Search options (Page 1 of 2)

Yeah of course, that's why it would have to be the sale price that increases by $5-$10 to pay for the extra discs (assuming my estimate of disc costs was correct... for all I know it may be way off).

As far as I can tell Mill Creek isn't in the business of colour grading. If I had to bet money on it I'd say they are probably just doing the disc authoring & printing/pressing.  I could be wrong.

I wouldn't even mind the lower 4.5mbps bit rate of the Mill Creek release if they had at least used the same masters as the Universal version.  As an example I just bought this knowing full well that one of the reviewers has mentioned the reduced bit rate due to the lower disc count compared to the original 2008 CBS release.  Although if it has serious mastering errors it's going back on ebay.

You'd think over time the image quality would get better, but with DVD they are actively making it worse.  Which really hurts since DVD is already suffering at 720x480 standard definition, and there is often no better option due to the streaming services being even worse (except of course for shows that were remastered in HD at great expense to the studio).

Given the surprising popularity of DVD some 28 years after its conception, maybe they should consider coming out with a "DVD+" spec for better quality video while retaining backwards compatibility and cheaper production costs compared to bluray.   For example the DVD+ spec could specify a minimum bitrate and have predefined encoder profiles agreed upon by industry experts.  They could retire the interlaced format and require all video be stored on disc as progressive, and if generated from an interlaced master it would need to comply with a certain standard of deinterlacing & ivtc defined by industry experts.   The raster size should also be expanded to 852x480 for NTSC and 1024x576 for PAL to retain 4:3 pixel density at 16:9.

Those ones I did at 24fps with pretty standard x264 settings (-tune film -pix_fmt yuv420p -b:v 8000000).  Not sure why your box can't play it... out of curiosity could you try this 8mbps x264 that someone else did so I can learn hopefully about what makes files more compatible with devices - https://drive.google.com/file/d/1Q2o0Cz … sp=sharing  (353MB)

Sounds about right.   This is actually the problem I've been trying to solve the past few months with an Avisynth script which I'm calling "adaptive IVTC" (90% complete).   Basically you can throw any content at it and it will give you the best conversion from interlace to progressive without having to configure anything.   Optionally you can set it to a profile to tell it more info about how much of the content is film-based but you don't have to, it will still work at the default setting with high accuracy, and for the scenes gets it wrong it errs on the side of deinterlacing with QTGMC antialiasing which still looks pretty decent and in a double blind test can be hard to tell apart from IVTC.

But it's a transcode of a transcode of the DVD, so it will look considerably worse!

This is why I'm such a stickler about using Avisynth in realtime to watch episodes, as it's the only way to do video processing without any generation loss.   It also saves storage space - 5mbps for the DVD vs 10mbps for the x264.  And no need to burn your CPU for days to render the files.

The only issue with realtime playback is you need enough CPU power if you want to do a lot of advanced AI upscaling in realtime.  However Avisynth+ supports multithreading & if I enable it by putting a Prefetch(8,6) at the end of the script I posted above, it manages to play in realtime at around 90% CPU load on an i5-4570 (4 cores 4 threads @ 3.2Ghz) which is not a very powerful machine at all.   btw Prefetch(8,6) means use 8 threads and prebuffer 6 frames in advance.  I found that even on my 4 core machine, 8 threads still loads up the CPU better.

QTGMC softens the image slightly...

We can control the amount of smoothing/softening by adjusting the TR0, TR1 & TR2 parameters, but it still won't retain as much texture detail as not using QTGMC at all.

Stunt actor

Here's a clip of Luck of the Draw PAL DVD (remuxed, no transcoding, no processing at all)

https://drive.google.com/file/d/1xsgovx … share_link

You might notice their voices are chipmunked due to the 24->25fps speedup - this can be reversed with AssumeFPS(24000, 1001, true) to slow the video and audio back down to 23.976fps same as the NTSC version.

