I experimented with double passes and single passes. I have come to the conclusion that single pass is always better in terms of efficiency and quality. These AI models (I may be wrong) were never meant to be ran multiple times. You pick one that's appropriate, you run, and you're done.
With regards to old footage restoration, you are correct. AI can only do so much to restore what could be there, not what wasn't there. If a video is a good quality 480p / 720p, and I mean tack sharp good in those native resolution, you can do quite a bit with them. It's even possible to upscale to noisy 4k. But if they are not even sharp in 480p / 720p, don't bother.
Upscaling is like plastic surgery. A bit is good because no one can discern what's natural and otherwise. But too much and it looks plastic. Therefore, I stick to Gaia for all my needs. Iris is a close second but it is a bit too aggressive for my liking. And other models are simply inferior for realism imo.
Here is my work flow to 4k:
- If a video is sharp in its native resolution, I use Gaia HQ. It is more aggressive than Gaia CG, and I use it because I can really get away with it.
- If a video is not sharp/ is extremely grainy in its native resolution, I use Gaia CG. Gaia CG is just incredible. You can probably blow up an image from 720p to 8k no problem, and it'd be perfectly realistic to the eyes. It is the least aggressive AI model, and my go to.
- I'd consider most 720p / 1080p source OF videos sharp, and can pass with Gaia HQ. I hope that serves as a useful reference to you in determining what I mean by "sharp".
For frame interpolation:
- Apollo (regular not fast) is always the first choice. Chrono is good, but Apollo is far more superior when it comes to non-linear/sporadic movements. It introduces less blur, less warp after render, and a smoother, more natural overall image. You can 4x/5x the frames no problem. Check out my recent uploads in Faye Reagan thread for sample.
