A preview queue is planned for an update very soon. This will allow users to queue previews and navigate to other tabs without cancelling the current task.
6 Likes
It’s a process, after all, early adoption, these things. 
The question is - are we able to buy Blackwell GPU for a MSRP anytime soon? 
Every reviewer did post that rtx 5090 is not supported in tensor rt and not optimised at the moment.
So, from the past, maybe in a half year we will see the right performance … but even now … i did not expect this low numbers.
4 Likes
That is very Great News! … Thank You Tony
Setting an In/Out point on the timeline while loop playback during previews.
We don’t expect it to take that long, we’re working on having the optimized TensorRT models ready for launch day.
7 Likes
Great news… now we wait for the CFR fix 
2 Likes
We have not forgotten! I know this has been a frustrating one, and we’re looking into a true fix that doesn’t force CFR through duplicate frames.
On the topic, we have seen some reports that ffmpeg exports even without tvai_up are exporting as VFR. Has anyone experienced this with a Video AI ffmpeg build or a different one? Any data on this bug helps
1 Like
Max 25% better I predict…Nvidia - pay 25% more for 25% more performance, and use more energy!
50 series - what a disappointment.
1 Like
Great news. Now we’re waiting for (finally) some performance optimisations on Apple Silicon 
Especially Iris, but also other models are clearly too slow and do behave quite erratic in terms of speed.
If you want real performance, you need a discrete GPU. Not some Apple hyped up M1 chip etc.
3 Likes
Obviously you don’t really have experience with TVAI and AppleSilicon.
Of course those chips cannot compete with a Nvidia 4090 but are also MUCH less power hungry consuming at most 150W in TVAI for the whole machine (CPU+GPU+NE+RAM+Mainboard+Harddisk,…) and in desktop operations (Webbrowsing, Office or Movies) only 10-15W. Not everyone wants to replace ce the rooms heating by a PC with a 1000W power supply (or even a second PS just for the gfx-card).
And still those chips deliver MUCH higher crunching power than Intel or AMD integrated gfx. But it’s not about any “my Nvidia has a longer one than your apple” here but just about the discrepancy between possible speed on Apple Silicon and what we get.
A M2 Ultra should be somewhere between a 4060 and a 4070 performance wise and mostly is in various apps.
And TVAI also WAS at least on a level or a bit faster than a 4060 in earlier TVAI versions, so this is not just a theoretic assumption, but a known fact. Iris really was flying after a performance improvement Topaz applied somewhere around V3.3 - V3.4, being faster than Proteus and Artemis.
But then since about V3.4.1 Iris performance was seriously crippled after a bug fix that came up in the early Sonoma versions - so that it’s only about 60% of the possible speed now.
Just look at the MacOS benchmarks and especially the Iris 2x performance. Iris was and should be at comparable speed to Proteus and Artemis (and also is on both Nvidia and AMD gfx-cards), but it’s considerably slower.
All newer models seem to be quite low speed-wise as well.
Another thing:
Overall the performance of TVAI is totally “uneven” with sometimes setting RAM to 10% leading to better results than at 100% or in some specific settings TVAI even running faster in low-power mode. Hell, the Nyx 2x was even faster at times than the 1x here.
Also, even only comparing the number of cores between my M1 Pro with the M2 Ultra there’s no real scaling, even much worse than on PC (while other tasks scale very well).
Has anyone tried simply copying the ffmpeg.exe from the 5.3.6 build to the 6.0.2 build to see if this fixes the VFR issue?
3 Likes
I’ve noticed quite a bug with frame interpolation/slow motion - the preview of slowed-down video is still of the same time length, as the original one (7s, but it should be 14s), and also the interpolated frame previews don’t seem to be ‘fully’ viewable in the program currently.
Also, the preview of just interpolated (60fps, instead of 10) preview doesn’t seem to show all the frames of the video, when I scrub it.
I’m testing it on that Steam Deck 10 fps rotating dock rendering video.
1 Like
Crop tool has been broken. Can’t get the “original” or “custom” modes to retain the settings I input. Ridiculous.
1 Like
We need to go back to 5.0 interface. 6 is pretty horrible (both interface & functionality). Go back to 5.0 interface and keep it. It doesn’t need to change anymore. My 2 cents…
7 Likes
I’m still stuck using 5.3.6. It functions much better than the new 6.0 versions.
4 Likes
FWIW this behavior is manifesting while working on a DV file. Yesterday (pre-update) I got a full day’s work out of the software; this morning I’m rolling back to 6.01.
1 Like
Why is it when I use GPU Scheduling the quality of the models show different results with it on/off.
For example, when I use it on nyx 1080p video, the denoise is alright depending on how much it is but using the same video/model with GPU scheduling on changes from medium noise remove to completely removing it all, including any film grain.
Just confused why the model results are different, surely it would be the same but just faster by sharing the workload? Anyone else experience this?
GPU Scheduling can affect the amount of VRAM available to Video AI and cause the app to use a different tile size for processing.
1 Like