Also there seems to be more film jitter in Luck of the Draw.  That thing where the film frames are jittering slightly in the spool causing the whole image to move around slightly by a few pixels.  QTGMC repair was automatically correcting that with its stabilisation so you might notice it in that clip since I disabled QTGMC.  If  that bothers you it's possible to correct it with something like

# Image stabilisation - http://avisynth.nl/index.php/DePan
mdata = DePanEstimate(last)
DePanStabilize(last, data=mdata, initzoom=1.025) 

I think initzoom=1.025 means crop 2.5% around image edges as buffer zone for image stabilisation.

The latest version of TFM (v1.0.27) appears to be broken when feeding it 60fps via DoubleWeave(), which is what the script I posted here uses.   

For some reason v1.0.27 fails to fieldmatch a 60fps doubleweaved input, whereas the previous version (v1.0.26) succeeds.  This is the version I've been using since last year.     

I tried using all the fieldmatching modes but they all produce the wrong result - either failing to fieldmatch or producing the wrong output cadence (2:2:2:4 instead of 3:2).

Failure to fieldmatch means combed frames which triggers deinterlacing, so if you were using v1.0.27 you'd regularly be dropping to 240p on moving objects.

Had a look just now with just TFM.TDecimate to see the raw frames, and I can't really see anything exceptional apart from the 1:1 sections at around 11:00 - those will look terrible with stutter and combing as TFM wasn't designed for deinterlacing 1:1 sections (still working on my adaptive deinterlacing script to solve this... around 1500 lines so far).

The field alignment seems no worse than something like Eggheads, and there's definitely 480p to recover from the source after IVTCing with TFM.TDecimate, judging by the smoothness of diagonal or curved lines.

I am sharpening the image with MadVR though - subjectively I find standard definition unwatchable at 0 sharpness.  For maximum effect sharpening should be applied before upscaling otherwise the sharpener is looking at soft upscaled pixels which don't sharpen as well.

Is the issue that the image doesn't look sharp enough?   Or too noisey/grainy?

Hmm that's weird, CAS is almost free on my desktop system (i5-4570 4 cores @ 3.4ghz).  AMD claims it's "highly optimized".

QTGMC and nnedi3_rpow2 are very CPU intensive... try commenting everything out except CAS, toggle CAS on/off and compare CPU before/after (give it about 10 seconds of playback for buffers to fill... after that my CPU usage settles around 7-10%). 

If you're still getting high CPU try playing with the opt parameter (opt=0/1/2/3) in case the default setting (-1) isn't correctly autodetecting your CPU capabilities.  If it's a mobile CPU that might have something to do with it.  The slowest should be opt=0, fastest should be 1 or 2.

Avisynth can also suffer from bottlenecks depending on what order you call filters in, eg. if there is a slow filter higher up the script, that can grind everything under it to a halt, but I don't think that's an issue here.

edit: also QTGMC's stronger smoothing (TR2=3) is very expensive on my system - I'm seeing a 30% increase vs default (TR2=1).

Contrast Adaptive Sharpening seems to be decent - ported from AMD FidelityFX.

Preview: https://drive.google.com/file/d/10HEOoD … share_link

Would recommend viewing at normal viewing distances.

Script:

# convert to HD colour
z_ConvertFormat(                        
\ colorspace_op="601:601:170m:full=>709:709:709:full",
\ resample_filter="spline36",
\ dither_type="ordered",
\ interlaced=true )

# 60fps IVTC
DoubleWeave().TFM(mode=0, scthresh=100, micmatching=1, ubsco=true, mmsco=true,
\   display=false, slow=2, PP=3, metric=0, cthresh=9, MI=80, hint=false,
\   clip2=propdelete("_FieldBased").bwdif(field=-2, thr=2, edeint=nnedi3(field=-2)))

# QTGMC repair - high strength, no sharpening
QTGMC(InputType=1, TR2=3, preset="slow", EdiThreads=2, Sharpness=0.0, Rep0=13)

# Contrast adaptive sharpening - 90% strength
CAS(sharpness=0.90, y=3, u=2, v=2, opt=-1)

# 4x AI upscale to 1920p
nnedi3_rpow2(rfactor=4, nns=1, nsize=0, cshift="Spline36Resize")

# Downscale to 1080p
Spline36Resize(1440, 1080)

I'm not really seeing that in the screenshots I recently posted - is there a particular episode which is problematic?

