Scanning is nearly done, the print is very faded and beaten up, here are some rough samples from the beginning, middle and end of each reel.
Agreed. Ben’s cloak doesn’t look right though.
I am so keen to try this out, can hardly wait.
Yes, Ben’s cloak tends to come out too green. There’s actually now an option in the algorithm, that should improve the results for such inherently unbalanced frames. It’s a bit technical , but it involves increasing the bin size of the color intensities that you assume should be balanced. After increasing the bin size Ben’s cloak looks better:
…and after a gamma correction:
It’s interesting to note that Obi-Wan’s cloak does appear much less reddish brown in production photos, than most home video releases would have us belief:
It also shows why production photos are close to useless, his cloak is completely different in those two shots, and every photo will look different, even when the same photo is in two different publications.
Now if someone had just held up a colour chart in a shot…
Agreed. Ben’s cloak doesn’t look right though.
I am so keen to try this out, can hardly wait.
I really could not understand why I’ve been getting such inconsistent results. I calibrated the latest version of the algorithm on this frame:
As far as I can tell, it looks just fine. When I apply the correction to other frames, I’ve been getting color shifts, that were mostly yellow/green, but also blue. However, when I examined the raw scan I noticed the relatively large differences in hue, brightness, and saturation between these two frames, that are only a few shots apart:
These inconsistencies are most likely caused by the scanning process, as poita’s scan looks much cleaner, an much more consistent.
The print appears to have a different source spliced in at this point, it is completely different colour wise suddenly in my scans as well. The Spanish print is a dog’s breakfast really.
I like the idea of being able to use part of a frame as the correction source, Artoo is especially useful for this.
The colour alignment will make all scratches and dirt become ‘tri-colour’ as they are not ‘misaligned’ to begin with, so it does make quite a difference for a grindhouse style viewing.
It is helpful in deciding what is print dirt/damage and what is baked into the negative though, I’ve been thinking about whether this could be useful algorithmically for restoration.
A true grindhouse would have the colour misalignment, as that is what is on the print 😃
What resolution, filesize would people want a grindhouse to be, and would you prefer the intermission cards left in or removed? Colour aligned or not? (It won’t realign perfectly unfortunately, but can be improved a lot), missing frames left as is?
Cast your votes here.
And yes, more funds for more HDDs would be great, as I am currently working without a net, and it has already cost me time when the RAID0 working space decided to ‘unraid’ a few days ago, costing me a couple of days tedious work to have to redo.
Have you looked into workstation cards rather than gaming cards? I don’t know if you’d do a whole lot better with price v. performance, but just a thought.
Workstation cards are slower and more than twice the money. If the software used FP64 then workstation cards would be considerably better, but none of the restoration suites use it, so the Titan X is currently the fastest thing available within the affordability bounds of mere humans, the 1080 will make things more affordable and bring a little extra performance (9TF vs 7TF) where 8GB of RAM is enough, and the 1070 will bring good performance at a relatively low cost, i.e. slightly slower than the Titan X but less than half its price. (6.5TF vs 7TF).
The ‘twice as fast’ that NVidia are quoting is only for VR performance where the new cards only render high detail in the centre and low detail towards the edges, not on normal compute tasks.
Star Wars Sample: https://we.tl/VEyCxemjLh
to much yellow
No, I matched it perfectly to this colour reference: https://www.youtube.com/watch?v=08kLB221rys
The 1070 is slower than the Titan X on compute unfortunately.
It will only be up for a week, so if you want to check it out, please PM me!
Thanks for all the interest and questions. Don’t think me rude, but I am out of here now, so any other questions may go unanswered for quite some time.
If anyone wants to throw a pair of GTX 1080 cards my way to speed things along, feel free to do so 😉
It will be a combination of multiple prints for all three films actually.
So is this the kind of thing where you utilize multiple prints in order to achieve a higher net resolution than any individual print alone?
Partially, but it is mainly for other reasons.
The sad reality is the popularity of these movies means that any ‘complete’ release prints are actually full of splices and are missing frames, so you need multiple prints to have a hope of getting as close to a complete movie as you can.
As per above, some frames are scratched or damaged to the point where the only feasible repair is to use frames from other prints to help restore the damage
Noise reduction, having multiple prints is useful for scenes that are heavy comps. They are way noisier to begin with because of the extra generations required, and as they age, in some cases they get the dreaded brown noise that you can see clearly in the sandcrawler and Speeder shots for example. Combining multiple prints helps the signal rise up out of the noise.
Colour. Having multiple prints means you can make a better estimate of the original colour timing.
Damage decisions. Is that a fleck of sand on Kenobi’s robe, or is it a pinpoint scratch on the film? Multiple prints help you decide what is image and what is not.
So, some scans may not end up having a single pixel in the final mix, but still may be useful, some shots will be from one scan only, others may have 2 or more prints combined on any given frame. It is a painful way to work on one hand, and is a horrible amount of data to wrangle (another 5GB of SSDs would be lovely), but it does take away the guesswork and allows you to do more restoration and less interpolation.
It will be a combination of multiple prints for all three films actually.
