mt_lut(expr="x 1.8 ^",U=1,V=1).removegrain(4,-1).mt_inflate(U=1,V=1) 
mL = mt_logic(tmask.invert(),mE,"min",U=1,V=1).removegrain(20,-1) 
mF = mt_logic(tmask,mE,"max",U=1,V=1).removegrain(20,-1)

# SHARPENING
b1c = source.MCompensate(sup2, bv21)
f1c = source.MCompensate(sup2, fv21)
Sclp = smB.LSFmod(defaults="slow", preblur="ON", strength=100)
# Sclp = ContraHD(smB,source,b1c,f1c,0)
Tmax = source.mt_logic(f1c,"max",U=1,V=1).mt_logic(b1c,"max",U=1,V=1)
Tmin = source.mt_logic(f1c,"min",U=1,V=1).mt_logic(b1c,"min",U=1,V=1)
shrp = Sclp.mt_clamp(Tmax, Tmin, 2, 2, U=1, V=1)
sL = mt_merge(smB,shrp,mL,U=2,V=2)

# DEBANDING
# GFc = sL.f3kdb(sample_mode=2,dynamic_grain=false,keep_tv_range=false,dither_algo=3,y=64,cb=64,cr=64,grainY=48,grainC=48)
  GFc = sL.GradFun2DBmod(thr=1.21,thrC=1.5,mode=2,str=0.6,strC=0.0,temp=50,adapt=64)
Frs = mt_merge(GFc,sL,mF,luma=true,U=3,V=3)
Frs#.mergechroma(den)
YlevelsS(0,1.0,255,0,255,false)
 
# -- visualisations --
# stackvertical(source,last)
# interleave(source,last)

if SVSuper( " gpu:0, pel:2, scale:{up:2, down:2 } " ) => error " SVSuper: invalid "params" syntax"

[bmb]

hey their, im currently using this script

super = MSuper(pel=2,chroma=false)
backward_vec1 = super.MAnalyse(isb = true, blksize=16, delta = 1, overlap=8,chroma=false,search=5,searchparam=4,temporal=true)
forward_vec1 = super.MAnalyse(isb = false, blksize=16, delta = 1, overlap=8,chroma=false,search=5,searchparam=4,temporal=true)
MDegrain1(super, backward_vec1,forward_vec1,thSAD=20,plane=0)

i use it to denoise and remove grain generated from sharpening, it works great (the best iv used so far) it keeps the detail of the picture sharp and still theirs no noise or blocks generated.. my sources are bluray content.. i sometimes get slight noise and blocks (depending on the compression) i was wondering if you could translate it or offer a script better with this SVP and il test it out and report back.. thanks ...

[Boulder]

One small question: is there any harm in using a super clip created by MSuper instead of SVSuper when the vectors are converted to MVTools2 format after SVAnalyse?

[chainik_svp]

Boulder
There's no difference.

bmb

super = MSuper(pel=2,hpad=0,vpad=0)
vec = super.SVAnalyse("{block:{overlap:3}}")
backward_vec1 = vec.SVConvert(isb=true)
forward_vec1 = vec.SVConvert(isb=false)
MDegrain1(super, backward_vec1,forward_vec1,thSAD=20,plane=0)

didn't test it

[Boulder]

I tried to test your modded Avisynth build, but avs2pipemod keeps on crashing during launch so I cannot pipe the output to 64-bit x264. Opening the script in VDub has no issues.

[chainik_svp]

Boulder
tried version from here, works for me (without x264): "avs2pipemod.exe -rawvideo test.avs > out"

[Boulder]

Hm, it's possible that I have an outdated piping client (although I have had no problems using Avisynth 2.5.8 or 2.6 earlier). I'll get the latest one and retry.

[Boulotaur2024]

Originally Posted by Mug Funky

just dropping by to say thanks to all involved in the SVP project.

my canon workflow owes a lot to it

what i'm doing:

- shoot at as high a framerate as i can manage (720p60), at as high a shutterspeed as i can get a good exposure on (i like to shoot wide open, so it's no problem to get 1/250 or so). also, high fps means less rolling-shutter effect than if i'd shot at my target frame rate. less jelly = better.

- depanstabilize. goes without saying if i'm not willing to give up coffee.

- SVPfps to my output framerate*samples. samples working in the same way as in a graphics package.

- temporalsoften to blend only frames that correspond to my desired shutter angle wrt the output frame rate

- selectevery(samples, 0)

i'm able to output 2400/1001 with motion blur as if it was shot that way at 180 degrees

the interpolation quality is such that when it fails, i don't see weird things. usually just a little ghosting.

me happy

Hi Mug Funky, could you please post your script ?