Combing should not be an issue if you have TFM's post processing enabled and an MI value of around 80 to 100.  Unless it's 1:1 cadence stuff like the Mindgame sequences in Eggheads - that's what I'm currently working on to solve using a latching system where once 2 or 3 combed frames start coming out of TFM, the rest of the scene gets deinterlaced regardless.  And then after n latched scenes, declare the content type as video and start every scene deint latched until the latch is broken the other way, and vice versa for film based scenes.

Avisynth's built in Sharpen() is poor imo, but there are other third party sharpeners here.  Although many are 10 years old and don't support the latest 64-bit version of Avisynth.

I haven't played with any of them as I do all my sharpening in the video renderer (MadVR) which runs much faster on the GPU and has good customisation like "thin edges" and "enhance detail" etc. which I personally like.  Here are some comparisons...

1. Original , sharpened
2. Original , sharpened
3. Original , sharpened
4. Original , sharpened

edit: the above are all from Eggheads but none seem to be visibly affected by the field alignment issue.  Here is a shot from the same episode where I think the field misalignment just happens to be occurring where his eyes are positioned: https://lensdump.com/i/ES6w8K

Just checked that scene on my end and it seems the culprit is QTGMC's sharpness=1.0, as setting it to 0.0 or commenting out QTGMC gets rid of the aliasing on the radiator grille.   

Usually radiator grilles cause aliasing due to the grille looking like combing artefacts to TFM, triggering deinterlacing and its associated aliasing on moving objects, but even at MI=40 it's still not getting detected as combed so it's not that.

Looks like QTGMC sharpening should be avoided then.  I do all my sharpening in MadVR so that's probably why I didn't notice it - sorry.

I know EXACTLY what you're talking about.  Funny thing, the special effects actually look BETTER on the DVD.   They are animated at a higher frame rate (30/60fps on many effects which gets bodged down to a stuttery 24fps for the HD version) and the colour of the effects is better matched to the scene they are superimposed onto.

Yes!   My Sony DVD player for instance, when playing a PAL 2:2 cadence DVD (pretty much all PAL DVDs) it inverse telecines (weaves) to 576p.   But with NTSC 3:2 cadence DVDs like Sliders it just deinterlaces using BWDIF style deinterlacing (weaving only static elements, but anything that moves drops to 240p).  So I think that could explain the difference you were seeing.

Also for NTSC DVD's there are ones that are authored as "soft telecine" where the video is stored as 24p on the disc and there are "repeat field" flags in the video stream telling the DVD player when to add fields to make a 3:2 pattern for 480i output.   On these discs my Sony DVD player appears to ignore the repeat flags and just outputs 24p.  But Sliders DVD uses "hard telecine" where it's stored as 30i on the disc with the 3:2 pattern baked in, and the only way to recover 24p is to use inverse telecine (eg. TFM) which my Sony DVD player doesn't appear to support, so it just deinterlaces every frame BWDIF style.

I'm having the Lensdump issue as well - seems to be a problem with their site.  Disabling my adblocker seemed to fix it but now it's playing up again.  The reason I use Lensdump is cause they don't compress the images.   I'd imagine they are aware of the issue and will fix it soon.

Honestly I just think it's great that we have these kinds of tweaking capabilities - I'm constantly editing my .avs files and tweaking my MadVR tags to get things looking "better" for certain seasons of shows and it's great.

Interpolation is really just a synonym for upscaling - taking some samples like pixels and making a guess at what the pixels in between them might be based on the surrounding pixels.

The simplest form of deinterlacing is Bob deinterlacing which just takes each 720x240 field (60 of these per second in a 30fps interlaced file) and upscales them to 720x480, typically using Bicubic or Bilinear resizing in the vertical direction.  If you want to see what it looks like type Bob() in Avisynth.  The result is soft and flickery, but frame rate is perfect and combing artefacts are impossible.

