Thread Rating:
  • 0 Vote(s) - 0 Average
  • 1
  • 2
  • 3
  • 4
  • 5
[Released] STAR WARS: ANATOMY OF A DEWBACK (1997) re-framed for 16:9 screens
This is not a particularly sophisticated project, and indeed although it's watchable I'm not necessarily finished with it. It was basically my first shot at mucking about with AviSynth (with some advice from this forum, actually) because I couldn't see a better way of doing it.

This is the 1997 featurette ANATOMY OF A DEWBACK, about the process of doing ungodly "Special Edition" CGI modifications to the "look sir, droids" sequence in STAR WARS (1977). It was originally released in five very short (~5 min) "episodes" exclusively on the official Star Wars website (and yeah, this was 1997 so it was a really really awful 240p, 24fps RealPlayer stream) and although later released in its entirety on Blu-ray in both 2011 and 2020, they screwed it up pretty badly both times, which annoyed me enough to try to fix it up a bit.

The original 1997 web video is long since deleted, but I kept a recorded copy of the files and recently dug them out to have a look. The video appears to be 320 x 240 pixels (4:3) but this includes black letterboxing; without the letterboxing it’s more like 320 x 192 px, which is a somewhat unusual aspect ratio of 5:3. Although this is the native aspect ratio of 16 mm film, this featurette seems to have been shot on video so it’s probably more significant to note that 5:3 was used in some countries as an early “widescreen” format for a while, presumably as a compromise between theatrical 1.85:1 and 1.33:1 (a.k.a. 4:3) home video. This original version looks something like this, if you crop off the letterboxing from the top and bottom:

[Image: Wad9yEG.png]

On the 2011 Blu-ray, the featurette was for some reason encoded to display as (almost) 5:3 “widescreen” on a 4:3 television screen. The trouble is, even in 2011, those were a dying breed, and definitely aren’t anywhere near as prevalent in 2020. The result of this is that the vast majority of people will watch this on a 16:9 screen, but the “widescreen” image will not even come close to filling the display on account of being restricted by the 4:3 box. The actual image is a very rough looking 700 x 430 or so pixels, inside a 720 x 540 pixel 4:3 frame. It’s “open matte” to some degree as it hasn’t been framed correctly for this release, but it’s also skewed toward one side, with the left side not cropped enough and the right side slightly over-cropped compared to the old web video. It also appears to have been slightly squashed horizontally. That one looks like this:

[Image: BvPsCPu.png]

The 2020 Blu-ray is different again, with the image being about 720 x 440 but this time it’s been cropped much more noticeably on the right-hand side than the 2011 transfer was. As a result, it can’t be restored back to an accurate representation of the original framing, and to be honest, it looks like crap overall when compared to the 2011 version. It’s also noticeably stretched horizontally, from less than 700 px (I’m guessing 640 px) to 720 px:

[Image: mgfxntJ.png]

Since the least cropped (and least aliased) reasonably modern source seems to be the 2011 disc, I cropped and upscaled that (without sharpening the hell out of it) to fill a 16:9 screen, in order to ditch the letterboxing and attempt to fix (as far as possible) the slightly deformed aspect ratio. Since a bit had been cut off the right hand side, I also cropped a little bit off the left to recentre the image, adding equal borders on the left and right to fill a 16:9 screen and upscaling to 720p. Since the source was interlaced, standard definition NTSC, I deinterlaced it with QTGMC. Do not expect this to look like HD footage, because it’s not, and it shows… but it’s a heck of a lot better than the 56k web streaming version and is framed better than either of the official Blu-ray Disc versions as well.

Here’s the 1997 web version with the letterboxing removed, then the remaining frame upscaled to fill a 5:3 frame inside a full-screen 16:9 display, to show what the ideal framing would look like:

[Image: G4Tlhmu.png]

Then the same thing but cropped in slightly on all sides to better match the available picture information in the 2011 Blu-ray version (since the 2011 BD version is slightly cropped on the right as well, I cropped in on all sides to restore the original 5:3 aspect ratio):

[Image: Fn4R0qT.png]

This was used to work out the most accurate crop and size adjustments for the 2011 video.

