With GTX1070 GPU, and 6700K CPU upscaling a 1080p clip 2x to 2160p with Gaia and ‘1’ grain turned on, I get GPU load almost pegged to 100% and this for performance:
Argus monitor showing full GPU load:
With the same settings with Proteus model I get roughly double that speed but GPU doesn’t seem to be fully utilized.
Well it is absolutely good to know what the frame per second speed is with this version.
If it is a little slower, but the video output quality is higher then the slowdown is worth it in many situations.
But if it is slower and the output quality is not better (using the same process settings) then I’m not going to blow $150 on an upgrade and I am going to start to worry how long the 2.6.4 version will be supported.
I see that it downloads something before starting a project 1/2 the time. What is it downloading and is this always the same file or does it get updated?
I have a render farm and I don’t know if this is going to be a good solution moving forward. Plus I have no problem buying multiple licenses for each computer.
At $299 – this is not a “consumer product” – it is definitely a “prosumer product.” That seems to be how it is advertised.
I would certainly hold my money until much more progress is made on refining it. This is an unpolished beta at best (that’s being generous). That said, see what they do for Black Friday.
It’s downloading the model based on your input source and output settings. If you have a model already that fits the need, it won’t download it again.
Yes - I understand – I’m just wondering when that will cease for version 2.6.4 or is it using the same model for every version.
It is hard to determine if this is a good investment (as I am assuming that there will be extra licensing costs for multiple computers in the future, especially when they are now providing CLI).
EOL for 2.6.4 is not indicated.
This is a problem when you always have to rely on the company to continue serving you models for the software to function.
Wow - are you saying that your render speed is 1.76 seconds per frame?
I haven’t seen anything on a cheap Mac Mini M1 at a speed that slow. Most things I do range from .09 to .25 frames per second. And there is no noise, heat or significant power use.
I will mention, that you can use any previous version of the software regardless of age. I still open 1.5.3 on some instances, and all my upscaling is on 2.3 currently, which is not the latest release of the 2.x series. I would probably still use 1.5.3 except the time saving on 2.3 is significant so I try to replicate the same results I get in the early version to 2.3.
Already paid licence won’t expire for previous versions, you just won’t be able to activate new ones after the licence expires. You will not get any notice on EOL for 2.6.4 - at least I didn’t for any previous version. They just release new ones which you can choose to install or not.
Someone said earlier in the thread that they believe the same model is being used, but the issue on quality change is because of what appears to be a bug in the translation from GUI to the CLI - while you may not interact with the CLI directly, it still uses it in the background. That user is reporting the CLI command being made is wrong and if you manually correct it, the quality is the same.
Also note that Topaz supported CLI for a long time. I experimented with it myself for a while, was just easier for me to use the GUI. It was only recently they removed CLI, so I would hope licencing costs wouldn’t increase as CLI was supported for years before they removed it. This is the guide for CLI interface from a year ago. I just don’t remember exactly which version it was removed.
https://topazlabs2.zendesk.com/hc/en-us/articles/360046455552-Command-Line-Interface-Documentation
They have eliminated the ProRes 422 output option?
Yes
The M1 chip is quite powerful so this is not surprising.
Gaia is always slow. I can remember testing Gaia way back to really early versions, and it is always double if not more the time of the other models. It can give really good results in some cases, but it is not a fast model.
I use a 1080 currently still, and I upscale from 540 to 1080 and thats at least 1 second per frame, and normally closer to falling between 1.3 and 1.7, so 1.76 seconds for a 1080p to 2160p sounds pretty good. Unfortunately, Gaia has not been worked on a for a long time, including this release, so I don’t expect it to get a large performance increase anytime soon.
Ok - so based on many posts on speed for Mac M1 machines – Version 3.0.0. is a no go.
Supposedly 2.x and 3.x are using the same models, except some have had Tensor cores enabled in them.
I used VEAI to show the export command line, changed scale=0 to scale=2, opened the command prompt from within VEAI, pasted the command with scale=2, and it ran.
I checked the result, and it’s exactly the same as what I posted above (and below). There’s no change in the frames created by VEAI with scale=0 changed to scale=2.
So, the statement you made is not correct. Whatever is going on, changing “scale=0” to “scale=2” has nothing to do with it.
It’s not worth updating the below image, because the VEAI 3.0.0 result is pixel-per-pixel the exact same with “scale=0” changed to “scale=2”.
Again, I posted the source file that I used. Try it for yourself and see. (Imgur resized the image from 1280x960 to 1200x900 for some reason. Just use Irfanview to resize it back via Lanczos for comparison.)
(Again, it’s Artemis Strong Dehalo v2 (alqs-2), with a 2X upscale.)
(Right click → “View Image” for full size.)
test_sample.zip (745.1 KB)
This forum supports directly pasting images into your post, you don’t need to post them on sites like imgur first. Just drag the original image into your message.
It’s an animated PNG. The Topaz forum doesn’t support animated PNG’s… the frames don’t animate. But it works when loading the animated PNG externally.
Ok then nevermind, lol
But not by frame numbers! 
When I did checksum comparisons on the exact same model files from VEAI 2 and VEIA 3, the checksums don’t match. And the results (like I posted above) also don’t match.
I know at one point, we were told you could copy models directly out of the 2.x folder into the 3.x.
2.6.4 has the model manager in preferences, and you can download all models. For me, it didn’t work in one go… I had to go back in, open each model box, and select the models that weren’t downloaded completely. I did it one model at a time. So make sure you go in and open each model node, and make sure all versions of each model are showing with a check mark next to them. You may just want to do it one model at a time to insure it works best. I know that takes time, but it only needs to be done once.
For me, on a PC, this downloaded all optimized versions (480x384, 576x416, etc. etc.) for Open Neural Network Exchange for GPU processing (both the fp16 and fp32 optimized files for 1x, 2x, 4x-ox.tz) and Intel OpenVINO for CPU processing (fp16 optimized files for 1x, 2x, 4x-ov.tz). When the download was done, the model folder has 6,484 files at 190 GB.
VEAI 2.6.4 should continue to function like this forever. If you’re really paranoid you can block the program in Windows Firewall.
I believe the model files download from Amazon web services. I backed them up to pCloud… all 190 GB!
I’m actually an OpenVINO user, and VEAI 2.6.4 successfully determines and uses an OpenVINO optimized model (I believe it’s the 480x384 for me). In 3.0.0, the optimized model either isn’t implemented, or doesn’t work, and I only get the default 288x288 non optimized model for OpenVINO. It also runs much slower, and produces different results that I don’t like as much as 2.6.4.
You can “trim” the video with an in and out point, but then the trimmed version expands to fill the entire timeline so you can’t see where within the overall video you’re working.
The bigger problem is that if I wanted to extract 2 segments from a video, I would add the video twice in 2.6.4 and set the in and out points differently on each one. There’s no way to add a video multiple times in V3, however.