QTGMC deinterlacing is like Bob deinterlacing but NNEDI3 upscaling each field from 720x240 to 720x480, followed by some advanced techniques to smooth and antialias the image.  I believe it looks at the previous/next field as part of this process.  But it never weaves fields together - a test pattern of alternating 1px white/black horizontal lines always becomes solid white/black in each output frame, just like with Bob.  I would characterise QTGMC deinterlacing as the best possible version of Bob deinterlacing.  It tremendously repairs the field misalignment issue in the Mill Creek release.

QTGMC "repair" mode doesn't deinterlace at all, it just looks at progressive frames and applies its antialising magic to it.  Film grain is slightly suppressed but the 1px black/white lines are fully resolved.  Field misalignment issue is not repaired. 

It depends what the cadence is.  1:1 cadence is where every field is a temporally unique image captured from a new moment in time, like the Mindgame scenes in Eggheads.  Sports, news, talk shows and soap operas tend to use it.  If we weave every pair of 2 fields together on a 1:1 sequence the resulting image is combed:

top field        A-C-E-G-I-K-M-O-Q-S
bottom field     B-D-F-H-J-L-N-P-R-T
combed?          Y Y Y Y Y Y Y Y Y Y

(letters indicate temporally unique image)

To avoid combing on 1:1 we can use Bob, but a better solution is BWDIF (bob-weave deinterlace filter) which uses Bob deinterlacing for moving parts of the image and weaves any static parts.  Result is 480p60 same as Bob.  From memory this method of deinterlacing became mainstream around the late 2000's and was touted as "per pixel deinterlacing".

If the cadence is 2:2 (every 2 fields = a temporally unique image) like the spinning earths in the intro, weaving them results in 480p30:

cadence          2 2 2 2 2 2 2 2 2 2
top field        A-B-C-D-E-F-G-H-I-J
bottom field     A-B-C-D-E-F-G-H-I-J
combed?          N N N N N N N N N N

If the cadence is 3:2 (99% of the show) we get:

cadence            3 2 3 2   3 2 3 2
top field        A-A-B-C-D-E-E-F-G-H
bottom field     A-B-C-C-D-E-F-G-G-H
combed?          N Y Y N N N Y Y N N

TFM field matching applied to the above results in:

cadence            4 2 2 2   4 2 2 2
top field        A-A-B-C-D-E-E-F-G-H
bottom field     A-A-B-C-D-E-E-F-G-H
combed?          N N N N N N N N N N
duplicate?         Y         Y 

TDecimate can then remove the 1 in 5 duplicates to produce 480p24.

So we have a mix of 480p24, 480p30 and 480p60 coming out of the original 480i30 container.  If your goal is to preserve these original rates, they would fit best inside a 480p60 container, which is what I'm currently working on in Avisynth.  99% of the show is 3:2 though, so most people would probably be fine with putting it all inside a 480p24 container and tolerating the odd bit of stutter/combing on 2:2 and 1:1 sequences.

Yes, although I think the field misalignment issue is not too severe on the Universal NTSC discs.

Seems to be fine, although I'm not familiar with GaussResize.  The wiki says it has an adjustable sharpness parameter p so if you wanted to control final sharpness you could do eg.

nnedi3_rpow2(rfactor=4, cshift="GaussResize")
GaussResize(1440, 1080, p=30.0)    #1.0=blurry, 100.0=very sahrp

I'm just using the standard LanczosResize() with default settings which I believe is 3 taps. 

I don't normally do the upscale in Avisynth as it uses CPU so I prefer to let the media player do it on the GPU.  There's nothing particularly special about Lanczos, I just like it cause it's inexpensive and doesn't look objectionable to me.  Some people complain it's too ringy (overshoot artefacts) but I don't seem to notice it.

Depends on how good your TV is at upscaling 1080p to 4k - I'd expect it should be perfectly fine & I'd be surprised if it was bad.   Although current model LG OLEDs have surprisingly low sharpness at max sharpness setting.

For future reference all TIVTC and QTGMC dependencies are in a zip file here & I've updated all previous references in this thread to point to it as well.

