Film Restoration Experiment: Machine Learning in Action

In this second machine learning test, I aim to tackle a more intriguing challenge. As evident, this 16mm film suffers from advanced hydrolysis and vinegar syndrome, rendering the initial restoration efforts inadequate. While concluding this subject, a telecine transfer of the film, recorded around the time of the original movie's release, became accessible. This telecine has its own set of issues, notably a 29.97 frame rate. Typically, one would perform a reverse telecine to restore the original 24 frames, but this copy lacks the necessary flags for such a conversion, leaving us to deinterlace and manage frame interpolation.
The video also reveals a significant crop in the telecine, resulting in a loss of approximately 20 to 25% of the image.
What can be done with these two elements? I'm considering a method to transfer the telecine's cleanliness to the 16mm's resolution and clarity without the damage, and to reclaim some of the cropped edges.
Here, Nuke's CopyCat comes into play again; we utilize the cleanest frames of the 16mm (or portions thereof) to generate an appropriate training set for the telecine. After training, we blend it with the original 16mm to regain some lost detail.
As the video demonstrates, shots with minimal movement yield the best outcomes, specifically shot 1 and, to a lesser extent, shot 2. Shot 3 exhibits various artifacts on the lower frame due to the telecine crop, and shots 4 and 5, lacking sufficient "clean" frames for training, resulted in less accurate restoration.
This endeavor will likely remain experimental, as some shots fail to maintain consistent quality. Nevertheless, this approach suggests the potential of utilizing various media types to aid in restoring a project to its original form.

[@lemonlime8949]

It’s really cool how old film could always look beautiful but just didn’t have the tech to, shows how far ahead and behind they were. But seeing old music videos be remastered and looking incredible, yet different to digitally recorded videos, is a really cool thing.

[@fabiocolor]

Movies have way more flexibility and visual range than we usually think. Even old prints can look stunning if taken care of.

[@hesselfridyanto]

Nice!

[@Notelu]

Very interesting, but im assuming it only works on clips that have an alternate cleaner source? Also it seems like the edges become very messy. Still a really cool proof of concept though!

[@fabiocolor]

Certainly, additional training time and a more comprehensive set of frames are necessary to achieve production readiness.

[@silvesterstaloney8838]

This gonna blow soon.