Scientists publish their methods, but not while it is under development. As I've said in the past, I will post the script when the project is finished. The detail recovery is indeed a combination of details taken from neighbouring frames and deblurring. The deblurring also enhances existing detail in the frame.
DrDre said:
towne32 is doing a substantial cleanup of the GOUT Star Wars as we speak and should be finished in a few days, so I will be using that as a starting point for the rendering of the whole movie. In the mean time I will try to make a short sample for SRV11 and post a few screenshots for TESB.
Might I suggest to NOT stabilize the image until after the SR/AA process. My hunch is that the gate weave plays a factor in the recovery of image detail due to subpixel-level shift of picture within each shot.
If your crop is water, what, exactly, would you dust your crops with?
I'm not stabilizing anything. Just removing some of the more offensive grime/hair/whatever (manually, ugh). It's much better, but not complete as the GOUT SW is an absolute mess. But I already had this going so we can either use it as input or not at all. He has compared it to the straight DVD and says it looks good (mine is made straight from the DVD, encoded into lagarith lossless). If there does turn out to be any quality loss we wouldn't use it.
We might want to do the gateweave fix during the same rendering that he adds Greedo subs, later on. I'm not sure how Dre wants to do it: just particularly bad parts or just still camera shots would be easiest. If absolutely nothing else, I'd like to have the intro stabilized as the shakiness really sets the film off on the wrong foot. :) Doing that alone in adobe at the time of adding the subs would be trivial.
g-force said:
Okay, let me ask you this. How much of the improvement is from the temporal combination? How much is from all the other things that the script is doing? Scientists usually publish their methods, I have yet to see what the script is actually doing. I guess I just fundamentally have a problem with applying a bunch of denoising and spatial filters and claiming that it's all SR.
-G
Earlier in this very thread you agreed that SR is able to recover detail from the sub-pixel level on this source:
g-force said:
As AntcuFaalb correctly pointed out, and as you can see from every example in that paper, you need to have plenty of aliasing for this to work. Fortunately, the GOUT has plenty of that, at least in one dimension! For most sources that are scanned well however, the best you can hope for with such methods is just noise reduction.
-G
Then, in the previous comment, you said that SR contributes almost nothing:
g-force said:
I stand behind what was stated. Only a very small percentage of what you are considering improvements here are due to SR. What you have is a denoiser that leaves a sheet of slowly moving grain, oversharpening effects (sorry, de-blurring artifacts), and shimmering residual aliasing.
In answer to your question, SR is a plugin that (given the right source) can recover detail lost in a single frame. So whereas an avisynth script may look like this:
Source > Denoising > Anti-aliasing > Upscale > Sharpen > Downscale
SR slots in and does most of the anti-aliasing and up-scaling for you:
Source > Denoising > SR > [Further upscaling > Residual anti-aliasing > ] Sharpen > Downscale
And that's just plucked out of the air I'm not suggesting this is the exact order of filters used by DrDre, but just using it as a visual representation of where SR fits in the Avisynth script. It literally takes the place of the anti-aliasing filter. So compared to previous methods that were based on EEDI or NNEDI for the anti-aliasing it provides more detail and a more organic anti-aliasing.
[ Scanning stuff since 2015 ]
I am not sure that is right about the anti-aliasing. I think it is a separate component from the SR. I have been following this from nearly the beginning and the anti-aliasing was something that was figured out along the way after SR was implemented.
Regardless of what is making the picture look great (and I personally think a lot of it is due to SR (or deblurring) in combination with a great anti-aliasing method and a bit of sharpening), the end result was worth the journey.
The best way to think of SR is that it widens the scanlines and distributes the picture information across them. Where you once had a scanline of 1 pixel thickness, after SR you have a 2 pixel thickness scanline. This itself reduces aliasing - but the aliasing can be reduced further by using an interpolation method designed to de-interlace or reduce aliasing - like EEDI (enhanced edge directed interpolation) and/or NNEDI and/or QTGMC (and/or possibly even SR again). EEDI widens your scanlines, except by keeping the exact information present in the original scanlines and interpolating the information for the new ones. Your 1-pixel scanline is now 4-pixels, and depending on what you've done the width is either doubled or quadrupled. You then resize down to your output resolution.
[ Scanning stuff since 2015 ]
Thanks for that great explanation RU.08! I couldn't have done it better myself.
Here's a sample for SRV11 for those that are interested:
https://drive.google.com/file/d/0B8_LYKyZDiajam0wVHZRblFrVjA/view?usp=sharing
That was amazing in a way i have no clue how to put into words. Definitely still my favorite . shut up and take my money!
Preferred Saga:
1/2: Hal9000
3: L8wrtr
4/5: Adywan
6-9: Hal9000
Thanks! It looks really great and 3D. Compared with V12, I think it looks like there is slightly more detail, but has slightly more aliasing issues. Either one is great!
Since the rendering of these movies now takes a stunning 45 days at least, I've decided to release the socalled SRV10 raw first. These only take about a week to finish and have maximum detail, while reducing aliasing compared to the raw GOUT. This means the trilogy for this version should be finished in three weeks. I will also post the script for SRV10 raw at that point, such that you can make your own improvements if you wish. These versions will be available as 1080p mkv and DVD9.
Awesome! Cant wait to see what Empire and Jedi look like.
Thank you!
Pfft, 45 days is nothing. The CGI effects for Jurassic Park took on average 10 hours to render, per frame. There's about 4mins of CGI footage in Jurassic park, so that's 57600 hours (342 weeks/6.6 years!!) Mind you that was back in the day when computer hours were more expensive than they are now, these days you could afford to spend millions of computer hours for complicated effects and still achieve a short turnaround time (with a large enough bank (network) of computers of course).