I currently use QTGMC for my motion blur needs but MVTools does produce weird artefacts *, I'd be curious to see how SVPflow performs in comparison

* these aren't exactly artefacts in my case but more the typical MvTools "morphing effect" between scene changes

[kolak]

+1 for the script

[Boulder]

Originally Posted by Boulder

Hm, it's possible that I have an outdated piping client (although I have had no problems using Avisynth 2.5.8 or 2.6 earlier). I'll get the latest one and retry.

Yep, the latest version seems to work. Now some testing and if results are what I'd like them to be, I think I'm going to ask SEt if he could take a look at the modifications to put them in his 2.6 build

[chainik_svp]

The main idea is to limit each cache size (in frames), opposite to Avisynth approach when all caches uses one common "frames" pool somehow limited with SetMemoryMax value. This can easily lead to cache misses in general case BUT as far as I know that SVP's script is linear frame-by-frame only - it works.

So may be it's not a good idea to include such behavior in official Avisynth, but may be as an additional argument to SetMTMode()...

[Boulder]

I have some results. In a regular case with "DGSource-Downsize to 720p-MVTools based denoise", your and SEt's Avisynth build are pretty much the same what comes to speed. Your build uses something like 200-250M less memory (svp_cache_fwd=4 vs SetMemoryMax(512)).

In an "MPEG2Source-TFM-TDecimate-MVTools based denoise", SEt's build is slightly faster. The amount is small, 1-2%, and the difference would be even smaller if x264 was set to a slower preset (I used ultrafast to remove the encoding bottleneck). The difference in memory usage is the same as in the first example

However, a script with SRestore doesn't seem to work with your build as it crashes upon script loading in VDub and avs2pipemod. I'm not that surprised since it's a mean piece of work and if I'm not entirely mistaken, multithreading is a bit of a no-no although I don't remember anybody really confirming that. But doesn't the same apply to TDecimate as well?

[chainik_svp]

Originally Posted by Boulder

Your build uses something like 200-250M less memory (svp_cache_fwd=4 vs SetMemoryMax(512)).

720p is too easy and 250M is nothing I'm talking about situations when official Avisynth can't work at all. For example "canonical" MVTools-based interpolation script can easily run out of memory with 1080p and pel=2 in MSuper regardless of SetMemoryMax value.

[Boulder]

Hehe, that's true. I was afraid of trying to QTGMC 1080i as the tests would have a taken a veeery long time

[chainik_svp]

4k ?

[Boulder]

Yeah ----- when the day comes that I can buy a reasonably priced TV that supports the format

[Mug Funky]

for the two people interested:

function frcblur( clip c, int "num", int "den", float "angle", int "samples" )
{

num = default(num,24000)
den = default(den,1001)
samples = default(samples,7) # anything more than 11 is way overkill.  internally this number is doubled by default.
angle = default(angle,0) # 0 disables motion blur.  180 is standard cine camera.
multiplier = angle==0? 1 : round( samples/(angle/360.) )
parmstring = "{rate:{num:" + string( multiplier * num ) + ",den:" + string(den) + ",abs:true},algo:21,cubic:1}}"

c

spr = svsuper("{pel:2,gpu:1}")
vec = svanalyse(spr,"""{block:{w:32,h:32},refine:[{thsad:200}]} """)
sorathread()

svsmoothfps(spr,vec,parmstring, url="www.svp-team.com")

sorathread()
angle==0? last : temporalsoften(multiplier/2 -1,255,255,255)
angle==0? last : selectevery(multiplier,0)

}

this requires sorathread, the very easy to use threading plugin. i'm running 2.6 so i don't have much MT stuff available.

here's a script example, with stabilization thrown in:

QTinput("x:\things\MVI_1402.MOV",audio=2)

assumetff()

assumetff()
converttoyv12()

sorathread()
mdata = DePanEstimate(last)
DePanStabilize(last, data=mdata,cutoff=0.333,initzoom=1.1,addzoom=true,zoommax=1.5,rotmax=10.,tzoom=3.0,pixaspect=1.0,prev=3,next=3)
sorathread()

frcblur(num=25,den=1,angle=180)

requires the latest (still old) depan+depanestimate

i'm using QTinput because of the luma levels issue with ffms2.

[Reel.Deel]

Thanks Mug Funky! The result looks very nice.

For Avisynth MT I had to add the mt=x argument to svsmoothfps.

Sorry for being a little off topic but what are the issues with FFMS2 and luma?