The end result is a precisely 5:3 frame with black bars at the sides to fill a 16:9 screen at 720p, instead of a tiny 5:3 frame inside a 4:3 box in the middle (which would have been great when we all used 4:3 TVs but is extremely inconvenient nowadays). It could most definitely look better and the image is still not 100% accurately re-warped to match the exact aspect ratio of the web vid but then it's entirely possible that was wrong in the first place anyway; if I knew more about upscaling and had access to tools like AI sharpening (rather than simple AviSynth filters that create ringing artefacts worse than what you already see below) then perhaps I could achieve better results. I did try just about every resizing filter I could find any info on, and this appeared to be the one that produced the least artefacts. I've also left it in the original BT.601 NTSC colour space rather than outputting BT.709, which may or may not be wise (open to suggestions, would be easy to re-render out BT.709 instead).

Anyway, here's how it ended up after the above:

[Image: Fl3CWrr.png]

If you'd like a copy, or have any suggestions on how I might improve it further, please let me know! Thanks to TomArrow and Sabbelbacke for assistance thus far.
Thanks given by: PDB , Kynch
One hell of a roller coaster ride of emotions that was!
I appreciate the level of detail you went into explaining how you went from the original source to the final result. As you said, you can only do so much based on the material you have. It's sad to think that the original tapes this was recorded on could likely be reripped for a much better result, alas…

I'd love to see this, if you could PM me a link.
Thanks given by: pipefan413
I'm doing a test encode of this just now (overnight, so I'm not tempted to sit and stare at it instead of going to bed and sleeping like a normal person is supposed to do sometimes). I rewrote the script fairly significantly to change the way I did the upscale and aspect ratio adjustment, so that it would do only one rescale instead of two, but there was absolutely no visible improvement from doing so as far as I could see and it meant that the aspect ratio adjustment was less precise which made it look slightly less correct, so I've gone back to the original method. What I've changed, however, is:

1. I'm now using gauss rescaling instead of bicubic (honestly the difference is negligible but I think I prefer the very very marginally smoother result of the gauss)
2. I'm now resampling the colour from BT.601 to BT.709 because it was suggested that some hardware might ignore the colour space declared by the file and instead just assume it should be displayed in BT.709 because it had an HD resolution (720p)

Would be grateful for any opinions on the whole colour resampling thing, by the way. I stuck to the same colour space as source to avoid unnecessary conversions but if it is indeed the case that 720p should generally be BT.709 to avoid playback issues then perhaps this is indeed the way to go.

Will report back if I can see any difference on my own hardware but it already looked correct in BT.601 so I don't really expect to!



Color range                              : Limited
Color primaries                          : BT.709
Transfer characteristics                 : BT.709
Matrix coefficients                      : BT.709

Whether it's made any difference or not, I dunno yet. What has made a very small difference however is using the gauss rescale instead of bicubic, since the crap source seems to benefit a bit from the slightly smoother result.

By the way, does anybody happen to know why there are two different Portuguese language subtitles? I'm guessing it's maybe that one is closer to the language spoken in Portugal itself and the other is maybe closer to Brazilian Portuguese, but who knows which way round it is if so. I suppose you'd expect the first option to be Portuguese Portuguese and the second to be Brazilian Portuguese but that's an assumption. I want to add the subtitles from the 2020 source as well which I'll probably get done today, and a similar thing happens there with its Spanish subtitles; I guess one might be Castellano and the other closer to Mexican Spanish, or something like that.
Thanks given by:
Right so the BT.601 to BT.709 conversion I tried using AviSynth+ (specifically, ResampleHQ) didn't work. I might be using it wrong, but if I am then the documentation / function naming is daft, because when I set the source matrix to BT.601 and the destination matrix to BT.709 it makes the colour go absolutely mental. Skin tones turn pink. So, just as an experiment, I'm running an encode just now that doesn't use ResampleHQ at all and instead just does a simple crop, upscale and stretch for aspect ratio correction with the gaussian scaler (because it produces a softer picture than the others, which all sharpen and produce really bad artefacts because of the crap quality source I've got to work with here); I'm encoding in BT.709 just to see what x264 does with it. If the colour's accurate as is, good. If not, I'll have to figure out whether to just leave it as BT.601 as I originally did, or work out how to get it resampled to BT.709 without causing inaccuracies in the colour as ResampleHQ did.

