The Tesla V100 will be fastest for deep learning. But considering the cost of a Tesla V100, I would be absolutely sure it will work in VEAI before buying it. It would really blow to spend anywhere from $7,000 or more and find out it doesn’t work or perform as expected. And I’m not sure if Topaz devs are going to be comfortable giving you any kind of a guarantee (in writing) that it’s going to work. I mean, I guess they wouldn’t “guarantee” any specific hardware, but you might ask if they’ve done any testing or benchmarking with the Tesla V100, and what their results were.
Sure would be nice if, after they tune up the deinterlacing, they added a detelecine filter to get rid of those ghosting interpolated middle-frames when a 24fps video is deinterlaced.
Ben, that’s what I did. Take a look at this screenshot:
As you can see, there is no envelope icon – or a field where I could type a comment… 
That is strange, my profile settings allow for other users to send me messages.
I had this issue before, and it’s not an issue, i think you don’t have the autorisation yet to write a private message to an another user or a beta tester. i had this problem too when i first subscribed to the forum, and after a while, (not sure how much) the possibility to write was available.
for now, ask to benjaminrettig to write to you as he’s allowed, until you’ll be allowed to send PM at your turn.
So guys, something else I have noticed: Since you already know I am currently doing a lot uf upscaling work on “old” (sorry, but by todays standard they would be considered as such) SciFi series like Star Trek DS9 and Voyager and I noticed problems with star patterns. Almost every model seems to have difficulties handling space scenes / star patterns. Now I know the DVDs don’t give much detail and resolution, but when working with StaxRip and QTGMC (ok, there is additional filtering to compensate for aliasing and rainbowing) there may be a few less stars (some are apparently being lost when filtered), but they appear consistent and in white coloring. When I am working with the Deinterlacing models (Dione, no matter which one) the stars are flickering like crazy in white, blue and purplish. Also, sometimes they randomly disappear and reappear (perhaps those are the ones that are being eliminated by QTGMC due to inconsistent visual behaviour). In any case, that was the reason for me to not rely on Topaz-based deinterlacing just yet.
1 Like
I really hope they say how they upscaled their DS9 example. Increadable 4K video example.
It was a selling point for me. I have Voyager cause I wife loves the show, but not DS9 and I would love to upscale babylon 5. How bad the resolution is for those TV series is the only reason I haven’t bought them yet.
Is it legal if I pay you for your upscaling DS9 if I do own the series, but haven’t ripped it myself?
This upscaling takes a long time and spreading the workload between two or three people would make it go a lot faster. We could setup a FTP server or something.
Tesla V100 is more useful for the training itself. Not sure about any benefit from the convert part (upscaling). Not to mention it’s normally a data center card and you would need another card with DP/HDMI output.
Anyway Tesla V100 will become now obsolete with the new A100 and the 3rd tensor cores gen.
I don’t know if english is your second language, but what are you trying to say?
They just released 1.9.0
Do you need help with how to clean up or prepare your videos to be upscaled or are you just fustrated with how long it takes to upscale videos?
I personally find it frustrating how long it takes to prepare your video and its even very complex to deinterlace videos, increase there frame rate to 60fps, and some how not lose half the stars or make people look like aliens. I haven’t been doing this for 10 years, but for more than that I wish I could.
Could not create a new post in the discussion category, so posting it here.
If you seem to have a slow upscaling speed (> 1 seconds per frame) and also have an NVIDIA GPU- I encountered this problem and support helped me out.
Each time, ‘Initializing AI Engine’ took me 10-15 minutes! I knew something was wrong.
The errors shown in the log:
"
ORT Session creation failed
Loading default error reading file: C:/Users/user/AppData/Roaming/Topaz Labs LLC/Video Enhance AI/models/gcg-v5-gaia-fp16-384x480-2x-ox.tz
Unable to create or load model of type ONNX16 from path: gcg-v5-gaia-fp16-384x480-2x-ox.tz
BackendRunner cannot be initialized for model: gcg-v5-gaia-fp16-384x480-2x-ox.tz
Model Runner init failed for index: 0
…
ONNX problem: DML Init D:\2\s\onnxruntime\core\providers\dml\dml_provider_factory.cc(136)\onnxruntime.dll!00007FFE07F5287F: (caller: 00007FFE07F52917) Exception(3) tid(d574) 80004005
"
The solution:
Reinstall your latest NVIDIA driver (even if yours is the latest) and select ‘Clean Reinstall’ checkbox when in the installation wizard.
If this doesn’t work, try the same with the Studio-grade NVIDIA driver.
Proposed to support to add an error message which will warn the user that the GPU acceleration could not be initialized, with suggestion to reinstall driver using ‘clean reinstall’, and also stop trying earlier than 10-15 minutes of trying.
Hope they pick it up 
Also, kudos to Emily and the dev team for helping me out swiftly and professionally!
Much appreciated and keep up the awesome work!!
3 Likes
It would be nice to have such documentation for sure. As it is now, each user must go through a learning curve of trial and error, and only after a few weeks of experimentation does one begin to get a feel for the application’s capabilities, and the dos and don’ts.
I think he’s wondering why we have no current alphas or betas…as am I. The cycle has slowed down tremendously.
I’ve completed the 1st ever episode of OFAH. Spent 3 weeks removing dust, dirt and other video errors. Very laborious but looks MUCH better than official DVD release.
If this helps: I’ve been outputting 720x480 footage as 110% scale, which gives you a 720x528 render. So you aren’t losing any horizontal resolution that way. If I’m deinterlacing with DioneDV, I do that, then use that result to upres with ArtemisHQ.
Or has someone found a way to change the aspect ratio (“SAR” in bottom attributes bar) for imported footage files?
Unfortunately the sample I provided did not represent the whole story. As I have said, several issues have come up that were not that immediately visible from the sample, so my enthusiasm might have been a bit premature. Still, my work continues.
Unfortunately you can’t pay me for my efforts since that would be illegal (copyright and so on, even if it were just one fan to another).
they will all be good for VEAI
i think the rtx3090 is the most powerful on your list.
i got a 3090 and you must know, VEAI does not use the RTX30x0 Series up to their limits, yet
to get the most out of the card (100% gpu load) you need to run 2-3 parallel VEAI instances.
if i had the choice (with that knowledge today) i would save some/500 bucks and take the 3080
you really needed 3 weeks for 1 = ONE single episode???
sue me, but i think i get the same result (your screenshots) in 3hours with the right VEAI model
i upscaled the whole DS9 with 7 seasons (more than 140episodes)
in 4-5 weeks (ok not to 4k, just 1080p)
short: something is wrong on your side, or you really need a faster gpu
thank you for sharing this helpful infos with us
Help plz
What doing “disable model downloading”?
That will disable the downloading of models customized for your GPU. There’s really no reason to disable it unless one of the models is crashing. But a better solution in that situation is to restart your computer or manually delete the problematic downloaded model file or files and have it re-download the model again.
