Topaz Video AI - Starlight Mini Test

[tsw01]

The result is amazing.
https://www.youtube.com/watch?v=VPXtPlQGu6s

[JoeS99]

I did the "try it free" online (limited to a short clip) with some Wolverine scans of grainy 8mm film and I'm amazed at what it can do.
Hopefully there will be downloadable versions that are affordable.

[JoeS99]

Here's my test video.
Original footage first, then the Starlight enhanced.

[aramkolt]

Mini I think is the downloadable version that uses your computer's hardware, but takes much longer. Still kinda pricy, but my guess is costs will come down as time goes on.

[Alwyn]

Both = impressive.

@tsw01, what encode times are you getting?

[Selur]

Temporal coherence could be better, but it probably only filters one frame at a time without taking the others into account atm. .
=> diffusion-based upscaling models sure produce interesting stuff.

[JoeS99]

Is there some other app that even comes close with poor quality sources?

[Selur]

With conventional filtering: no change
Denoising itself can be done with SpotLess, MCTemporalDegrain addding something like BasicVSR++ can help, but the face reconstruction, would need something better that CodeFormer.
Other diffusion-based upscaling like StableSR, SUPIR and similar, should produce similar outputs, but might be slower. (assuming Topaz tuned their stuff)

Cu Selur

[JoeS99]

Well, I'm just starting to look into this for possible cleanup of old 8mm films.
Lots to learn, I'm sure.

[Selur]

Diffusion-based stuff isn't really cleaning but more like using chatgpt or similar, it's more guessing what could this look like, since it must look similar to stuff I know.
Same with stuff like CodeFormer,...
So be aware, that you might end up with unexpected results.
You should always check the result and not blindly apply machine learning stuff.
And be aware that diffusion-based upscaling and machine learning based stuff in general will usually be rather slow and/or require monster hardware.