I don't believe so, for example QTGMC(preset="slow", Rep0=13, EZKeepGrain=1.0) is working on my system with just those dll's I put on gdrive.  I double checked by removing all other dll's from my C:\Program Files (x86)\AviSynth+\plugins64 folder and putting only those ones on gdrive in there. 

What error message do you get when doing eg. QTGMC(preset="slow", Rep0=13, EZKeepGrain=1.0)?

I do believe QTGMC may have some extra dependencies when using "very slow" and "placebo" presets though - those should be available in "optional plugins" here: http://avisynth.nl/index.php/QTGMC#Requirements

Yes, so you'll need:

1. TIVTC dll's: https://drive.google.com/file/d/1IgIze3 … FOW5F/view
2. QTGMC dll's: https://drive.google.com/file/d/1c1Duyt … YnXZx/view

But I forgot about avsresize.dll - that one is needed to do the colourimetry conversion from SD colours to HD colours if you are upscaling in Avisynth.

The first thing I'd do is confirm you've got Avisynth up and running by making that hello world Version() clip.

Once you've got that running in MPC-HC, the next thing I'd recommend is set MPC-HC's audio render to "MPC Audio Renderer" as that seems to be required for compatibility with Avisynth audio (in MPC-HC options -> playback -> output -> audio renderer -> MPC Audio Renderer).

Also in MPC-HC options -> internal filters -> source filters -> tick Avisynth.  This will make MPC-HC use its own internal LAV filter to open .avs files, which in my experience is better than the default Microsoft one you will probably get if it's unticked.  If you have any troubles, right click the video image while an .avs file is playing -> Filters -> copy filters to clipboard, and paste the result here.

If you can get to this stage then you should be good to go with writing Avisynth scripts and previewing them in realtime without having to spend hours transcoding to mp4.   QTGMC's "medium" or "fast" presets may be useful too.  Don't forget to enable multithreading with Prefetch(cores) at the end of the script, as QTGMC needs multithreading to be fast enough for realtime use.

I'm currently working on a script that will default to outputting 60fps deinterlaced frames thus preserving frame pacing.  Then if a deinterlaced frame is similar to the TIVTC'd frame within a threshold, swap it for the TIVTC'd frame.  This should give the best of both worlds - TIVTC 480p resolution on film sections (most of the show) and autoswitching to 60fps deinterlaced goodness on video sequences (like Eggheads mindgame, intro sequence, etc.).

edit: here's a preview.  Still a work in progress, trying to adjust the thresholds to handle the switching behaviour on low motion

For completeness here's a short clip of that scene with vs without QTGMC repair - the difference is quite astonishing: https://drive.google.com/file/d/1D4PEbu … share_link

Just to clarify regarding comb lines or "combing", those aren't defects - interlaced video is supposed to have combed frames.  It doesn't necessarily have to contain combed frames, but it's expected and normal if it does have combed frames.

The way I'd summarise it is like this: 480i30 is the same as 480p30 in that it still has 30 frames per second of 720x480, except with 480i30 each frame may or may not contain two different images inside it - one in the even lines and the other in the odd lines (top field and bottom field respectively).  If such frames are displayed "as is" then it looks combed as we're seeing two different images from two different points in time weaved together. CRT's would display each field seperately one after the other so the resulting image wouldn't look combed.

Not all frames in the source are necessarily combed though, eg. the spinning earths intro animation is effectively 480p30 - both even and odd lines are from the same image, so it doesn't even need any processing at all and will look identical to 480p30.   The majority of the show uses a pattern of 3 progressive frames followed by 2 combed frames - a lossless method of storing 24p inside a 30i container. 

In an ideal world the entire show would just follow this predictable 3-2 pattern (cadence), but in practice the source may contain a hodge podge of different cadences and/or random cadence breaks. eg. some sequences where the camera is panning with the 3-2 pattern, and then they'll overlay credits half way through a frame, momentarily breaking the 3-2 pattern. Some frames either side of a scene change may have unexpected combed frames due to splicing half way through a frame.  In Eggheads when they're playing Mindgame, all frames are combed. 

