I am working on a film restoration that needs some fixes done during the over-picture end credit crawl. Just blurs and some marker removals, nothing crazy, already knocked them all out in textless but I am struggling to find a not-painful way to pull a matte for the (moving) credits on top in the texted scans.
Texted and textless are two separate film elements that have unique wear and deformations enough to where I can’t get away with a difference matte. On some of the darker shots I can get close with a luma key but not the shots with lighter backgrounds, and it’s really only a half solution even then since I lose the drop shadow doing it that way.
My next plan of attack is to try to find a single clean-ish frame of each credit, 2d histo each until I have something usable, then track that matte to the credit in the texted scan.
I’d, uhhhh, really love to avoid doing that for every credit to say the least. Any ideas?
Wow, that sounds, um…juicy, to say the least. I don’t see how you get through this without having at least an element of a retyped credit roll for the problem areas. I mean, you’re a magician, not a wizard.
This might be a bit convoluted and I have no way to test it but here goes anyway. For both the textless and the texted: Blur, Negate, comp the negate with the original at a 50% blend. This will give you a high pass for both texted and textless. Do a difference matte between the 2 highpass versions. With some tweeking of highs and lows you might get a usable matte of just the text. Since your situation is so unique, I have no way to test this. It’s just a theory.
You might want to try it without the negates as well.
You might also consider 2d transform-> stabilize and track the y (and x, if it’s wavy enough) with auto scale turned on. Do this for as many chunks as you’re comfortable with and compound them to get clean mattes of the type in big strips that you can then re apply to a clean BG.
This sounds like a great use case for AI. Not that I really know anything about how to do it but I would guess that Nuke Copycat could be trained to identify and matte the text. That’s probably not very helpful but just a thought.
Definitely go back to the neg assembly, scan it and redo the titles if that’s an option. The quality will be better than the print/inter-negative you’re working from now and you won’t have the insane headache of doing what you’re purposing. The end product will be far superior.
I’ve had to do similar things…paint out legals/supers and update copyrights where the generic didn’t exist. What you’ve been tasked to do is…well, kinda wreckless and setting you up to fail big time. Right?
