Thank you for the detailed information.
I have located the log files and .tz files.
I looked at the contents of the log file for a bit and then closed the file.
It seemed difficult to me and I couldn’t understand the contents.
Has anyone noticed any audio stuttering issues with output .mp4 files in 1.7.0? These artefacts weren’t present in 1.7.0.3b as far as I can tell, but I upscaled three videos with 1.7.0 release and they all have small amounts of stuttering.
Not a major issue as I can re-mux the audio from the original, but not sure what would be causing it, as don’t think much changed between final beta and release.
2 Likes
It work great on RTX 2060 but encountered one issue.
When 4x (ghq5) scaling of 720x480 video with FAST model, it is actually slower than default model.
Default Stock Model: 5 hours to complete
Downloaded Fast Model: 10 hours complete
What wrong ?
Btw no issue on 2x (ghq5) scaling of 720x480.
Seems like it’s always using the 288x288. In the logs I can’t find something about the used model.
Cuda param set to: 2
main onCompleted Sat 2020-11-07 05:42:14
--------- tbackend.modelInfoMap [object Object]
TPostProcessing::rebuildList(): Post Processor Loaded “undefined - undefined” from user file.
I don’t get it. I think I’m missing something 
As I said upthread, look in the log for the text “Loading time for model file:”
That will be immediately followed by the file it will use. Look at the couple of dozen lines before that and see if it tried to find some other file.
Before replying I checked with Topaz to make sure they were OK with me posting my methods, and they said it was OK. The process is fairly easy, but lengthy, so here it is:
First step is to determine which files are correct for you, then second step is to create a list, then the third step is the actual download.
There are 4 variants (at least that I know of) for the models for different GPU types, and the files have the following endings/extensions:
-ov.tz (openvino)
-ml.tz (coreml)
(fp32) -ox.tz (onnx)
(fp16) -ox.tz (onnx16)
You will know as soon as you look in your models folder which one VEAI is picking for you - all of the files for 288x288 will be there in 1x, 2x, and 4x versions. For me it is fp16 -ox.tz
There are 9 HxW variants of each file, and three size versions, except for thf which has only single unsized version. The HxW list is:
256x352
288x288 (you don’t ever need to download this one)
384x480
480x384
384x672
448x576
576x384
576x416
576x672
Basically you can use the 288x288 files to build a list, repeating file names with the other sizes, plus addding the URL. Here is an example from mine:
http://veai-models.topazlabs.com/alq-v7-gnet-fp16-576x384-1x-ox.tz
I built the entire list in a text file during the beta, and then when beta 1.7 became release 1.7 and alq and ahq changed from v6 to v7 it was a simple find and replace.
For the download I use a program called jDownloader (free) and it will pull from the Windows clipboard, so just select the files to download and copy. I’m sure there are other batch downloaders available. If people have favorites please chime in.
Ask if there are parts I need to clarify.
4 Likes
Thank you very much, was looking at the wrong spot…
For upscaling of 720x576 (DVD) to 1080p (1920x1080) it’s loading the 288x288-2x model (200% (2048x1152)).
So I guess there’s no way for faster upscaling of DVD content? 720x576 is larger than any available model. Am I right or is there something I missed?
Waiting for DEINTERLACER AI 
4 Likes
Doesn’t sound right. Here, 720x576 to 1080p loads 576x384-4x file. Did you look up a few lines and see if it even tried to get another model? Which model sets is your system using, ox.tx or one of the others?
Hmmm. On my end it’s checking this (guess searching the best suited model)
Block: 288 288 6 0.5 0.2
Block: 480 384 4 0.977778 0.777778
Block: 384 480 4 0.977778 0.777778
Block: 352 256 9 1.40556 0.955556
BLOCK: 288 288
[N]-v[V]-fnet-fp32-[H]x[W]-[S]x-ox.tz FILENAME: ahq-v7-fnet-fp32-288x288-2x-ox.tz
Loading default model file: C:/Users/****/AppData/Roaming/Topaz Labs LLC/Video Enhance AI/models/ahq-v7-fnet-fp32-288x288-2x-ox.tz
That’s about it. This happens on two seperate machines. (This log is from a 1070ti)
Where tor read full faq about artemis,gaia,theia, what is the difference?
1 Like
I hadn’t looked at those, and I see that it does appear that it is saying (lower number) that 288x288 is better for yours (although it is not clear why it is using 2x instead of 4x). For me, there are two better numbers:
Block: 384 576 2 0.166667 0.0666667
Block: 416 576 2 0.255556 0.155556
Block: 288 288 6 0.5 0.2
Block: 480 384 4 0.977778 0.777778
Block: 384 480 4 0.977778 0.777778
Block: 352 256 9 1.40556 0.955556
BLOCK: 384 576
This is with a 2080ti
I have a 2070 laptop that doesn’t have 1.7 loaded yet; if it gives different numbers I will be back to report.
Glitch in FAST MODEL ghq-v5-gaia-fp16-480x384-4x-ox.tz cause major slowdown on RTX 2060.
I mistyped 2070 in my prior post - the laptop is a 2060, but the block files and chosen model are identical to the 2080ti. I wonder if the lack of tensor cores in yours is the reason?
If you are just doing 2x upscale, set VRAM at max, you will get better speed.
1 Like
Don’t set Max VRAM too high if you are doing 4x upscale on a 6GB card. It will slow things down.
1 Like
Just a note, the only models included with a default offline VEAI Windows install are fp16 ox 288x288. It will need to download the 288x288 OpenVINO and/or 288x288 fp32 ox models, if the user needs them.
Tested VRAM in medium low fixed slowdown issue. However performance no much different from STOCK model.
What is VRAM requirement for 4xghq?
Actually 400% vs upscale 2160p, are there any different in quality or waste storage for 2160p?
I’ve downloaded the latest version which finally recognises upscaling videos over 30 minutes to 4k! However, my 4k conversions Have jerky playback (as if it’s skipping many frames. Is this something to do with Compression Factor? What do I need to change to get smoother playback (evreything is perfect when upscaling all the way to BluRay)?
Thank you, great info.