Dealing with all of these cadence changes presents a massive headache if your goal is to figure out which fields can be weaved with others to extract as many progressive frames as possible.  TIVTC does this quite well, but inevitably ends up with leftover combed frames which contain fields that couldn't be weaved with any others due to the aforementioned idosyncrasies.  This is where TIVTC's "post processing" comes in - it detects and deinterlaces any such combed frames leftover after the field matching process.  This relies on combed frame detection which is prone to false positives or false negatives.   But when a combed frame is detected, it uses motion compensated deinterlacing which maximises resolution by weaving parts of the image which are static, and upscaling ("interpolating") parts which are moving.  So anything that moves basically drops to 240p, but anything static has full 480p resolution.  The drop to 240p on motion causes aliasing, shimmer and moire, but QTGMC's "repair progressive" mode (InputType=1) does a remarkable job at cleaning this up and making it look like 480p (except for moire patterns). 

Separate to this is QTGMC's deinterlacing mode (InputType=0) which is intended to be used on its own and not in conjunction with TIVTC. In this mode, moire is present even on static elements, because QTGMC deinterlacing never weaves fields to progressive frames - it is a completely novel solution which upscales every field from 240p to 480p, and then does some very clever blending with neighbouring fields, using motion vectors and neural network antialiasing on edges to smooth everything out.  The actual amount of resolved resolution on static elements is somewhere around 360p according to a test pattern.  On moving objects it visually appears similar to 360p as well, although I haven't measured it.  This is great for low quality sources which never contained 480p resolution to begin with (especially the Mill Creek DVDs as it blends and smooths the field alignment issue) but for higher quality sources I prefer its "repair progressive" mode (InputType=1) applied to the motion compensated deinterlacing.

For Season 1 I will probably use TIVTC + post processing motion compensated deinterlacing + QTGMC "repair progressive" mode.  The repair progressive does slightly suppress film grain though - on higher quality sources it does tend to look a little "processed".  But it does get rid of aliasing and shimmer on those occasional deinterlaced frames to the point where I have a hard time telling whether I'm looking at TIVTC'd frames or deinterlaced frames, except for high frequency patterns which still have noticeable moire.

Sorry, looks like I had a brain fart and misread what you wrote

That's true, but then you will have some combed frames on certain sequences like during the intro, or certain scene changes that weren't spliced correctly, or when they're playing Mindgame in Eggheads.   

If occasional combing doesn't bother you then you should use it as it retains full 480p resolution on 24fps and 30fps sequences which make up 98% of the show. 

If TIVTC had more reliable combing detection this wouldn't be an issue.  We can probably improve it for this episode by playing with the combing detection thresholds (cthresh and MI parameters) but in my experience this only tends to provide a choice between false positives or false negatives.

But it's still an improvement from no post processing at all, so try setting MI=200 which eliminates false positives from that street scene (no moire).  The highest number of combed pixels for that scene was 195 inside the square by the window, so setting MI to 196 or above would avoid moire on that scene:

Just be aware you may get other scenes that could be combed and have less than 200 combed pixels and those would go undetected (false negative) and remain visibly combed.

To get that debug overlay in Avisynth: TFM(PP=1, display=true)

Here's that first qtgmc preview again, but this time tagged with SD colours - specifically "-colorspace 6 -color_primaries 6" in ffmpeg (6 = rec.601/smpte170m)

qtgmc 720x540p60 Rec.601 tagged.mkv

I'd expect VLC should read the tags and do the conversion properly.  Without the tags VLC probably just assumed nothing and didn't do any conversion, which is fair enough.

Yep, it's way off on the old SD one, and I'm pretty sure it's cause your media player isn't converting SD colours to HD colours to suit your display mode which uses HD colours.

But that's my fault cause I assumed that media players would see the file's resolution is SD and convert them to HD colours, as that's how mine operates - https://imgsli.com/MTY1NDcz

I'll run off another copy with the mkv file properly tagged for SD colours and then it should look fine on your end and all media players.  Sorry about that - good catch.

edit: btw, the colour is still slightly different on mine too, and I think it's cause my media player uses slightly different formulas to do the conversion compared to the Avisynth filter I used.

