I’d install without any problem, and also it performs without any problem, will say that processing speed got not worse than 3.0.0
I’m using
Windows 11
AMD Ryzen 7 5800X with 32GB and SSD 970EVOplus @4.5GHz
ASUS RTX 3060 with 12GB and 522.25 driver
I upscale 2 videos in parallel 960 ->1920 with Proteus Relative to Auto and individual preset (0.32spf each, 0,23spf when only one), H.265, MP4
CPU load: ffmpeg @18% for each video (24% when only one)
GPU load: 90%, memory 6.8/12GB (90% and 3.2/12GB when only one)
So far nothing special to be reported.
Geforce Experience optimization still not supported for 3.0.1
As mentioned in the Facebook page, I don’t think its a good idea to have feedback for the newer release version in the same thread as the older version - I am not sure how lost this feedback will actually be in the noise.
However, in relation to initial impressions, issues, or bugs in 3.0.1:
- Preview is “better” but still having problems. The left and right images are not linked - they are not the far off each other, but its weird. The left image will move but the right image doesn’t, the right image will move and the left image doesn’t.
Overall, on scene changes you will see preview comparisons like this, which are clearly not the same on left and right:
- The model loading is still strange. I can click preview, and it “downloads” the model, but according to that status bar, it is apparently “downloading” the model everytime I hit preview, even if I have done nothing but click the preview button for a second time, IE no changes to settings.
Additionally, each time it preview it takes as long as the first time to preview, whereas once models were downloaded for that sequence in V2, initialisation of preview was almost instant. This one seems to keep doing things in the background first.
Its not a large issue, but it means it doesn’t load anything as fast as the previous app, rather still slower.
- Export as image sequence auto fills a name for the export, as in Topaz Is putting this name in. I left it as in and tried to export and it errors out, so there is a bug here: If auto filling the name, make it valid, or leave it blank.
- This next one makes me happy, but also needs work because it took me a while to decipher what was going on - the behaviour from Topaz was bizarre to say the least.
When you choose the same export folder and name for image sequence, I am glad to see that it actually does overwrite the existing sequence. So finally, after a long time of this issue, I can re-export over the top of an existing sequence if I feel like changing the model - thanks.
However, the GUI in Topaz makes no sense at all what this happens.
I am going to explain exactly what it is doing, as this took me a bit to work out, hopefully this won’t confuse anyone.
There is a preview window (top right, shows result), a progress arrow that is tied to that window and a blue progress bar underneath the arrow on the timeline. The below image shows the arrow and progress bar in processing.
When you start an export that overwrites, Topaz does two things:
- It loads the image sequence saved in the export folder and uses that image sequence of the preview.
- It starts processing the new images with the blue status bar and overwrites the old images.
In other words, the Preview window and the Arrow have nothing to do with the new export you just requested. Additionally, as the loading and preview of images goes faster, what it does is display them much faster than the blue progress bar. It will only show you the actual export image sequence if or when it reaches an image in teh sequence that does not exist.
So if you export the first 100 frames of a sequence that is 1000 frames long, cancel the export (or was from a trim), and then export the whole sequence using a new model, it will
- Display the existing image sequence as a preview, too fast for the processing to catch up (old images)
- When it reaches the 100 frame mark and no images exist, the Arrow will stop and preview will freeze until the Blue Bar catches up to the frame number
- At that point, it will start loading the new image sequences and the Arrow and progress bar will sync back up again.
- I know quality of models is one thing that appears to be on the backburner, but as this is the first time I have been able to do a direct comparison in a while, this is a comparison slider as an example of degradation in model quality.
On the left is ALQV8 in V2.3. On the right is ALQ (no idea version) in V3.0.1. Both input sequences are identical. The only different setting in V3.0.1 is Grain as the new settings are weird and hard to work out the equivalent from V2.3.
