logo Sign In

DrDre

User Group
Members
Join date
16-Mar-2015
Last activity
18-Apr-2024
Posts
3,985

Post History

Post
#764195
Topic
Star Wars GOUT in HD using super resolution algorithm (* unfinished project *)
Time

Here are the first screenshots for super resolution v4. I'm rather proud of the result. It has enormous detail, but little noise or artifacts. I think it achieves a natural high definition look. What do you think?

These screenshots are all from the Tantive IV scene. I think you all would agree that was a pretty noisy scene. Here's what I've been able to do with it:

http://screenshotcomparison.com/comparison/122567

http://screenshotcomparison.com/comparison/122568

http://screenshotcomparison.com/comparison/122569

http://screenshotcomparison.com/comparison/122570

http://screenshotcomparison.com/comparison/122571

Post
#764062
Topic
Star Wars GOUT in HD using super resolution algorithm (* unfinished project *)
Time

@ NeverarGreat

The complete movie is being processed as we speak. I decided to go for slightly less denoising. The Tantive IV scenes had very little noise compared to the source, but in close ups details like facial structure were somewhat smoothed away. It did look very clean though, so maybe some of you will prefer that version. The good thing about all the added detail, is that you still have a good amount of detail compared to a standard upscale even if you denoise rather strongly. So it doesn't have that cartoony look that many of the denoised GOUT upscales have. I will post comparisons for both versions later today. 

Post
#763897
Topic
Star Wars GOUT in HD using super resolution algorithm (* unfinished project *)
Time

Inspired by the discussion with RU.08 I decided try to reduce the noise. It appears that due to the high resolution of the upscale, much of the noise inherent to the upscaling method can relatively easily be removed using an advanced filter, without a significant detail reduction. Here are some comparisons between the noisy and denoised SRV3: 

http://screenshotcomparison.com/comparison/122096

http://screenshotcomparison.com/comparison/122097

http://screenshotcomparison.com/comparison/122098

Here are the comparisons with the Avisynth Spline64Resize:

http://screenshotcomparison.com/comparison/122116

http://screenshotcomparison.com/comparison/122117

http://screenshotcomparison.com/comparison/122118

The noise is now generally less than for the Avisynth Spline64Resize and less than for the source material. 

Here's a denoised video sample:

https://drive.google.com/file/d/0B8_LYKyZDiajY1YwNGFHZmdEZUU/view?usp=sharing

The detail is retained, but the cost has been reduced. Thanks to RU.08 for keeping on the pressure. ;-)

Post
#763892
Topic
Star Wars GOUT in HD using super resolution algorithm (* unfinished project *)
Time

@ RU.08

I understand, the problem is I don't find the loss of detail to be acceptable. You don't like the noise enhancement, that's fine. The price of reconstructing detail is that the "detail" in the noise is also enhanced (the haloing is actually part of the source, but is enhanced). However, the cost of your method is equally large, but you accept it, because it's part of the source. To me that's not good enough. The reference should be the high resolution source (in this case the print the GOUT is based on). The focus is put on undesired noise enhancement. However, to me what was lost in creating the GOUT dvd's is just as important. I've provided a way of regaining to a large extend what was lost, at what I think is an acceptable price. As I said, it's a matter of taste. Your upscale has less noise, but lacks detail. Mine has detail, but has noise. 

Post
#763879
Topic
Star Wars GOUT in HD using super resolution algorithm (* unfinished project *)
Time

@ RU.08

The last one looks much better indeed. Yours certainly looks cleaner with less artifacts. It certainly is a lot better than a simple Spline64Resize. Compared to my method, the downside is that there's much less detail. Personally I'm not a big fan of sacrificing detail to gain a smooth image, but that's just a matter of taste. 

My goal is to obtain a high resolution version of the GOUT, warts and all, so I've accepted that the final product will have excessive grain, and a number of other flaws inherent to the GOUT.

I'm aware that for many of you who are expecting a completely restored original trilogy, my grain filled high definition GOUT isn't good enough, but it's what I set out to do. As far as GOUT based upscales go, at least there's now another choice on the menu. Mine is a little rough around the edges, but I do believe it's the most detailed GOUT upscale to date, and I guess that may be of interest to some.

Post
#763876
Topic
Star Wars GOUT in HD using super resolution algorithm (* unfinished project *)
Time

@ RU.08

A closer inspection of our methods also reveals a large amount of distortion in your upscale, compared to the Avisynth Spline64Resize. Just look at Vader's buttons. Apparently, the algorithm's desire to simultaneously sharpen edges, while smoothing noise, causes it to take liberties with the shapes in the image.  

Avisynth Spline64Resize vs "proper' upscale (zoomed 3x)

http://screenshotcomparison.com/comparison/122072

Avisynth Spline64Resize vs super resolution v3 (zoomed 3x)

http://screenshotcomparison.com/comparison/122073

Post
#763870
Topic
Star Wars GOUT in HD using super resolution algorithm (* unfinished project *)
Time

@ RU.08

Although your upscale is much smoother and certainly removes many of the artifacts in present in the source, it introduces an unrealistic smoothness to the frame. In that sense it literally is a smoothed and cleaned up low resolution image on steriods.  There is none of the detail or structure present that is native to high resolution images. In a sense you eliminate both the gain and the cost of my methodology, but essentially introduce a different cost. I guess it's a question of preference. Personally, I prefer mine (big surprise), but I can understand why you would prefer yours.  That's why I would indeed describe yours as a "proper" upscale, but mine as a "proper" high res reconstruction. 