New discovery...QTGMC has a separate "repair progressive" mode which can provide a significant improvement when applied to all frames post-TIVTC.  This mode doesn't deinterlace at all, it just "repairs" progressive frames in a way that reduces deinterlacing artefacts, aliasing, moire, shimmer etc. without smudging out any resolution in the source. I ran it through a resolution test pattern and 1px alternating white and black lines retain full contrast.

The downside is that it doesn't provide any benefit to video that was already deinterlaced by QTGMC, but it greatly reduces post-TIVTC deinterlacing artefacts such as:

1. "Eggheads" the moire on the wall shingles are suppressed - clip
2. "Prince of Wails" the Jeep radiator grille is now perfectly rendered
3. "Last Days" opening street scene with the tree branches is perfectly rendered

Because it only works post-TIVTC, you have to forfeit the perfect 60fps frame pacing that you would get with QTGMC deinterlacing only, such as the scenes in "Eggheads" where they're playing Mindgame (comparison: QTGMC 1080p60 , TIVTC + QTGMC repair 1080p24)

Anyway, I'm super stoked with this discovery as it means I don't have to worry so much about false positives with TIVTC's combing detection, as those frames will come out looking nice after being repaired by QTGMC.   I'm also using it on Rugrats (streaming service version) and it beautifully smooths out all the deinterlacing artefacts baked into it.

It's a side effect of field interpolation, which is what deinterlacing does, and QTGMC is a deinterlacer that suffers from this as well.   I had mentioned it here (check out the sleeve on her shirt).

The only way I could suppress it is to blur/blend pixels with the one above/below (which is the left side of that 4th image comparison) but then the image becomes quite a lot softer to the point that TIVTC alone would be the clear winner imo.

edit: just checked that scene with TIVTC and it's no better, cause the wall texture has alternating 1px patterns which gets falsely detected as combing, triggering TIVTC to switch to deinterlacing mode for those frames, resulting in moire.    Setting the combing detection threshold higher isn't an option as I've already set it so high that there are some false negatives creeping in.   The way this would be handled by a restoration studio would be to flag that scene as progressive to force deinterlacing to not kick in just for that scene.   A better solution would be to have better more reliable combing detection in the first place.

Yeah, it's probably better for us to produce 1080p files as it takes out of the equation the unknown upscaling quality of the end user's media player or TV, which could potentially be poor.   

These are the scalers included with Avisynth: http://avisynth.nl/index.php/Resize

The most popular ones seem to be BicubicResize, Spline36Resize and LanczosResize.  They can be tuned for sharpness:

BicubicResize(1440, 1080, 0.0, 0.75) means upscale to 1440x1080 with blur=0.0, sharpen=0.75

LanczosResize(1440, 1080, taps=4) is sharper than LanczosResize(1440, 1080, taps=3), the latter of which appears very close to Spline36Resize.

There is also this Sharpen() filter included with Avisynth: http://avisynth.nl/index.php/Sharpen

There are a bunch of third party sharpeners here http://avisynth.nl/index.php/External_f … Sharpeners which I'm sure can do a much better job than Sharpen(), but I haven't delved into them as I do my sharpening through the media player (MadVR).

If we upscale to HD we must remember to convert the colours to HD otherwise they will be quite off:

(animated .png)

In Avisynth...

# convert colourimetry from NTSC to HD 
z_ConvertFormat(                        
\ colorspace_op="601:601:170m:full=>709:709:709:full",
\ resample_filter="spline36",
\ dither_type="ordered",
\ interlaced=true )

AI upscaling:

nnedi3_rpow2(rfactor=4, nns=1, nsize=0, cshift="Spline36Resize") # AI upscale 4x (720x480p -> 2880x1920p)
Spline36Resize(1440, 1080)                                       # downscale to 1080p

Preview:

QTGMC + AI upscaling (1080p60).mkv
TIVTC + AI upscaling (1080p24).mkv