Topaz Video AI v3.0.11

When the authentication is expired, I want it to stop with an error instead of trial mode.
Since I mainly use the CLI, when the authentication is expired, I sometimes waste time without realizing it.
I had a watermark on a 2 hour video that took 2 days to process, I wish it would have stopped with an error 2 days ago.


I find this watermark business quite annoying, tbh. Either put a text on your screen, in big, fat, bold letters, saying your output will have a watermark on it, or don’t even start the export. Don’t leave ppl hanging for 2 days, for crying out-loud!


Installed without any problem, same performance as before
(AMD Ryzen 5800X @ 4.55 Ghz; RTX 3060)

Proteus auto 0.71 spf; 24% CPU; 22% GPU 6.3/12GB
Artemis HQ 0.57 spf; 25% CPU; 30% GPU 5.5/12GB
GAIA HQ 0.51 spf; 15% CPU; 74% GPU 4.9/12GB

when running all 3 in parallel 11-48% CPU; 82-100% GPU 11.6/12GB
Proteus auto 2.39 spf
Artemis HQ 1.80 spf
GAIA HQ 2.87 spf

1 Like

uninstall and delete the folder
When I changed the driver to an old one, it was fixed for some reason

I finally installed this one after only doing some tests on some of the beta versions couple months ago and I have to say I’m somewhat disappointed. Coming from 2.6.4 I feel more comfortable going back to 2.6.4 than staying on 3.0.11 so hopefully I can give some pointers on what I feel is wrong.

Starting out with the UI, I feel like in general it feels worse though it does have some improved features. In terms of responsivity it feels more responsive than it was in the beta version I tried but yet still feels worse than 2.6.4. The 2.6.4 UI is easier and quicker navigate through, though I’ll say that I think the individual menus of enabling the certain features you want instead of just a list of models is a good improvement. I also think that the ability to change the amount of time you want to preview is a nice addition but at the same time the previewing now feels like a hindrance compared to 2.6.4.

Previewing is quite laggy on playback, scrubbing the timeline is laggy and the time it takes to load and process after clicking preview takes too long making iterating between settings slow. While it does show the time it takes to process a frame or fps (which ever you chose in the settings) that number is just replaced by the export time once it’s finished processing the preview, it would be nice to have that number shown even once processed like it did 2.6.4. In 2.6.4 I quite liked the ability to 4 way preview and compare multiple models in different views that you can zoom and scrutinize without clicking between them, I don’t know if that’s possible in the new UI but it isn’t obvious to me how to do that or if it exists at all. I also think that now that the previews are processed as files (maybe it was way before) it seems that possibly it’s slower because it’s actually taking the time to export each preview than just maybe saving it to ram or something. I don’t know how the previous preview was implemented I just know that it was faster.

Processing times are also disappointing. I can understand that moving from 8bpc to 16bpc operations can slow down the processing time but was under the assumption the general improvements of 3.0 along with TensorRT models might make a significant improvement to the performance but yet I’m left with performance that is either the same or even worse than 2.6.4. I don’t know if I’m even using the TensorRT models but I’m just using whatever model it auto downloads when I click on the model I select. I’m currently using a 4090 with a 5950x with 64gb of ram and previously had a 3090. The performance difference between the 3090 and 4090 is not worth comparing since it largely performs the same with marginal differences that could be attributed to multiple variables but I assume that it isn’t even properly utilizing either cards considering that I’ve seen much worse cards perform exactly the same.

Model quality remains the same but I know what I’m getting when I use VEAI and personally I’ve never really had many issues with the quality of the models aside from the issues I made clear in my post about the color handling in the beta. Though obviously better models would be nice and probably expected on a major version update. I know some have mentioned about Nvidia’s “RTX Video Super Resolution” though personally if its anything like their MAXINE effects then it isn’t really close to the fidelity that VEAI produces. Perhaps the Artifact Reduction might be useful but their SuperResloution or Upscaler that I tried wasn’t that impressive. *(Edit, after looking at the RTX Video Super Resolution more it looks too good to be true, If you compare the differences between the original and upscaled frame there’s details from the original game that aren’t even present in the original image. Perhaps it’s sampling temporally throughout the video and that the original might look better in some parts but Nvidia just cherrypicked the most compressed frame out of the video.)

I think the general sentiment I have is that after using one version of VEAI for a year and moving onto the new version I don’t feel like much has changed and even feel like it’s gotten worse in some aspects like performance or that there are feature’s that I’ve lost which of course is going to feel bad. I understand that a rewrite was needed but personally I’d rather use 2.6.4 until 3.x is in a state where it objectively feels better to upgrade because right now I don’t see much reason to pay for the upgrade when it feels like I’m losing more than I gain.


720x480 source mkv file
Processing with artemis medium model to 1920x1080.
If video is treated as 1:1 then processing is around 15 fps.
If video is treated as 4:3 then processing is around 5 fps.

3 times slower!

Its better to process 1:1 and then resize it with another software to 4:3.