However to get a better sense of the quality perception of the upscale/reconstruction, it would be interesting to see yours in motion and compare it to mine. I would say take the low res sample I put up and upscale it with your method for a proper comparison. 

Post
#763856
Topic
Star Wars GOUT in HD using super resolution algorithm (* unfinished project *)
Time

Here is the first video sample of Star Wars SRV3:

This is the low res source material:

https://drive.google.com/file/d/0B8_LYKyZDiajVG1UWGhfcTMyTkU/view?usp=sharing

This is the high res reconstruction:

https://drive.google.com/file/d/0B8_LYKyZDiajRENkTjdKTGFMcG8/view?usp=sharing

For full quality just download the samples.

It's not Harmy's Despecialized Edition, but considering it's based on a single low resolution source of rather poor quality, I think it looks pretty good. 

Post
#763706
Topic
Star Wars GOUT in HD using super resolution algorithm (* unfinished project *)
Time

Yes, it does look like a film scan in bad need of cleaning, but I guess that's precisely what the state of the GOUT was back in 1993. These defects were not visible on Laserdisc, but certainly come out now.

However, this is not a restoration job. The task I set myself was to reconstruct the high definition master of the GOUT. This will of course not result in a perfect representation of the three films, but it is one way of obtaining a high definition version of the original trilogy. As far as I'm concerned that more or less what's been achieved (aside from some imperfections like aliasing). 

There are other people on this forum who are far more qualified to perform some sort of restoration. 

Post
#763655
Topic
Star Wars GOUT in HD using super resolution algorithm (* unfinished project *)
Time

The post-processed Super Resolution GOUT Star Wars is finished. Here are the first screenshot comparisons. I have to say it looks amazing, aside from some aliasing issues present in the source material. I never expected it to look this good...

http://screenshotcomparison.com/comparison/121967

http://screenshotcomparison.com/comparison/121970

http://screenshotcomparison.com/comparison/121971

http://screenshotcomparison.com/comparison/121972

http://screenshotcomparison.com/comparison/121973

http://screenshotcomparison.com/comparison/121976

http://screenshotcomparison.com/comparison/121978

See the reconstructed detail in Vader's mask in the last comparison.

Post
#763631
Topic
team negative1 - star wars 1977 - 35mm theatrical version (Released)
Time

I'm kind of new to this forum, and catching up. Came across it when I discovered Harmy's despecialized editions. I just want to say I'm amazed by the amount of work (and years) that's gone into restoring these 35 mm prints. Great work fellas! As much as I love the work Harmy has done on combining elements from different sources to reconstruct the originals, it's equally wonderful to see the progress made at restoring the originals from an analogue source.  

Post
#763607
Topic
Star Wars GOUT in HD using super resolution algorithm (* unfinished project *)
Time

To quash any discussion about whether actual detail has been added by super resolution v3, here's a screenshot comparison that shows the super resolution v3 and the scaled difference between super resolution v3 and theAvisynth Spline64 Resize:

http://screenshotcomparison.com/comparison/121938

If the added detail was "micro contrast in the grain structure", the added detail should be random, since film grain is random. As we can see Obi-Wan and Luke's shape in the difference image, this is very clearly not the case. 

Also, both Luke and the background are out of focus, as the camera is focused on Obi-Wan. As there are hardly any details in out of focus film, no details should be added in those areas, as is the case in this example. 

Post
#763593
Topic
Star Wars GOUT in HD using super resolution algorithm (* unfinished project *)
Time

@ RU.08 & Laserdisc Master

Although I'm sure you're right to a certain extend, the process of combining the frames is not as straightforward as you seem to suggest. In principle the differences in the elements of the frames due to differences in depth are taken into account (as are lighting issues, scale issues, and a number of others). The reconstruction is a statistical prediction on what the actual high res frame should look like, combining information from different frames in a non-linear way. This process will by definition be imperfect and have a cost as you say. The cost may be loss of depth, but, depending on the source material, may also lead to undesired depth enhancement in some places.

There is a misconception that super resolution is all about adding micro-detail. Although this certainly is one of the aspects that makes it a powerful technique, it's main objective is to get a more accurate representation of the high resolution frame. In doing so it reveals more micro detail, but also removes many of the artifacts created while compressing and downscaling the original to a lower resolution. You can get hung up on maintaining shapes in the low resolution frame (sorry Laserdisc Master ;-)), but this assumes these shapes are good representations of the original high res material, which is not necessarily true.

In my opinion combining the information from different imperfectly compressed low res frames to reconstruct the actual shapes and depth of elements in the original high res frame (whether they be micro details or larger shapes), is to be preferred to smoothly interpolating a single imperfectly compressed low res frame, which by definition contains much less information about the actual shape and depth of objects visible in the individual low res frame.