Analysing details in video to improve upscaling

VEAI should be able to analyse the whole video at first, and learn things so it can later do a better job at upscaling.

Like: a video shows a street sign in the background. But later in the video, the video shows the street sign zoomed in. VEAI can now learn what the street sign looks like in high detail, and can upscale it much better now then just “guessing” how it looks.

Same with faces. Learn how faces look when zoomed in - then apply what you learned the whole video. And so on.

This is not my idea btw. I saw a video on YT where a research paper was doing that - with spectacular results. I think it was on 2minutepapers - but cant find the video. But it’s proof that the method works!

That would be a great feature! :slight_smile:

1 Like

then upvote please :grinning:

3 Likes

I would love to up vote but your original post does not offer the heart symbol to click on, so I just replied. Anyway your reply offers the heart to click so I do. :blush:

there should be a button VOTE under the vote-counter…

Left of the title

Oh you are right. It s on the top left while the heart icon is on the lower right. Quite confusing. :grinning_face_with_smiling_eyes:

up-voted! Hope this would make skin less “plasticky” if VEAI can grab details from other scenes!

I think you need to go into the professional products section for something like this. There are other products out there, but at a premium.

meh it’s a great idea but the reason it’s so plasticky is too much noise removal and compression reduction. I do not have that problem at all using Proteus and it becomes very apparent with high levels of noise removal and compression reduction. Don’t use the presets and you’ll be fine.

1 Like

When I get alien faces - I take the original video and create a mix of both the old and new video. When combining both images at the right percentage, I get fuzzy detailed faces but also some structure to the details. It provides a much better and believable video.

Adding random grain or noise seems to not create a better image to this particular situation.

1 Like

Thanks r0llinlacs & michaelpfost, I will experiment with what you are suggesting.