Is one of the Star Wars prints you’re working on a Tech IB?
Yes.
It will be a combination of multiple prints for all three films actually.
I’ll record the process for a scene when I get back to the trilogy.
I didn’t explain myself properly, I was trying to get across that the original colours on set (assuming 5500K lighting) are not necessarily what was captured on film, or intended to be captured, and that what was originally captured on the neg may have very little bearing on what was on the release prints.
Even in the Star Wars days, scenes were colour timed to help tell the story, so what the set looked like lit with 5500K lights may have little to do with what is delivered to the audience.
This video gives an example, the opening shot is relatively neutral, but what is delivered to the audience will be very different depending on the story being told by the scene. Although the scene may have been shot under neutral lighting, it may not be intended to look that way in the final product. (again, this doesn’t change how effective or useful the tool is!)
www.editorslounge.com/special-coloringwithalphadogs/colorstory.htm
This is not to understate the usefulness of DrDre’s amazing tool, it is extremely useful and works very, very well, but a few people seem confused by exactly how it works. It can’t on its own restore a scene to exactly how it looked in the cinema, especially if there are no neutral scenes in the reel. If there are neutral scenes, then it will restore them to as if they were shot under (approx) 5500K light, which then adjusts the rest of the scenes and maintain their grade, whether neutral or not. This is its great strength. You then have a start point for doing the final grade which is great.
However, even the most neutral scenes in a film won’t have been lit in such a way that they perfectly matched the Illuminant E lighting used for the calculations, so it becomes a great base for then restoring the grade from that understanding and starting point. It may sound like a quibble or a criticism, but it isn’t, it is just important to understand that restoring to the Illuminant E lighting setup is different to restoring to what was captured on the day. Knowing this lets you use the tool to get to a start point that is consistent that you can then grade the movie from, which is fantastic.
What it does do is incredible, and extremely useful, especially for films where there is partial colour fade, or in the case of Star Wars where they have gone back later and digitally regraded the film, and you want to unearth what the initial grade may have been, and for a multitude of other uses.
I can’t thank DrDre enough for continuing to develop this, it is going to be a mainstay in the toolkit for all of us.
Sorry, I just meant that I balanced the sequence by eye, out of habit/years of balancing shots. I didn’t spend more than a minute on the colour adjustment, just did it without really thinking, like changing gears whilst driving.
It might match the print, it might be miles off, I don’t have the print here to check.
It is finally up, anyone who has offered to help, please drop me a PM with your email address and I will add you to the list.
Ok, excellent.
So basically it gets us back to the colours on set, but not necessarily the colours captured on film, or to the original grade.
It does give a great neutral starting point to work from, I am keen to try it out.
Go for it, but keep in mind I have not done a colour grade yet, and this does not match the print very closely.
Excellent quality! Thanks for sharing your progress with us! And thanks for the read-along records! I had both of those as a kid and the big record for ROTJ. I didn’t remember the voices being so different from the movie.
I’m glad you enjoyed them, here is the last one I have
Mac: https://we.tl/OoWYrD4K16
Windows: https://we.tl/srDZ35ycvt
I guess back when you could only see the movies in the cinemas, you didn’t notice the different voice actors quite as much.
I accidentally uploaded the file without audio, so had to re-upload it.
I actually don’t have a list of people who offered help, the link has been sent to thxita, he will pass it on once the upload is complete, it shouldn’t be too much longer now.
However, I had to use infinit as the file was larger than 2GB with the audio, and I needed something with resuming as my internet craps out all the time, so you will need to get infinit if you want to take a look at it.
It is free, https://infinit.io/?referral=poita and if you use my referral, we both get an extra half gig of storage which would be very useful going forwards for this project, and for the Original Trilogy restorations, as I need to transfer large files to people who are helping me find glitches in my work as it goes along.
You can of course not use my referral, or not use infinit, it is just what I will be using going forward, so would mean things will appear there more quickly than elsewhere.
This will be all from me for a while, I need to get on with THX and also find some paying work, as I really need a few more SSDs to make it faster and less painful to work on, and family stuff is pretty overwhelming at the moment.
Thanks everyone for the feedbacj, I will post here again when I have something to show, but it may be a while, so don’t be concerned if I go quiet for a month or two.
To round out the trilogy, here is a shot from Jedi.
Again, there are some issues, but this gives some idea of how it will look.
Dialling up the accuracy of the motion detection brings the speeder back. Also did a hair pass and stretched the signal a little to reveal more shadow detail.
(the green glitch frame happened during upload for some reason)
As mentioned these are just a test, not something requiring feedback at this stage. I’m happy for everyone to comment and pull them apart, but these clips are not going to be used as is, they are only offered up as an example of how I work.
I’ve also been asked by a few if they can have a copy of the scripts, I probably should have used different terminology. I have written these as part of Nucoda’s development system, they are not AVISynth scripts, they are more like plugins and the deflicker on the last compile was just over 3000 lines of code, the motion estimation is around 12000 lines and the de-dust is about 4000 or so, and they have dependencies on other specific plugins.
