Iris 1 and 2 are such wonderful methods to smooth out and detail footage, but the face recognition can ruin it at times… would it be possible to have a slider from 0-100 (or a on/off switch) to determine how much, if at all, face recognition is applied?
I made a similar request a few months back. I think the main issue I’ve discovered is that if the material is noisy it will confuse the upscale. I’ve been seeing better results by first denoising with Neat Video, keeping the video interlaced. You can then deinterlace using Hybrid QTGMC and converting the fields to frames (for me 25 to 50fps). With a cleaner source without temporal noise VEAI is better able to distinguish the details in the image. I’ve also read that converting to square pixels within VEAI is better. You can of course work with Iris to deinterlace if you don’t have Hybrid.
The nostrils in the image you supplied have become eyes, so there are still some strange quirks –but I do think overall, Iris LQ is much better in the current version and able to do things the other models can’t achieve. You might consider doing two passes, one with Iris and another with Proteus, and then mixing them together in your editing software. Hopefully the Iris model will improve. It only came out earlier this year so let’s hope Topaz can refine it further.
being noisy is the fast track to getting “monster faces”, but even with good footage, iris will create havoc when the “face” it recognised is too small, has too little information.
Lol, the nostrils. In my opinion Iris is the best model, but yes, it goes too far on faces, especially the eyes. A slider for the facial enhancement, and a sub slider for specifically the eyes would be perfect.
I would not be surprised if this request comes off frustrating to several people at Topaz—considering Iris is built off of Proteus with face enhancement added. Why don’t you just use Proteus? Well, the settings don’t line up. You don’t get the same results if you put in what works well in Iris, into Proteus. It’s probably more complex than that.
Anyway, a slider would be the best way to go about this.
i´m not familiar with the history of these models, but even apart from faces, the results of proteus and iris are substantially different. anyway, it would be unreasonable to expect that a handful of models can cover anything and everything out there. potential video footage is too diverse for that, so i think the current evolution of these models is just 1 step of the journey. no need to be frustrated at anything, because in the end, this whole process is going forward very visibly and in a promising way.
As requested quite a few times now, it would mostly be sufficient if Iris just left small or very LQ faces alone.
Similar as it is with TPAI where face recovery automatically unselects some faces it thinks are not appropriate.
Perhaps, if it’s possible, add a slider that would control the amount of “Iris enhancement” on faces only, in isolation of other parameters, as outlined in the first post.
TOTALLY agree. The Iris model is excellent, but if the video includes many faces, such as a shot of an audience, it will try to put faces on all of them resulting in unacceptable and ghoulish images. I would love to have a slider to make Iris enhance faces ONLY when they are big enough that it can do a respectable job.
Did you try Iris LQ 2x upscale with high Recover original details (>80)?
This diminishes those zombie faces compared to IrisMQ and even to Proteus quite a bit.
Worth a try. Thanks.
I gave it a try. It seems the trick is to use Iris LQ not Medium. Here are two pix showing the old way I was doing it and the new way per your suggestion. On closeups there is much less improvement, but it is still some improvement. Thanks for the tip!
Not only the faces are MUCH more natural, also the clothing. IrisMQ has this ugly artificial oil-paint look.
That’s why I like Iris LQ so much. It really does clean up bad sources very well (and also does a quite decent job on deinterlacing at least for my sources) and with the right settings increases sharpness and details - but does only few “bad” things at that.
I do prefer a little less details over more but mostly “fake” details.
My preferred settings for lq SD material:
IrisLQ 2x upscale - this is also important as the 4x model of IrisLQ has some of the misbehaving of the MQ model. So if you really want 4x upscale (often not really worthwhile) I’d suggest doing two consecutive 2x passes.
for these scenes, proteus is much more suited.
Nope. I used to use Proteus for that material but IrisLQ simply is much better in many aspects so I’m still reencoding vids I already upscaled with Proteus before with Iris LQ.
Have to thoroughly test ProteusV4 though - when I have time.
i said “scenes” for a reason. i´ve turned to making several encodes of footage material and use them on a scene-by-scene basis in my video editing software (which has scene detection functionality), where i make one top layer (e.g. iris) transparent whenever its results are not desirable, and let the layer beneath it (e.g. proteus-encoded) shine through, in whatever transparency percentage it makes sense to do so.
proteus has saved me from many ghoulish faces in crowds, because sadly, no matter how far you turn up the detail slider, the face recognition in iris will go haywire -eventually- and if you have nothing to fall back on, the “believability” factor in the final footage can be destroyed easily.
(having said that, right now i am working on footage where NONE of the models properly recognise even huge human faces like (like 25% of the screen size) and can make basically nothing out of it, even with some intense denoise/deblock prework.)
I had five one-hour original 8 mm → VHS → DVD. So, that’s two generations removed from the original. I created a preset using the Iris filter LQ set to Auto back on June 9, 23, with frame stabilization, auto-crop and added grain to the resultant video. I had come up with these settings after extensive previews. The results were acceptable, and I was happy to edit the resultant video in Premiere Pro. Recently, I ran previews of one of the videos using that same preset. I was disappointed that the same preset now had unacceptable results. I just assumed running the same preset that I created in June would have the same results that I ran in late December. I can’t remember whether there was even a LQ or MQ setting for Iris back then, but my preset shows LQ. I don’t remember the VAI version I was running at that time.
Looking forward to the new Proteus 4 that I’ve heard so much about.