[ Scanning stuff since 2015 ]
I'm more than happy to wait 45+ days per movie for the best results. Looking forward to them.
If there is anything we can do to help you, for example rendering parts or donating for your electricity bill, I'm in.
Edit: There seems to be an Avisynth TCPServer function for distributing clips over the network. However, the usability is in question and some people have reported that it did not increase speed.
The problem might be that you have to wait for the second machine to finish its script until you procede with the one on your first machine. On the first view I would say that I cannot confirm that, but I'm not so experienced in Avisynth.
Maybe someone can find out? All threads about this are pretty old.
Darth Id on ‘Why “Ben”?’:
And while we’re at it, we need to figure out why they kept calling Mark Hamill’s character “Luke Skywalker,” since it’s my subjective opinion that his name is actually Schnarzle Shnuzzle. It just doesn’t make sense!
Damn you George Lucas for never explaining why they all keep calling Schnarzle “Luke”!
Damn You!!!
Is anybody with the know-how interested in colour-correcting these? All that beautiful recovered detail with the awful GOUT colours seems like a crime.
Either Harmy's or TeamBlu's versions could be used as a reference as both have much improved colours.
George creates Star Wars.
Star Wars creates fans.
George destroys Star Wars.
Fans destroy George.
Fans create Star Wars.
Turisu said:
Is anybody with the know-how interested in colour-correcting these? All that beautiful recovered detail with the awful GOUT colours seems like a crime.
Either Harmy's or TeamBlu's versions could be used as a reference as both have much improved colours.
I've got settings that I'm going to use for my personally slightly modified color correction.
But I think Dre's actual release should be unmodified. Not that it can't be drastically improved, but you will never have any consensus or agreement on the matter. One of the main reasons I'm looking forward to this is, despite absolutely loving TeamBlu and Harmy's reconstructions , I personally am not 100% satisfied with the color in either of them. They both blow away the Bluray, but I do prefer the GOUT (which can be made to look a little less washed out without significantly changing anything). Other people love them. Other people love KK650 or age's or other LUT options. Can't satisfy everyone, so best to leave it.
Harmy himself has said that there was a misunderstanding about the source he was correcting it to and that he would do it differently. Blu's is attractive, but a reimagining.
Edit: Actually, I love Harmy's earlier color (1.0?) and TeamBlu V3. Not surprising, for me, since they're more GOUTy.
kane1138 said:
I'm more than happy to wait 45+ days per movie for the best results. Looking forward to them.
Agreed. After seeing V11, everything before it I am no longer excited about. V12 is great too. I would vote to save the week and put it towards the 45 days.
Regarding color correction, I am glad it will remain untouched. Unless it is done carefully, it can ruin everything to various degrees. I personally prefer the GOUT over what I saw of others because of the blown out details.
Turisu said:
Is anybody with the know-how interested in colour-correcting these? All that beautiful recovered detail with the awful GOUT colours seems like a crime.
Either Harmy's or TeamBlu's versions could be used as a reference as both have much improved colours.
Team Blu's '97 Special Edition preservation is also a good color source, and I would argue that it is the closest of the projects to the original coloration of the film from 77, albeit with somewhat muted colors. But pretty soon there may be a much more definitive color source in the screenshots Mike Verta is planning to release. So I'd say hold off on color correction for now, good things come to those who wait :)
You probably don’t recognize me because of the red arm.
Episode 9 Rewrite, The Starlight Project (Released!) and ANH Technicolor Project (Released!)
By popular demand: 45 days it is... :-)
Yay!!! :-)
thorr said:
Yay!!! :-)
Seconded.
Far out. Looking forward to it. :)
Perhaps we could get the script for SR10 raw since it won't see a release? ;-)
Preferred Saga:
1/2: Hal9000
3: L8wrtr
4/5: Adywan
6-9: Hal9000
Here's the script for SRV10 raw.
orig=AviSource("Star Wars.avi")
orig=ConvertToRGB24(orig)
edi0=nnedi2_rpow2(orig,rfactor=2, cshift="spline64resize") # edge directed interpolation
edi=edi0
sr1=SR(orig,1424,548) # super resolution in RGB space
sr1=ConvertToYUY2(sr1)
yuy2=ConvertToYUY2(orig)
sr2=SR(yuy2,1424,548) # super resolution in YUY2 space
rev=Reverse(orig)
sr1r=SR(rev,1424,548) # super resolution in reverse direction in RGB space
sr1r=Reverse(sr1r)
sr1r=ConvertToYUY2(sr1r)
yuy2rev=ConvertToYUY2(rev)
sr2r=SR(yuy2rev,1424,548) # super resolution in reverse direction in YUY2 space
sr2r=Reverse(sr2r)
s64=Spline64Resize(orig,1424,548)
s64=ConvertToYUY2(s64)
edi=ConvertToYUY2(edi0)
Average(sr1,0.5,sr2,0.5,sr1r,0.5,sr2r,0.5,s64,-2,edi,1) # combine SR in both directions for additional detail, and combine details from SR with nnedi2 to reduce aliasing
ConvertToRGB24()
sr=last
edicd=ConvertToYUY2(edi0)
edicd=ConvertToRGB24(edicd)
Average(sr,1,edi0,1,edicd,-1) # correct colors
nnedi2_rpow2(rfactor=2, cshift="spline64resize")
Spline64Resize(1920,816)
Are you encoding your file in parts and then joining them? That's what I would do with a long encode like that...
[ Scanning stuff since 2015 ]