is upscaling worth doing for deleted scenes?
Yeah, if deleted scenes are only in SD, I would call that a last resort and consider it worth it.
But also try multiple methods. That's where there might be something to really talk about. I think masking together a mix of different upscale attempts would produce some surprising results. This is not a silver bullet one-button solution but an experiment to try.
A "standard" or general non-specific upscale or two. Maybe one using a model intended for animation and one from Topaz or something popular and considered overall decent. Watch through both and take note of which shots look better in which method and what still looks blurry. Experiment with blending the methods to see if that improves anything. These would be used as a basis or a bottom layer. Then the goal is to mask in elements of other upscale attempts and add back in lost detail
Remini, despite being an android app, when it works, can pull off some crazy levels of detail for faces. However, it works best on photos, so it doesn't necessarily have the same level of quality for video. So you would want to use it on still frames from the video, however, these can maybe become keyframes for EbSynth.
EbSynth can apply styles to video, and may be useful in applying detail from the keyframes to the scene. If you are able to get EbSynth to apply Remini's "HD" face details onto the upscale it may save time. Or it might help bridge the visual gap between frames generated by Remini. The more keyframes the better the end result will look. But it's still not done. Remini is not great and is even at times bad at upscaling stuff that
isn't faces, and some faces with not a lot of starting detail will end up horrific and will need to be blurred or removed. And EbSynth is not great at fast motion. So the output, even when ideal, won't be immediately useable.
Once you have your footage from Remini, whichever way you end up generating it (with or without EbSynth) put that on a top layer. Mask out everything except the faces you want to keep. Maybe use unsharp masks or lower the opacity on the faces that are almost useable but not quite. Use a blend of your other upscales based on what looks best to you underneath. Depending on the footage, you might even be able to substitute elements like skies or close-ups of props or hands or something with new or repurposed footage. Finally when this is all together, compare with "true HD" footage specifically to get color match as close as you can. This is the final step, once color is as close as possible, you've done all you can. It won't be perfect but it'll be a lot better looking than stretching the SD footage and calling it a day, it might even be passable.
I'm open to talking about this more and elaborating or helping to brainstorm. Let me know if it is of interest.
