Change the workspace/temp data folder to something different.
For example C:/topaz_temps.
Then you can clean all of this files out by deleting this one folder regularily
Since video AI V5.0, The Proress export in mov container has this big issue(exported mov file canāt be imported into adobe premiere or Davinci). I try it many times and find this issue cause by sound track data. without sound track output, the mov file can inported into Premiere without any problemā¦
Unfortunately setting Audio to āNoneā doesnāt help for me. It still says āthis file has an unsupported compression typeā. Are you using a Mac with Apple Silicon?
No, I use PC windows.
3840 x 2160 is ā4K UHDTVā, 4096 x 2160 pixel is "4K DCI " from display Industrie, not film, but all is labeled under ā4kā
Ok 4:3 - 4k does not exists, but we need it and vertical resolution 1440 or 2160 is same, many displays has 2160 or 1440, so you have a native pixel resolution then.
multiplier 4x should not removed, but just one fix factor is not enough. Not everyone likes upscales to 8k, and it makes no sense upcale your 1080p movies to 8k watching on a 4k TV and my 480p uplscales ends in 3413 x 1920, I donāt like that.
There are limits, Topaz canāt invent new things that didnāt exists. 4 pixel noise even looks flat when you scale it up to 8k, so itās just smoother becaue more pixels are there but looks not better. Round about 200% upscale max gain is reached. You can clearly see that on all other models.
The reason why 1080p to 4k instead 1080p to 8k with Rhea looks worser is the 4x model is better AI trained and thatās the only reason I (must) use it.
You can also see what the film industry does, they donāt scale 480 pixels up to 4k, they do 1080 then, or when they do you have a worse 4k compared on the level of other 4k content. Analog film remaster yes they do, but this are not pixels, thatās another matte
To respond quickly
-
There is a mathematical reason for x2, x4, /2, /4 because many algorithms or functions uses much odd numbers (bilinear / bicubic / anisotropic ). If you go into even numbers, you have to interpolate and random decide with number will be chosen. Introduce this notion in video scheme and it produces aliasing or artifact.
-
Why there is a purpose to 1080p ā 4096p, because by doing this you create what is called a āmasterā. I donāt force people to do this way, but if you want the best upscaling result, itās the way to do. (It is by this way movie industry release āupgradedā mid-old movies, 1080p)
As we donāt have acces to original films and scan it to 4K/8K masters, the only way to increase a lot the original quality is to use AI and btw Rhea or RXL to its limits, save the result in the best format your stuff make it possible meaning ProRes 4:2:2 / 4:4:4 or FFV1 4:4:4 10 seeing 12b
Then you can start adding HDR with a great BT.2020 tone map, remove the mosquitos, find edges and so on. When such job is done you downscale the footage to your video broadcaster
(My case is 1920x1080p ā Rhea 7680x4320p ā Hyperion / Artemis / Gaia ā Theia / āLanczos3ā 4096x2304p ā Crop 72p up, 72p down ā Compress 4096x2160p)
I did not find Rhea performs better doing 1080p ā 2160p than 1080p ā 4096p. The reason you may feel the result āfineā is because TVAI is using the Lanczos3 algorithm to downscale the rendered image to desired output, and Lanczos 3 is known to be the best and sharper algo. for it
Finally to reply what you ādont likeā, I suggest scaling your 480p ā 576p with Theia or a mathematical model Lanczos ? Callmut ? and then use Rhea to get your 2160p
To me and dev. team may confirm or not, Rhea is trained with x4 content, and is not simply ābetter trainedā.
2 Likes
Yes but it theoretical, 4x is not the only one fixed upscale factor. Anything that can be divided, or corresponds to 16,32,64,128,256 bits etc. does not produce floating points.
In fact Rhea is the only model that is best optimized for 4x only, but when 4x is the law, why Iris, Proteus and all others are not? When I upscale with Iris I know I donāt have to go 4x it gives me pending on source good results using variable scales.
āMy case is 1920x1080p ā Rhea 7680x4320p ā Hyperion / Artemis / Gaia ā Theia / āLanczos3ā 4096x2304p ā Crop 72p up, 72p down ā Compress 4096x2160p)ā
You can do that but Iām sure most donāt wont this time-consuming pathway. Now I do 4x too and then I downscale to 4k, but I hate it, Topaz should do that for me!
Of course we can set different scaling with Topaz, but itās not optimized, thatās my point. What I want is optimal trained models to Upscale 480p to 1080p, 480p to 2k/4k, 1080p to 4k (and 8k we have already)
If Topaz does that, Iām sure neither of us will do any more the downscale with Lanczos.