-
There is a minor, unimportant typo or message error when in application updating - no file even loaded, there is no export in progress:
https://i.imgur.com/nJgdJTD.png -
Ongoing - Still need to find a solution for Avisynth import.
2 Likes
The files are big – and when you close it properly, it rewrites them for some reason (which can take a minute or two and require more SSD). I have found it faster to just force halt the software and reboot.
1 Like
you can find the link of 2.6.4 here :
1 Like
- if you want to be refunded if you did an upgrade to 3.0, you can write to the support and ask for a refund ! there is a 30 days refund policy at Topaz.
- 2.6.4 had not future and couldn’t be updated. all the features that people were asking for 2.6.4 and we’re starting to see a bit in this 3.0 version was not possible on 2.6.4. That’s why it’s leaded to this 3.0.
- is this 3.0 version full of bugs ? yes.
- is this 3.0 version was not ready for such an early release ? yes
- are the devs and the team behind of it are aware of the situation, and understand the overall frustration since the release ? yes.
- does the team didn’t listened to all messages told in the beta ? no. it just take a small look at he software to see that a lot of request done in the beta thread and before have been implanted. Av1 decoder/encoder has been implanted in this 3.0.1, that’s why it ask an upgrade of the nvidia driver to the last version.
- will the team will go back to 2.6.4 and improve it ? No, no matter what you’ll do or ask, it will never happen.
- will all 3.0 issues be fixed ? we’ll see !! i hope, like everybody here.
7 Likes
Hi 
it will not be supported before a long time. nvidia will certainly only put it in Experience once a stable release will be out for 3.0.
Yes, 2.6.4 speed might be better than 3.0.1. However due to H.265 support in 3.0.1 output files are only 50% at same quality. Therefore 2.6.4 is obsolete for me.
2 Likes
It’s obsolete for you because you like the new version 3 simply 
1 Like
Well, there is still a good mileage to go before I’ll really “like” what VEAI does deliver:
e.g. face refinement, pause / resume, better understanding / description of features, etc. …
1 Like
For my part, I will still use 2.6.4 or 2.3.0 with the old models when version 3 will always be slow. When version 3 is fast, then maybe I’ll really get into it. For now version 3 is a regression of VEAI and I will always say that for now 
1 Like
Speed is an important argument:
When I started to use version 3 betas I upgraded my PC to improve processing speed.
Before it took ~140Watt and 0,17spf, now it’s ~270Watt and 0,24spf for same input video. That means 170% increase in resource consumption.
Therefore I expect VEAI sometime to improve speed back to 0.17 spf (for same video) to recover resource consumption a little bit.
VEAI 3.0.0.8 is very speed for me, after this version, it’s very slow, i don’t understand.
1 Like
3.0.1 crashes a few seconds after I try to export anything. All the same settings as 3.0
if not done yet, check your driver update, minimum requirement for the nvidia driver version has changed ! (they updated the changelog)
1 Like
I’ve installed the latest driver (RTX3090) and completely uninstalled/reinstalled Video AI, and it seems to be working now… just ~50 times slower than before :-\
on which source and which filter you have the slowdown ? we found an issue with 4/3 source, which must be enhanced to 1440x1080 for example, instead of 1920x1080. the add of black bars make the rendering to be very slower. a normal upscaling is a bit slower, but not 50 times slower (exemple of 720p-> 1080p or widescreen SD material as source).
i can’t talk about other filters, i don’t use them, except chronos but didn’t had the opportunity to test it yet.
50x slower? Are you sure?
Upscaling a 1280x720 Mediathek video which content 960x720 pictures plus black bars left and right, I cropped first the pictures content without black bars, then upscaled the crop with bars to 1920x1080, and I’m surprised: it runs faster by 0,04spf (=-10%) than direct 1280x720 to 1920x1080.
1 Like
Thank you for explaning it to him…
At least you tried…
1 Like
lol yeah, i didn’t want, it’s useless, but at least the message can be read by other people which could write some questions close to this.
1 Like
