Thief: The Dark Project intro briefing video, upscaled to 1080p via machine learning

[Goldwell]

I have seen quite a few older games have their cinematics and textures upscaled to more modern resolutions using machine learning powered applications. I decided to give one of the Thief videos a shot and this was the result. Took just under 2 hours to process.

 

https://www.youtube.com/watch?v=Zs4CYxB0PGI

[chakkman]

"You have talent, lad". Always nice again.

[grayman]

Very nice!

[Goldwell]

And hey just for fun I also put the Thief successful mission loading screen through it too.

 

https://www.youtube.com/watch?v=R-QeMjSxtLI&feature=youtu.be

[jaxa]

Good work. I think you've found the next 40 uploads for your channel.

 

Now if we can only use machine learning to enhance the audio (more prominent hiss/artifacts in T1 briefings than T2 IIRC).

Edited June 1, 2019 by jaxa

[Goldwell]

Good work. I think you've found the next 40 uploads for your channel.

 

LOL I think that would piss off a lot of my subscribers as I usually only post 2-3 times a year if that.

 

Also it takes a long time to produce just one video. But I have done one more and I think i'm done for the time being.

 

This one is a little more localized though! The TDM logo video upscaled to 1080p (no audio)

 

https://www.youtube.com/watch?v=KcNH-rg8SVE&feature=youtu.be

 

 

Now if we can only use machine learning to enhance the audio (more prominent hiss/artifacts in T1 briefings than T2 IIRC).

 

I just played around with that for a little bit.. and it turns out it is possible. I don't want to reupload the briefing video but this is the audio track with the hiss removed: https://vocaroo.com/i/s0G6DHFPmWbZ

[zergrush]

Marvellous work

[OrbWeaver]

I suppose machine learning is the next big thing in image scaling and video compression. Codecs of the future will not simply analyze macroblocks and decompose them into frequency components, but will analyze textures, shapes and lines and reconstruct these at the desired resolution.

 

After all, if you're looking at an image of gravel, you don't actually care if every stone is exactly the same as the source material. If the codec can reconstruct a similar-looking gravel texture this will look perfectly fine for the viewer.

[Oktokolo]

Looks like the artificial neural network processed the separate frames of the video. Whenever ther is movement, i see some edges "wobble" a lot.

 

Is there more information about the layout of the used ANN and how it got trained?

[Goldwell]

Looks like the artificial neural network processed the separate frames of the video. Whenever ther is movement, i see some edges "wobble" a lot.

 

Is there more information about the layout of the used ANN and how it got trained?

To be honest I dont fully understand the coding and process behind it. I spent a few hours google searching and found a few applications and played around with it.

 

Your assessment of the frame by frame is exactly right too. I was watching the temporary folders it created and thats what it did. Dump 1400 odd or so frames into a folder and then one by one upscale them. In the end it packaged them all back together.

[ERH+]

Now the frame rate sticks out. Is there a similar process for adding additional frames in between existing ones?