Youre using 4096 instead 3840(<-TVs native resolution), we can discuss that. There are pros and cons. Personaly I want to play the TV in native resolution and I donāt want to crop and keep a second video as a master. My 80TB NAS is now almost full.
Just as I was saying Rhea was trained with x1 and x4 input thatās all
Yes I agree, my approach is ultimately a workaround for what the software doesnāt do but should.
Concerning the resizing algorithms, I completely agree that many people donāt know them and use Lanczos without even knowing it.
For the āTrue 4Kā DCI 4K, my case is I admit particular, my diffuser (a video projector) has a native matrix of 4096x2160, thatās why Iām trying to āscratchā the 7% of additional pixels
Otherwise indeed 99% of the material is in UHD, because only the high-end video projectors from JVC and Sony have this resolution. (And the cinemas of course, since Sony provides the matrices)
To me, the model that get closer your need is Proteus. I am not sure Topaz work a lot with <SD footage. Their video samples donāt go in this way. But Proteus has some limitations due to being trained as a āgeneral purposeā, it is even very slow.
oh yes it is! and other models like Theia are very slow too and needs also a quality update.
My method is 100% off, just donāt renew, Iāve had enough of this ongoing series of bug-fixes and unnecessary (for most) added features and user interface messing-around that is Topaz Video AI - who are still to this day having to fix over a dozen bugs in a fortnightly release. Has anyone added up all the fixes they have to push out every year? It must be dozens or even hundreds.
I have and use an old version that ended up reasonablky bug-free - but they continued messing around with it as above so Iām keeping that, and itās still good for what I bought it for: UPSCALING and slow motion of old poor quality video.
Add to that the appalling new EULA and itās plain to see that Topaz care not one jot about many of those who bought it for simple needs in the past and by now it looks like it will never change. So itās stick with the 100% off for me, and continue to use my old fit-for-purpose version that does what I bought it for, alongside my trusty bug-free NLE for video. Horses for courses and my TVAI horse has run its course.
Perhaps, however, there is hope? Topaz have for the first time set a policy of maintaining multiple versions (standard and PRO) so those who want/need the extra features can pay more for the privilege. Now, IF Topaz were were to introduce, maintain and occasionally update models (not interface or āfeaturesā) on just one older and stable version without all the later bells and whistles, my views might change as that could be an entry-level lower cost version needing very little maintaining by the devs. It could even be limited to the original Artemis and Chronos models.
This currently imaginary lower tier TVAI-Lite is all that many users will ever want or need, while more advanced users could pay for the additional tiers. IMO there is room in the market for this third, lower cost TVAI variant, so weād have Lite, Standard and Pro. I hope Topaz will consider doing that as It might bring some of us back into the fold we have now left and could even provide them with a base level of additional income to help support the higher level variants.
7 Likes
Hello.
TVAi 5.3.3 is just another soft so-called patch update, IMO.
Going off the feedback here, from each release from 3/4Q-23 up to today, users are still having YEARS-OLD issuesā¦ e.g. monster faces, player issues, ect.
I was going to give Topaz executives ~1st of this month to address their potential EULA reform, a CLEAR product roadmap for the remaining of '24 - 1H25 (thereās only 2 feature updates left for this year, folks [at best] from Topaz, and so far, IMO, it doesnāt look promising) butā¦ who am I kidding. Topaz, IMO, wonāt answer the call. Just more broken promises awaits, IMO.
So, Iām done here untilā¦ againā¦ to ~4Q25 - 3Q26.
However, I might just lurk around time2time before that because overall, I am a technophile. 
2 Likes
Hi there,
Just a quick question. Does the latest release try to detect what model would be most appropriate? I never know which to use myself.
Kind regards
Yes it does to some degree, by highlighting them. Otherwise, besides reading through the posts here for tips, you have to experiment.
Updated information
Using version 5.3.2 the suggested models appear after you turn on the enhancement switch. Theyāre highlighted by gold circles with a black star.
Thank you, Mike. One last thing. Can previews still be looped? Iāve not used the app in over a year.
On mine
Windows 10 Build 22H2
Intel CPU
Nvidia 3070 GPU
Using
VideoAI 5.3.2
Yes preview looping does work. Thereās a loop selection box that appears after your first preview render in the previews status area.
It also appears if you set in and out points and select to render that area.
As far as I know the 3840 x 2160 are often told to be 4 k as well. Even my LG OLED TV was sold as a 4 k screen.
