First, let’s talk about Dehalo. Your team seems to love it–it’s on everything, but it sucks.
It only over-blurs the image, grays out the blacks, and does more overall harm than good to the picture quality, which cannot be repaired once it’s done. It looks like a greasy camera lens. My professional opinion is that Dehalo should be a separate slider (like add noise), used to solve specific problems in very specific cases. It should never be part of “Auto” mode, AND should absolutely NEVER be an integral feature of the model’s baseline–which brings me to the problem.
Apparently the Iris Interlaced models (MQ and especially LQ) have a baseline-level of dehaloing BUILT IN that cannot be turned off! This has to be some kind of mistake. You can’t get rid of it even when the manual dehaloing level is set to zero. The Iris LQ model is the only one that is able to correct the line-skewing from analog signal noise on a VHS recording I am restoring–and that part looks GREAT…but the built-in dehaloing is DESTROYING the drop shadows on the CG text (see below).
This shouldn’t be happening. If de-halo is automatically included as a baseline in the model, you need to include a checkbox to allow users to turn it off in situations like this. Please? Thank you!
Repaired line skewing…but, with destroyed CG drop-shadows and over-softening.
I’m not certain that those shadows are being removed due to dehaloing. Regardless, this is why I requested an Iris HQ model. Iris is exceptional, but it’s impact is frequently too intense. If the team was able to eventually release an MQ version of Iris, they can produce an HQ version.
In your case, it may be best to run that video through a program that does a better job of handling line-skewing first before running it through the video AI.
Any suggestions as to a software that actually does line skewing?
Vertical jump is another one I deal with too (not related to this issue but a restoration issue just the same).
Yes, and my guess is the biggest obstacle to a fix is they still haven’t taught these models to recognize text. If it could do that then the shadow removal could be avoided/minimized, or the shadow could be enhanced or corrected…not sure what the problem is. Right now unclear text even gets misinterpreted as something that looks like Thai characters. Big problem.