The other issue is the aspect ratio itself. My original reference was old web video (which is how this featurette was originally released, exclusively, until the Blu-ray box set in 2011). The new methodology is to look at the frames of the actual film footage in the featurette and calculate how to frame them as precisely as possible as 2.39:1 inside a 1280 x 720 pixel box. To do this, I cropped off at little as possible from the top and bottom of the frame to avoid over-cropping the camera-shot footage (which make up the majority of the featurette), and cropped off only a few pixels from the left and right to get rid of only excess black in the 2.39:1 film footage, then tried various rescales until I got 2.39:1 for the actual film picture. The camera-shot footage contains less resolution that's intended to be visible than the 2.39:1 stuff, so unless I also do an additional crop off the sides after calculating the right aspect ratio, there will be slightly untidy sides on the camera-shot footage. I had initially opted to remove those and crop in on all 4 edges to preserve the web video aspect ratio, but one of the main problems with this is that it means the name and job title text that appears to identify LucasFilm staff is waaayyy to close to the bottom edge (it's already stupidly close without any cropping).

So, before any additional side cropping, this time I've ended up with an overall visible picture frame of 1248 x 720 pixels (with 16-pixel black vertical bars on the left and right to make it up to 1280 x 720 at 1.78:1 a.k.a. 16:9). Inside that frame, the film footage when it appears is 1248 x 522, meaning it's at its correct ratio of 2.39:1. Here's the "Look sir, droids!" moment:

[Image: look-sir.png]

And here's a camera-shot bit with a name label as well to show how the rest of the footage looks (although I forgot to add the black bars at the sides before taking these screenshots to make them up to 1280 pixels wide, the ratios should be correct as long as the 2.39:1 stuff was encoded into the source video master at the right ratio relative to the other footage). Firstly, here's the "open matte" version without cropping any extra off the sides, so this is the same framing as the above screenshot of film footage:

[Image: David-Tanaka-crop-open.png]

Here it is with the sides tidied up. First, with only just enough pixels to get rid of excess black, which means the crop is uneven (10 pixels off the left, only 4 pixels off the right):

[Image: David-Tanaka-uneven-crop-10-4.png]

Finally, here it is with equal cropping on the left and right to keep the picture centred. However, obviously this means that more picture is lost from the right hand side (and the top of the left hand side, since it is curiously not a straight line up the left for some reason or another):

[Image: David-Tanaka-even-crop-10-10.png]

I think I might actually be leaning toward the "open matte" version with no destructive cropping now. Although it does mean that the sides look less tidy for the camera-shot footage (which makes up the majority of the featurette), it also preserves more of the visible picture. I'm doing a test encode like this just now, for which I've also simply left the colour matrix alone in Avisynth+ but told x264 to encode as BT.709 (I would have expected that using ResampleHQ to convert from BT.601 to BT.709 would be more correct but evidently that wasn't the case, so I'll see if this is better or ends up the same). EDIT: Yep, it looks like that did the trick, the colour looks like it's correct at BT.709 now. I think I'm happier with the crop now as well, but I might also do a more cropped-in version for a cleaner look as an option. I may even do one that crops even less to leave the top and bottom exposed like the sides are, but there's not really much being cut off at the moment (I think it's only about 1 or 2 pixels at top/bottom and they're those messy edge pixels you tend to get from tape recordings; I can only really see anything missing in one shot where someone has pointed the camera at a monitor and the title "STAR WARS" at the top is cut off either way but a tiny bit less cut off in the source than in my cropped encodes).

Thoughts? Which crop do you reckon is best?
Thanks given by:
I'm always a fan of retaining as much information as possible. I can understand your wish to maintain the centre-axis consistent throughout.

Happy to hear the BT.709 colorimetry is now playing along!
Thanks given by: pipefan413
The question now is which stretch is the most "correct"...

The new method more closely resembles the 2020 Blu-ray ratio than either the 1997 web video or 2011 BD but it assumes that the film bits to which the ratio was targeted were inserted into the original release edit in the right ratio in the first place. If not, it might mean that I've now overstretched the image horizontally, for example.