And deinterlace is still not fixed:

on the right where it says PRESETS, VIDEO, FILTERS, click on ENHANCEMENT, then change VIDEO TYPE from progressive to INTERLACED. Then you have to know for sure if its TOP FIELD or BOTTOM FIELD. You can check and know with 100% positive by opening your video file into a tiny app called “mediainfo”. It will give all the specs from the file itself, including if the interlacing field order, because its either Top first or Bottom first, this is critical to know. Topaz tries to AUTO select top or bottom, but I always set it manually to be sure. My interlaced changed to progressive come out looking like it was purchased that way. You can also see the correct FPS from “mediainfo” meaning variable or constant, and the rate itself which is usually 23.97 regardless of what topaz thinks. If its a BBC video, its always 25fps. If the file is VARIABLE frame rate instead of CONSTANT frame rate, you need to first change the video clip to be CONSTANT before putting it into topaz, then topaz will handle Interlaced perfectly. I’m not entirely sure that Topaz can deal with variable FPS rate videos. Variable is more common with older DVD’s and far less common on blu-rays. I use HANDBRAKE to first change the frame rate to constant IF its a variable frame rate video (not common). In handbrake, just put the bit rate 20-30% higher than the DVD is and its basically lossless. Be sure to use the exact same codec for the audio in handbrake so it does not change the audio at all. Then let TOPAZ do its magic on it, and it will be smooth and flawless. I usually change my conversions that came from interlaced land and variable to constant, then deinterlace using Dione, then change the frame rate in topaz when done (a second render) to 59.97fps using Apollo. Apollo is better than Chronos, it can deal with frames that are blurry, where Chronos does not deal well with blurry frames, like 20-30 frames with a fast pan.


Seeing this in changelog - “Fixed TensorRT models not running on 4000-series GPU” - will this also be fixed for 3000-series GPUs?

It seems like the client has issues scanning through .mpg/.mpeg format videos- some I’ve worked with outright can’t be scanned through with the preview window and some will only recognize half the video and will only preview half of it. This also leads to export estimation times being incorrect as it’s not sure what to guess as the model works.
I’ve seen it say this, or in the case where it recognizes some of the video to scan through it will give an estimation to the model reaching what it thinks is the end point, and then will say the video is at 100% processing while it finishes the rest of the video.
I am assuming this is a bug because despite the client, the export works just fine.

Does your 3080 run @ PCI-E 3?

I’m very happy so far with 3.0.11 4090 performance looks to be improved.
Black Desert online 2x gaia CG completed in 4 minutes 48 seconds

Slower than “4 minutes 17 seconds to complete 2x upscale Gaia computer generated Video AI 3.0.2 = 0.11076 sec/frame
Black Desert Online-Wheel_235677003_gcg5” which I was only able to get briefly before the performance was reduced until v3.0.11

IMX50 PAL to 4k50p dione 2x deinterlace at 0.23spf. IMX50 PAL to 4k50p dione 2x deinterlace at 0.07spf. Serious speed increase over the previous version but I don’t have direct comparisons but PAL to 4K was in the region of 0.35->0.6spf varying for seemingly no reason.

Also on my first test with PAL it seems like the 1 frame blend on scene transitions isn’t happening anymore. That is a huge improvement if it’s consistent.

There’s a small issue where my one GPU is being listed 3 times while it had mistakenly only been 2 in the previous versions. I don’t think it effects anything but just FYI

7950x (-10 curve)
128GB ddr5 4600 (62GB/s, 82ns)
MSI Suprim X 4090 (stock 93% power limit)
Windows 11
Nvidia studio driver 527.56

1 Like

I converted the same video in TVAI (a boxing video) and found the motion smoother via chronos fast as opposed to Apollo which looked kind of jittery in motion. Maybe that was a one of relating to the source, as usually Apollo is my go to.

1 Like

Function to update presets is missing. Currently I have to create new preset if I change something on the actual preset. Can this be added? Similar to what Lightroom has in export presets maybe? Right context menu? and I believe “Output Settings” are not saved within the presets, correct?

I have a lot of DVDs that show frame rates of random numbers like 21.489 in MediaInfo. I think it’s because they are marked as variable frame rate. Either way, I always convert DVD sources to a constant frame rate before loading them into TVAI. Mostly because VEAI couldn’t handle them correctly.

1 Like

Mine does (X299 chipset), but I’ve never had issues with that on 3090.

I plan to buy a new graphics card recently, and noticed that the new 4070TI bit-width is very low. If I use this card, will the rendering speed be slower than 3090 or 3090TI? 4090 is too over for me, Is that 4080 more better than 70TI?

Thus. I checked the proteus model in this version on the RTX4090 and … no changes. Just as slow as previous version 3.0.
Version 2.4.0 runs over 50% faster.


4090 performance is still super slow at around 2.5 fps with proteus. I do not want to run 3-4 parallel jobs at a time. I want to run one job and I want it to be as fast as possible.

2 things that I’ve noticed in this version:

  1. Gaia HQ, which has always been quite a bit slower than the newer AI models, is actually faster than Proteus Auto. In this case, Gaia HQ - 4.8 fps, Proteus Auto - 3.8 fps.

  2. The alignment of the preview is off more often in this version.

1 Like