Hi all, I’ve been playing around with trying to upscale an 80’s sitcom from some pretty mediocre 480i DVDs. So far I find that I get the best results by deinterlacing in another program and then choosing Proteus. Has anyone found some decent starting parameters for old multicam sitcoms that were filmed on videotape? Typically we’re dealing with a mix of closeups on faces, couch-size shots (side-by-side conversations of 2-3 people), and room-size shots, all with staged lighting – plus the occasional establishing shot, which no one cares too much about. Proteus on auto seems to work beautifully for closeups and pretty nicely for the couch shots, but the room shots tend to be very rough on faces and small details. (Of course, I’m sure it would be best to upscale every shot separately, but I’m not that dedicated.)
My input quality is poor enough that upscaling by more than 2x seems to do more harm than good, so I’m only aiming to achieve about 1080p at best. Even then I may need to add moderate grain to obscure the cloudiness or bad detail in the wider shots. Anyway, thanks in advance for any hard-won experience you may have to share!
If the other tool you are using to deinterlace also has the option to denoise, have it do the denoising too. Don’t turn it up so much that it removes all the little details. Try to reduce noise as much as you can before they all get blurred out.
Pick a scene that is most problematic.
I cannot get Proteus Auto nor Relative to Auto to not create nasty looking fake details.
Set it to Proteus Manual.
Start with Revert Compression at 50.
Leave everything else at 0. See if that looks good.
If not, try moving it up or down by around 10, then narrow it down.
Once it looks acceptable, slowly try moving the Anti-alias/Deblur toward deblur. 10 points is a lot for this slider. Do 1 or 2 at a time. Once it starts creating little patterns on things like dirt, leaves, rocks, concrete, small patterned cloths or rugs, you’ve pushed it too far. Back it off 1 to 3 points.
You can try adding some points to Recover Details, if you add too much, it tends to make dark spots darker. Like far away eye sockets. If you add any points to Sharpen, it starts to do the same thing that the Deblur slider was doing. You can try to counter that with some points in Dehalo or back off deblur. More points in Dehalo blurs the image.
My main goal is to make sure far away faces don’t look unnatural or creepy. Then make sure those fake detail patterns don’t get added.
For a show from 1966, the settings I settled on are this:
Revert compression 50
Recover Details 20
Reduce Noise 0
For a more modern show from around 2005, they look like this:
Revert compression 50
Recover Details 5
Reduce Noise 0
Well in that case, I can share what I’ve been using in Hybrid. In the QTGMC settings for deinterlacing you can set the denoise settings, or in the denoise tab. I have ‘final temporal smoothing’ set to 3 and EZDinoise set to 2.00 with the preset on Very Slow. I have the denoiser set to KNLMeansCL. (This is for the 1966 show.)
There’s a lot of settings in there, and I have not taken the time to research if I’ve picked the best ones, but they keep more detail and remove more noise than the VaugeDenoise filter in ffmpeg.
temporal smoothing at 3 is high… whatever you apply as parameters, you lose detail in the movements, the best I have found is to select “keepgrain” at around 3 to increase the detail, leave the temporal at zero, dehalo with hybrid also and the mldegrain filter, strenght: 1, treshold: 0.80. then the “filmgrain” (glsl) filter: 0.01 and upsaler from 480p to 720p using lanczos (it has better quality than bilinear, bicubic etc). with that, my video sd come out clean, even if there’s still a little noise, it’s much cleaner with the artemis hight and medium models (it keeps all the details in addition to blurring the image a little)
My goal is to clean it up so that Proteus can do it’s work. I admit I have not tried all the options. Settings like ‘keepgrain’ sound like the opposite of what I’m trying to do, so I didn’t try them. I compared these results with the vaguedenoiser filter from ffmpeg, and the end result after Proteus was clearer.
Artemis models always create fake detail patterns… but maybe they wouldn’t if I added the “upsaler from 480p to 720p” like you suggest.
Are you upscaling to 4k or 1080?
480 to 1080 processes faster than 720 to 1080 because it uses different versions of the AI models for those resolutions. From movies I have processed, I know for sure that Artemis models produce different results depending on what resolution version gets used.
Here’s an example: Left is VEAI running the ‘2x’ version of Artemis High, and right is TVAI that defaults to the ‘4x’ version.
I switch to maximum HD. The more we go up in resolution afterwards, the more we lose details as in your example. After all, it depends on the filters used with Hybrid. It’s also good to pass Artemis aliasing before another model.
Thanks to both of you…I’m trying both methods on my input.
@ssbroly : When you say dehalo with Hybrid, are you just checking the DeHalo_alpha filter and leaving everything default? And for MLDegrain threshold, it looks like it can range from 1-1024, so is your “0.80” = ~819 or 80?
Finally, pardon the novice question but I can’t figure out which encoding mode in Hybrid to use for maximum quality. Without upscaling, I was just specifying the bitrate of the original file. If I’m upscaling, then I imagine I need to multiply that by 2.25, but I’d guess it’s simpler to just choose the ideal factor/quantizer? (I tried a factor of 1, but that definitely overshot the ideal bitrate and wasted file size.)
For the MLDEGRAIN filter, 0.80 is apparently enough, after you can always try the increasing value. For Dehalo filters I use Blind Dehalo 3 and Fine Dehalo. For the output, I do not know much but I found that prores were not bad, I still use prores xq 4444 but I read in the post after that carlito talks about ffv1 and avi.
Awe but I just got it in DVD! Hehe. Mostly I only get Blu-Ray versions if the DVD version I already have is lacking in some way. Like widescreen stuffed into full screen. That’s the worst, and TVAI doesn’t help much.
VEAI 2.6.4 doesn’t have a CLI option. The TVAI 3.1 documentation is here.
When I want to figure out how to use a model, I do a short run in the GUI then look at the command and change it up how I need.
I never used VEAI 2.3… the documentation should be somewhere still though.