On: RealFill: Image completion using diffusion models

[drcode]

When will they re-release all the old Star Trek TV shows in 1080p resolution and 16:9 aspect ratio?

[ShakataGaNai → drcode]

There are already applications like https://www.topazlabs.com/topaz-video-ai and https://tensorpix.ai/ -- So it doesn't seem unreasonable that some of these deep learning models could upscale all these old TV episodes to at least 4k.

I'd love to see a combo of this Google tech and AI upscaling do the same for Babylon 5. They had shot the actors in widescreen format, but the CGI spaceships were only rendered in 4:3 and the files have been lost.

[dragonwriter → drcode]

This requires other pictures of the environment to use to infer what should fill in the gaps, which will not exist for every shot in those series. (TOS and TNG were already rereleased in 1080p, though.) I suppose you could use outpainting to construct the rest of the scene in one frame, and use that as the reference for other frames in the same shot.

[pbhjpbhj → dragonwriter]

A lot of the shots are on the same set, so you'd want the system to use a whole series (season) as samples.