Iris tries too hard

I’m using TVAI to upscale 4:3 DVCAM to 16:9 HD. Iris does a great job, but with one flaw. All of my footage features airliners and they have titles on their fuselages and registration numbers. If the aircraft is sufficiently far distant that the lettering concerned is not clear enough, Iris makes it up. That ‘scribble’ is then visible in the upscaled video.

I’ve tried 0% and 100% recover detail, but with no success in curing this habit. I should add that I’m not a video expert, being self-taught. Does anyone have a solution to Iris doing this? Or, what’s the view on the best alternative to Iris for this kind of process?

Are you needing to deinterlace the video? Can try deinterlacing with Proteus or Dione models and then upscale with a different model like Gaia, Artemis or Theia in a second pass.

Hi Kyle

Many thanks for your reply. Yes, I do need to de-interlace. When converting from 4:3 to 16:9, I upscale to 1920 x 1440 as that then allows me to centre the subject within a 1920 x 1080 edit. My feeling is that Proteus has the edge over Dione. I’ll use that just to de-interlace and then upscale to HD @ 1920 x 1440 with one of the others you mention. Does that make sense, or is there a better approach?

Best regards


Try that and let me know how the outcome looks and then we can make some changes based on what you do not like about the preview or export. Start with Gaia and then we can look at some of the other options. There are some good tips in this forum post by other users for working with older video formats for upscaling and getting improved results. Not all of them will apply but they can be good starting points.

Iris and Proteus (even in auto mode) do antialias, sharpen and disolve large pixel block. It’s a more invasiv method than Gaia model. This brings you some amazing results but I also noticed this doesn’t work if there aren’t enough pixels for objects you have on wide view. It also sharpens details that are far away that should not be sharpened then looks bad. Would be ideal scene based different models are used, but I think that will be difficult to implement.