BUG: estimated vs actual processing times not accurate in some scenarios
Processing a 720p/23.19 fps source video, 2x scaling, 60fps (there is no 2x fps output option), chronos fast, artemis medium quality, estimated processing time shows this:
Then the counter gets to zero but processing continues, with no indication how much longer it will take, until it’s actually done.
Once processing is actually completed, actual processing time displayed seems accurate:
My guess is it has something to do with selecting a non-multiple value for framerate values. I picked 60 fps because there is no 46.38 option. This needs to be fixed because it will appear to most users that the processing has failed/crashed/etc. Thanks
EDIT: processing the same clip but keeping the original frame rate has no issues. so this is definitely chronos related.
Total Bandwidth : (x8 link): PCIe 4.0 = 16GB/s. Aka, like half a BluRay worth of data per second. There’s no way in hell TVAI moves that much data per second! A few megabytes, tops. These high transfer rates are needed for games (and even then the practical difference between x16 and x8 is negligible).
What causes this? I did a preview and it looked great – a little stabilization and then Proteus. But after completing pass 1 of 2, where I think it is stabilizing, when pass 2 of 2 starts, the screen looks like this. It’s a terrible waste of time. I saw someone post about this on 3.0.6. Its it going to be fixed?
Probably true. But they are not dreadful, they have their uses. Far from being perfect, just as Topaz is not, each has its benefits and flaws. I remember when these “dreadful” rip offs were doing face refinements before Topaz had anything similar, and in some cases it still does a better job than Topaz. But to each his own I suppose.
So is the culprit Stabilization then? I just don’t understand how it can look so great in the Preview but look like a homogenized rainbow in the export. Very frustrating.
It looks like, yeah, what exactly? In my case it looked like a 8 x 4 raster of the same image repeated all across the board, at 99% blur, for each and every frame. I can’t even describe what’s wrong with it, but it needs fixing.
But then, it would first have to identify and track the face including while turning away e.g. for scale, shear and so-on as well as accounting for any foreground masking of all or part of the face(s). Then, if fully obscured or turned away, it woul;d have to immediately pick it up again when usably visible again.
That sort of tracking needs huge computer resources, is painfully slow, and often needs serious amounts of manual intervention and corrections applying. If you google some examples of say Mocha Pro on youtube you’ll see what I mean. All this additional to the processing TVAI is doing on the rest of each video frame.
IMO, this kind of specific face intervention just isn’t a practical possibility with present consumer technology and if it were, it would likely at least double the cost of the software and would not in practice allow anything other than short clips to be processed.
You mean AVFS 1.0.0.6 for windows from 2015 or you manged to find something more stable? In my experience that AVFS easily handle 10-20 minute files, but crashes on longer clips…
Hi. I’m actually using the version that came packaged with VapourSynth R48 64Bit. I’m sure there are newer versions being included in more recent releases, but this one has been working fine for me since 2019. I don’t have any issues using it for high resolution files, only tested up to Ultra HD and about 3 hours run time.
Oh and it might have been fixed in more recent versions, but the R48 version tends to choke up on multichannel audio being in the scripts. I just remove the audio from the script and mux it later, which works fine since I don’t want TVAI touching audio anyway.
You can get R48 and new releases if you wanna try them from Releases · vapoursynth/vapoursynth · GitHub , just pull the AVFS.exe out of the release archive and place it where your existing AVFS.exe is.
Mine has a similar raster, just fewer and larger quadrants, which I suppose may be dictated by the size of the source material?? I have chapter markers in my source - hope that isn’t the cause. Most likely there is a difference in the TVAI processing code between the Preview process (which works) and the Exporting process (which fails with this blurry raster look). This problem happens numerous times and is preventing me from using TVAI. Also, in my case on a iMac 2020, 64 GB RAM and AMD Radeon Pro 5700 XT 16 GB, the process will eventually freeze and crash the computer about 3/4 or so through the job. I was hoping that uploading my logs to Topaz would help resolve the problem, but it’s been going on for several 3.x versions with no word on a cause or fix. Do you think this is caused by using Stabilization?
its not "maxing out " (=too much data for the given bandwith). The chart shows which PCI-E Speeds are chosen - in idle mode, a card can slow down to PCI-E mode 1 with 1 Lane (= “1x” ) in order to save power (not many cards go as low - it depends on the driver, the bios, the board, the chipset, etc…)… As soon as bandwith is required, a card shifts to the max level supported (in modern systems with only one PCI-E Slot populated that would be 16x on PCI-E, Gen 3, 4 or even 5, depending on “how modern” your system is…
bandwith of even 8x Gen3 is so much - if that would be the bottleneck in TVAI, then something is wrong… One can easily pass the whole GPU output over a 8x GEN3 connection from one GPU to another in FullHD and more than 60fps.
yes… Reducing the zombie factors in very low quality source material is not something that will happen next easter - and Topaz can´t do much about it, they have to rely on researchers to combine techniques which are developed seperately at the moment - that will take some time.
Issues which can be adressed by Topaz should be on top of the to-do-list, I am totally with you.
Same here on a 3060… TVAI does not load the TensorRT models correctly, fails on them, finally jumps to the onnx variant and this one works…
With no model on no system I have gotten the TensorRT to work so far - seems broken for now. Latest drivers, latest Win10, several systems, several installs,… nada
Addition: Gaja worked … but speed was not “RT-like”
There’s no way in hell TVAI moves that much data per second! A few megabytes, tops. These high transfer rates are needed for games (and even then the practical difference between x16 and x8 is negligible).
