All new VHS AI Model

Currently, all we have to work with is Iris LQ and MQ, and while LQ can be effective on mid shots and close ups, once the shot zooms out to a wide shot, faces start to get distroted bad. And using the MQ setting has little to no effect when working with VHS settings as far as upscaling 2x goes. There really needs to be something in the middle of these two.

Or better yet, a better AI model that will act more aggressively when face recognition is strong with a lot of pixels to work with, but then loosen up once the shot zooms out and it has less pixels to work with, a smarter model I guess is what I am asking for.

VHS footage is the toughest to deal with because it has the least amout to work with, which is why it needs more attention.

I would merge these if I could.

1 Like

ur basically asking for a quick fix, i agree but
thats lowkey a lazy request, idk if ure only talking about upscalling or also fixing analog bugs cuz u said it was the most time consuming restoration work

topaz is designed for upscaling good sources so if I were to enhance a vhs source with topaz I would first do a clean vhs scan, idk why im getting into this but im just showing u a workflow around if were not getting a analog vhs model, first I record the vhs using a hdmi vcr output player, then import the footage onto avisynth and start tweaking it using qtgmc turn off the additional sharpening, crop the black bars if there are any, finedehalo it, covert it to yuv 444, then downscale it to 333x240 thats the vhs resolution and then export with a lossless codec with a dar that resizes it to 320x240 under playback, use topaz at lq up to 640x480 and then use mq iris at 100%, this should be good enough for vhs footage

Oh my gosh…where to begin. First of all you are making a lot of assumptions here. This isn’t my first time at the rodeo, I am not just using some random component output vcr and capturing to some compressed file thats full of artifacting. And some of what you are suggestioning (hdmi output vcr) is NOT what you would want to do. Even if they made one that has that, VHS is an analog signal that would need to be converted to digital using the cheap capacitors within the vcr, and that is all not good for signal preservation. And as far as those HDMI signal converters and upscalers, BELIEVE ME when I tell you those are all garbage as far as giving you the best capture. Those converters use the same tech as the older HDMI TVs that provided component inputs, and unless you like the fuzzy and muddy look after it converts it, then to each there own. Been there, done that. The best way to capture VHS is first and foremost, you need a high end decent VCR deck, both JVC and Panasonic made the best you could get, but in either case you are looking at $1000 deck minimum. I use arguably one of the best you can get, a Panasonic AG-1980 which has a built in Line TBC, and I use S-Video output which is what you want to use for the cleanest signal when capturing VHS, and I run that through an external TBC, then to my capture device. I capture at a CBR that I control which is, again, what you want to do. For VHS, 5-6 Mbps is more than good enough to give you clean and near lossless footage with no artifacting from compression, at least at the native VHS resolutions that is. I then do my first pass of deinterlacing with Handbrake, but DO NOT try to double the framerate in handbrake, you will end up with multiple repeated frames, so I always set the framerate to match the same as the video. Then I bring it in to Topaz and use the Topaz deinterlace preset, only now I change it to progressive and at this point I double my framerate. It will take a while, but you will end up with super silky smooth and clean footage at the native VHS resolution, again maintain the same 5-6 Mbps and don’t do any upscaling yet, you are just deinterlacing still. The end result will be as clean of VHS footage as you can get, trust me. What is missing here is like I said in my original post, we need a better and smarter AI model that can address the native VHS resolution in the changing conditions of the camera shot. Topaz does AMAZING with closeups on VHS, and I do mean AMAZING!!! Even some mid shots look good. But it really struggles with the wide and zoomed out shots on VHS. Look no offense to you sir, but there is nothing “lazy” about my approach. I have been at this a LONG time when it comes to VHS.


ok sorry lol

(btw i dont use handbrake i use qtgmc)

yes vhs generally aint sharp at all barely holding details, they look soft

I have been working with 80s camcorder footage shot on VHS captured professionally with a TBC. With 5.02 I thought I’d look at using VEAI for deinterlacing. I’ve always gone the Hybrid route before. I’ve found that the Dione models are very noisy which is still the case. Now we have the ability to work with both Proteus and Iris and deinterlace natively with those.

My conclusion is that having all of the original data available may have advantages in terms of how VEAI reconstructs frames from the fields (converting them into 50fps progressive from 25fps interlaced PAL). I’ve noticed that the Iris model working with interlaced video might be better at cleaning artefacts and video noise – as some of the details seem sharper and more defined then working with a Hybrid or Handbrake deinterlaced source. I found that I get the best results using ‘relative to auto’ and dialling down things like ‘reduce noise’ , ‘de-halo’ and "antialiasing,’ into the minus area. Iris has a tendency to soften too much and the de-halo can be too strong.

Anyway, I think it is worth checking out the new models which have been trained on analogue video sources. It’s easy to be dismissive of the software and always worth looking at things again as hopefully there have been improvements.

I have good experience with Handbreak deinterlacing from 50i to 50p! Yadif/Bob and constant 50 fps makes smooth videos. If you play the result frame by frame, every frame is different, no double frames. Converting Video 8/Hi8 or DV footage.

For me, I am using VHS 29.97 to start with. So I found that Handbrake deinterlacing using Yadif/Default at 29.97 (not doubling yet) first then when I bring into VEAI, I set deinterlaving to Progressive and then double my framerate here to 59.94 at the same resolution output using the Proteus model, and I get the best looking, smoothest footage with little to no repeated frames. Not sure why this works out, but it really does for me.

With a VHS source, by not doubling the frame rate while deinterlacing, you are throwing away half of the original frames. That’s just the physical way VHS is made. If I remember correctly, the way the read head works, it can only read enough information to display one interlaced field at a time. Why make TVAI recreate those frames again?

I am of course only talking about things that were recorded directly onto VHS like with a camcorder. If the original source of what’s on the VHS was film, it’s going to have lots of duplicate frames. TV, I’m not sure about. You would need to check for duplicate frames after deinterlacing.

I hear you, but for what ever reason, whenever I tried to deinterlace trying to double the frame rate like you said, I would always get a FLOOD of duplicated frames and stuttering effect in motion scenes, and I know that normally means you have to change the field order, but it would still happen. The biggest problem was the repeated frames and stuttering effect that always seemed to happen.

When you let TLAI recreate the frames using its advanced software to do so, yes it takes a little time, but it does a FANTASTIC job at making it smooth and best part is, no stuttering and hardly no repeated frames, at all.

This is one of those cases I think where the software can do just as good of a job if not better when dealing with this level of resolution.

For those of you who want to start with good quality, here is a great project that came out of the videodisc decode & domesday duplicator projects.

Yes, it does involve hardware.