Bug - Oversharpening over-time on still scenes with upscaling for newer ai models

Like the title says. If you use any of the newer ai models on an upscaled video, during scenes where an element stays on the screen/still for a few seconds, like title text, a logo, or a background, over-time the still object will become extremely oversharpened and start artifacting.
It seems to be more prominent with higher upscales, so upscaling from 480p to 4k makes it much more visible than 480p to 1080p, but its still there.
If you want to test, load up a video with a still background or something like a logo in it for a few seconds, upscale it to 4k and use the Artemis Alias & Moire v10 model. You’ll see it happen. It does not seem to happen with older versions of the models, only the newer ones. Artemis Alias &Moire v9 does not have this issue, only v10.

Actually it’s oversharpening all the time, not just on stills. For VEAI v2.3, this affects many models released after v1.8.1. I was preparing a post about it, but haven’t had the time to finish (it’s long).
Briefly, sometime after v1.8.1 new models were released with json options about what kind of artifacts a model might have to deal with (my observation). For those who still have VEAI after 1.8.1 and up to 2.3, look at the jsons with Notepad++ (or equiv.). There’s a line near the top like:

  "vidArtifactType": ["Compression", "Noise", "Blurry", "None"]

I believe this is telling VEAI that a video has these issues.
“Compression” means to remove compression macroblocks. Leave this alone (it’s needed).
“Noise” means to apply extra noise reduction.
“Blurry” means to apply extra sharpening.
“None” is a placeholder and has no effect on the video (it might be needed for the code to not crash).

I opened each json and looked for this line (some don’t have all these parameters).

I changed all the “Noise” to “None” (most of my vids are virtually noise free). Leave it alone if you have a noise issue.

I changed all the “Blurry” to “None”
WHOA! Massive difference in video output. No longer any blocky oversharp features. Video now looks much like Artemis v9 (all flavors). If you want more sharpening, use CAS or vsMSharpen in Avisynth in your post-VEAI script (those are typically the best for most video IMO).

What version of VEAI are you running?

I’ve been wondering if this issue has been fixed post 2.3 or still exists. I refuse to use anything past 2.3 since I don’t know if it’s been fixed. Json access is gone post 2.3, but Taylor Bishop posted that jsons are still used internally and external access would return (as you may know Taylor is no longer with Topaz so this may not happen).

So my VEAI 2.3 experience is very different from most folks. Because I have better tools IMO, my VEAI videos look more realistic than most (not bragging) and those with real (generated) film grain from FilmConvert, you can’t tell by eye that they have been touched by VEAI.

Note I jumped from v1.8.0 to v2.3, so I don’t know anything about the other VEAI versions.

I’m on 2.6.4 and its still very much present

Well that’s disappointing. So most folks are not seeing how awesome the new models (especially Dione) really are.

For SD, Dione outperforms QTGMC (for 1080i, QTGMC still rules). I’m not kidding. You want a sharp 720p from 480i, for many videos Dione Interlaced Robust v3 or Dione Interlaced Robust Dehalo v1 can do it. Keep in mind this is with high quality interlaced VOB sources. And lots of video from the 80s and 90s don’t have enough detail to get a natural looking 720p (VOBs won’t help). Don’t try for 1080p. It won’t look natural (well maybe if you add tons of film grain via FilmConvert, no one will notice the overupscaling…).

I hinted at it before but the “Killer” progressive model for 720p and 1080p is a modded Dione Interlaced DV v2. This typically outperforms Artemis v9 on slightly blurry or overcompressed video without the blobby cartooning of Artemis Dehalo and Artemis Strong Dehalo (and no amplifying of dark creases). You will not need post sharpening (with CAS or vsMSharpen) on most videos as this modded model outputs video at maximum sharpness already.

Change in dione-dv-2.json to:
“interlacedFrames”: 0,

“vidArtifactType”: [“Compression”, “None”, “None”, “None”]

Rename it to Dione Killer Upscale (or whatever). This model is strong. Sometimes too strong and the results don’t look natural. Only use it when your usual others have failed.

My usual order of testing (for progressive) is Artemis HQ v9, Artemis MQ v9, and Dione Killer Upscale. If the video is very blurry, Artemis Dehalo or Artemis Strong Dehalo may help. I’ve given up on Proteus. Whatever Proteus does, I can do better with the models I listed plus DPID, and CAS or vsMSharpen afterwards. I have yet to see any of the newer Artemis models work better than my favorites. They don’t work at all or when they do work, it’s not without artifacts.

Does anyone know how to have this modded model coexist with the original Dione Interlaced DV v2?

I added some more details about my prior post here:
https://forum.doom9.org/showthread.php?p=1972496#post1972496