I think for a start I'll overlay the "look sir, droids" shot from something like 4K77 or the official release to check if the actual shape of the image (regardless of the rectangle it's sitting in) actually does match with this new methodology. If it doesn't, then I'll adjust until it does and see how the whole featurette then looks. If it does, then I guess it confims that I adjusted the film bits correctly but still doesn't actually tell me if the camera-shot talking head footage is right or not. Hmm.
Thanks given by:
I did some more thinking (and a lot of comparisons and calculations) and ended up deciding to revisit this with a number of changes:
  • I originally encoded in BT.601 even though I upscaled from NTSC @ 480i (60000/1001 fields, effectively 30000/1001 frames, per second) to 720p (60000/1001 whole frames per second due to interpolation during deinterlacing). The reason I did this was that I perhaps simplistically assumed that fewer conversions would be better, but the trouble is some players will just ignore the metadata and instead assume all SD content is BT.601 and HD (e.g. 720p) is BT.709, meaning the video was getting its colours displayed inaccurately depending on the player. This time, I converted to BT.709 to circumvent this problem.

  • I adjusted the crop and scale a bit. This time, I didn't target it toward the original 1997 web video (which is cropped in compared to the 2011 source) but instead compared different frames to various references, including the web video and various versions of Star Wars / A New Hope (e.g. 4K77, the new Blu-ray, etc.) to check that different shots were as close to the correct shape as possible. The trouble is that it appears that the different shots in the featurette, which are from a number of different sources with mixed frame rates and aspect ratios, have not quite been framed 100% accurately in the first place, so if I adjust one shot to make it align bang on with the film then it'll make others look worse. In the end, I settled on a slight compromise, which was ultimately very very similar to the shape of the original encode I did (and therefore very similar to the web video) but much less aggressively cropped so that no actual picture was lost from any of the 4 sides at all, just black bars.

  • I tested a number of scaling methods, including first upscaling by a factor of 2 (nnedi3_rpow2) before downsampling that to my final resolution, but everything that involved multiple rescales always looked terrible to me so I instead went for what I felt was the better approach of doing only a single rescale so that as few separate image manipulations took place as possible. In the first "finished" encode (which I'll call v1.0), I had used BicubicResize since it appeared to be a reasonable compromise between the very soft GaussResize and just about everything else (which all seemed to way over-sharpen the picture, accentuating how noisy and generally terrible it was at source), but I did two resizes, one to correct the pixel aspect ratio and another after cropping. This time, I cropped first then upscaled using ResampleHQ set to the Gaussian kernel since I decided it does look a bit better with the softer result, which helps to obscure how bad the source already was rather than making the noise and ringing artefacts even more apparent like other methods did.

  • In addition to the "open matte" version with the messy edges, I did another one that cuts more off all 4 sides to give a tidier looking frame (a bit closer to v1.0).

  • I encoded at a significantly lower bitrate of 9.8 Mbps instead of 20 Mbps seeing as the source is crap anyway and it honestly doesn't look any different to me from what I checked. I could probably drop the bitrate further with no detrimental effect but seeing as the frame rate was doubled compared to the original I didn't want to push it too low (for the same reason I left the buffer size massive even though I was dropping the bitrate, which is probably completely unnecessary and possibly obtuse).

  • I replaced the 2011 BD version's 3 chapters (which were largely arbitrary, for whatever reason) with 5 new chapters representing the original episodes released online on the official website way back in 1997.


v1.0 vs new version with least cropping vs new version with more cropping (notice colour is different in v1.0, he looked a bit jaundiced):

[Image: big-G-15000-dewback-WIP.png]
[Image: big-G-15000-dewback-NEW.png]
[Image: big-G-15000-dewback-TIDY.png]

Spline36 vs Gaussian resampling kernels for the rescale (notice increased sharpening with Spline36 compared to much softer Gauss):

[Image: big-G-15000-up-Spline36zoom.png][Image: big-G-15000-up-Gausszoom.png]

"Look sir, droids" moment in 4K77 vs new ANATOMY OF A DEWBACK version:

[Image: looksir-94056-4-K77.png]
[Image: looksir-94056-dewback-narrow.png]

The v1.0 AVC encode was about 3.67 GB, whereas the new AVC encodes (with the two different crop options) are approximately 1.79 GB each. The source MPEG-2 stream was 1.27 GB.
Thanks given by: Kynch

Forum Jump:

Users browsing this thread: 1 Guest(s)