2020-09-21, 07:45 PM
Some may know this movie is one of my favorites of all time.
After years of searching and waiting, I managed to get hold of a 35mm trailer scan of a Felidae trailer.
I then took the same trailer from the DVD (frame-perfect match), cropped the 35mm scan to fit the DVD trailer and set up these two sources for AI training. This is a trick I learned from someone else some time ago, thanks to that person!
In the end, I want to use the resulting AI model to upscale the entire movie. Now, the trailer is actually letterboxed where the DVD itself is anamorphic, but I figured I could give it a try anyway, and so far the results seem promising.
This is what I got after something around ~24 hours of training (good results can take weeks though!):
![[Image: 50lKynCs_t.png]](https://thumbs2.imgbox.com/07/50/50lKynCs_t.png)
DVD on the left, AI upscale on the right. The low contrast of the output has to do with how I set up the training, pay no attention to that.
Here's a direct comparison between Bicubic upscaling and the AI upscaling with slightly adjusted contrast:
http://www.framecompare.com/image-compar...n/7GGWWNNX
Already I think it's very promising. You can see it seems to have the most issues with emulating the 35mm grain, the imitated grain comes out way too uniform and synthetic looking. However this was actually much worse at around 5 hours of training and it's been improving ever since, so I'm hopeful that it will eventually improve to the point of being fully acceptable.
If that doesn't happen, maybe because the DVD just doesn't have enough information to properly reconstruct any reasonable semblance of grain, I will try adjusting the training set by temporally degraining the 35mm scan for the upscaling. That should almost perfectly preserve the image content (as opposed to spatial degraining) while getting rid of the grain, meaning the AI will no longer have to emulate the grain and could just deliver a clean picture. I could then try and add the grain with some other method later.
Anyway, I'll keep this training going for at least a few more days and see how it improves.
After years of searching and waiting, I managed to get hold of a 35mm trailer scan of a Felidae trailer.
I then took the same trailer from the DVD (frame-perfect match), cropped the 35mm scan to fit the DVD trailer and set up these two sources for AI training. This is a trick I learned from someone else some time ago, thanks to that person!
In the end, I want to use the resulting AI model to upscale the entire movie. Now, the trailer is actually letterboxed where the DVD itself is anamorphic, but I figured I could give it a try anyway, and so far the results seem promising.
This is what I got after something around ~24 hours of training (good results can take weeks though!):
DVD on the left, AI upscale on the right. The low contrast of the output has to do with how I set up the training, pay no attention to that.
Here's a direct comparison between Bicubic upscaling and the AI upscaling with slightly adjusted contrast:
http://www.framecompare.com/image-compar...n/7GGWWNNX
Already I think it's very promising. You can see it seems to have the most issues with emulating the 35mm grain, the imitated grain comes out way too uniform and synthetic looking. However this was actually much worse at around 5 hours of training and it's been improving ever since, so I'm hopeful that it will eventually improve to the point of being fully acceptable.
If that doesn't happen, maybe because the DVD just doesn't have enough information to properly reconstruct any reasonable semblance of grain, I will try adjusting the training set by temporally degraining the 35mm scan for the upscaling. That should almost perfectly preserve the image content (as opposed to spatial degraining) while getting rid of the grain, meaning the AI will no longer have to emulate the grain and could just deliver a clean picture. I could then try and add the grain with some other method later.
Anyway, I'll keep this training going for at least a few more days and see how it improves.
I love this movie. It's gotta look good.
![[Image: VHvXaNSu_t.png]](https://thumbs2.imgbox.com/84/71/VHvXaNSu_t.png)
