Film restoration, though it doesn't technically fall under "restoration".
I run a site [0] that has a lot of old and indie films. Some of these are really great, but terrible quality, because they are very old with very few copies in existence.
Proper restoration is a crazy time consuming process, and often requires you have access to some of the analog sources, and the machines to read them.
But... What if instead of a full restoration, you focus instead on just making it watchable? Allow some quality loss to trick the human eye into seeing what is important?
The best result so far is on The Great Train Robbery [1], and I have a brief breakdown of the process over here [2]. (I also put together a soundtrack for it, using a MIDI synth).
However... Currently the process is _slow_.
Restoring a full-length feature film, like I'm trying with the Blancheville Monster, will take months, despite it running in parallel to cover as many frames as the CPU/RAM combo can handle. (The Great Train Robbery took about 2 weeks).
Anything I can do to speed up the process is helpful. Unfortunately, the most time-consuming step (up to 20 seconds a frame, but usually around 1-2sec), is also the most important and I don't think I can optimise it. (The simplify step that uses k-means clustering).
The process is also not perfect yet. There is still too much quality loss. It's watchable, easily, but could still be better.
I run a site [0] that has a lot of old and indie films. Some of these are really great, but terrible quality, because they are very old with very few copies in existence.
Proper restoration is a crazy time consuming process, and often requires you have access to some of the analog sources, and the machines to read them.
But... What if instead of a full restoration, you focus instead on just making it watchable? Allow some quality loss to trick the human eye into seeing what is important?
The best result so far is on The Great Train Robbery [1], and I have a brief breakdown of the process over here [2]. (I also put together a soundtrack for it, using a MIDI synth).
However... Currently the process is _slow_.
Restoring a full-length feature film, like I'm trying with the Blancheville Monster, will take months, despite it running in parallel to cover as many frames as the CPU/RAM combo can handle. (The Great Train Robbery took about 2 weeks).
Anything I can do to speed up the process is helpful. Unfortunately, the most time-consuming step (up to 20 seconds a frame, but usually around 1-2sec), is also the most important and I don't think I can optimise it. (The simplify step that uses k-means clustering).
The process is also not perfect yet. There is still too much quality loss. It's watchable, easily, but could still be better.
[0] https://sixteenmm.org
[1] https://sixteenmm.org/s/thegreattrainrobbery_2020
[2] https://sixteenmm.org/blog/20200318-Filmscope%20Progress