Questions about Video AI

Hi,
I like to take a lot of old videos and convert them to VR. Don’t worry about that, thats not the main gist of my questions. I can handle that aspect on my own…however, I might ask something semi-releated, later.

  1. Can you input video that is better quality from a different film to act as a reference when you want it to increase the resolution/sharpness/quality of people in the video you want to upscale?

  2. When going from low res to high res, does it make sense to go to a mid level resolution, and then go to a high resolution? Or is it fine to just go from low res to highest res?

  3. When you go to a resolution, is there an advantage to running the video output/export through a second time to get it even better looking? Or is a single pass the best you can get?

  4. Can you choose to output to non-standard resolutions? A lot of times I like to make my “1080p” videos 1920X3414 .

  5. If you have a clip that has a really strange frame rate. Can it fill in the difference to smooth the frame rate out to a normal 30 fps or 60 fps?

no, no, yes, yes, maybe. :slight_smile:

  1. No

  2. This depends. On some videos you’ll get better results by doing twice a 2x upscale, for others going directly with 4x upscale can lead to a better result (especially since the 2x and 4x models for e.g. Iris are noticeably different in what they do).

  3. Mostly yes. You can improve the output sometimes by running TVAI more than one time; e.g. for VERY bad quality videos I found that it sometimes helps first cleaning them up with Artemis and then do a second round with Iris to reduce unwanted artifacts as those crazy “Zombie-faces/eyes”. Also I had quite decent results with doing Iris first followed by Nyx 1.
    In other cases the output can look too “processed” / smoothed when doing two AI runs (even more so than with a single one). So you have to try yourself depending on the respective source.

  4. yes.

  5. No real experiences here. But I guess this would likely be a “hit or miss” thing.

Could you elaborate on the iris x2 vs. x4 differences?
one produce better quality then the other?
if x4 is the superior one, what would be the best way to x2 upscale using Iris?

I wouldn’t call one or the other “better”, just different.
Iris x4 can bring more details (but also more artificial plastic looks) than the 2x upscale model, at least with Iris V1.

Since I do like the “natural” appearance of IrisV1 together with it’s nice cleanup of bad quality material, e.g. VHS and old interlaced sources I mostly prefer the 2x upscale look.
As always, YMMV

1 Like

Any findings of yours on the same nature as regards to Iris v2?

Did not really test this yet as I’m mostly using Iris V1. And even there the differences seem to have diminished with the llast model updates.

Maybe I can provide some samples later on.

1 Like

Here are examples that show that Iris 2x and 4x upscale models do different things other than simply having double/half resolution:

IRIS V1:
2x upscale:

4x upscale:


has artifacts in the “d” but better/noticeably different details in the plants.

IRIS V2:
2x upscale:


strange artifacts below the “dm”

4x upscale:


No artifacts under “dm”. Better / very different depiction of the plants. But the letters of the “dm” are somewhat distorted (maybe only seen in a direct A/B comparison).

3 Likes

Now I see what you mean. each one excels at one area but inferior at another

1 Like