Topaz Video AI Alpha v3.4.0.1.a

Topaz Video AI Video AI Beta
1

Hi All,

Here is an alpha build with some major changes to the processing backend.
There are NO GUI CHANGES FROM 3.3.9, so please only report issues related to processing.

Downloads: Win | Mac | Linux

Changelog from 3.4.0.0.a

  • Fixes Iris not working on windows
  • Fixes estimate button not working on windows

Downloads: Win | Mac | Linux

Changelog:

  • Quality Improvement for first few frames while using enhancement models (Except Gaia).
  • Speed up for Iris on Mac Ventura
  • Fixes first few frames repeating in frame interpolation
  • Login now uses proxy settings
  • Minimum number of frames needed for processing is 4
  • Processing on Intel CPU machines (12th, 13th gen) no longer slowdowns when app is minimized or out of focus
  • Fixes image sequence issues when non ASCII characters are in the path
  • Potential Fixes
    • ANE process using up all Memory on Mac
    • Processing failures when exporting large number of frames
    • Intel Arc/iGPU processing issues

Known Issues

  • On Mac Iris slower on higher resolution output
  • Intel ARC GPUs may get black screen while processing

Please let us know if you are still having any processing related issues.
Post logs/problems videos here
Thanks for testing

7 Likes
3

Please don’t shoot the messenger: estimate settings for Proteus and Iris doesn’t seem to work (works in 3.3.9).

1 Like
4

Well there is a reason it is an alpha, will check on Windows. Seems to work fine on the Mac
Finding bigger issues on Mac for Iris :slight_smile:

1 Like
5

For me (RTX 3090, Win10 22H2, Studio Driver 536.67) Iris model doesn’t work at all, no matter the mode (Auto/Semi Auto or manual in GPU mode) so I’ll upload logs to DropBox. :slight_smile:

6

Preliminary first short test:

YES! Finally this seems fixed.

And another YES. Speed is considerably up here using Iris on Apple Silicon, from about 12 fps up to 20 fps for SD → FHD upscales.

Upscaling SD content still is much faster using 10% RAM than with 100%
Here is the time needed for a 15 secs preview on the Mac Studio M2 Ultra with this new Alpha version (upper preview with RAM at 100%, lower with 10%):

Bildschirmfoto 2023-08-02 um 23.17.42
Bildschirmfoto 2023-08-02 um 23.17.422826Ă—252 47.9 KB

And here for comparison the same task on the exact same file with the 3.3.9.1.b:

Bildschirmfoto 2023-08-02 um 23.24.29
Bildschirmfoto 2023-08-02 um 23.24.292538Ă—170 33.3 KB

Here the lower preview is with RAM set to 100%. As you can see the RAM filling bug absolutely killed preview times with that in older versions.

Also, live Previews do seem to work again even for 1080p → 4K upscales.

All in all a BIG thumbs up for this version :+1: :+1: :+1:
Way to go.

2 Likes
7

OK, one not so good thing:
1080p → 4k Upscales with Iris unfortunately are dramatically slower with this version.

This is with the new alpha:

Bildschirmfoto 2023-08-02 um 23.45.02
Bildschirmfoto 2023-08-02 um 23.45.021270Ă—96 15 KB

And the exact same task with the 3.3.9.b.1:

Bildschirmfoto 2023-08-02 um 23.41.07
Bildschirmfoto 2023-08-02 um 23.41.071296Ă—106 12.8 KB

So, since the 3.4.0.0.a is much faster for SD upscales but much slower for HD: maybe you could mix the code using the new for SD tasks and the old one for HD upscales?

8

I agree on the 10% vs 100% VRAM slider situation, especially when using lower resolution input. This is due to the multithreading that happens when the input is split up. If you do not see any quality issues it should be fine. Quality issues can happen on videos with fast motion.

We are studying quality to figure out if we can change the size selection accordingly.

1 Like
9

We are noticing the slowness as well and looking into it. Added it to the known issues for now.

4 Likes
10

The lower speed seems to affect all Iris upscales to anything higher than FullHD. So upscaling SD to 4k or 8k is also considerably slower in the new alpha than it was before.

Still, my most used case is SD->FHD and there this new alpha shines.

But integrating both codes from the new and the old version and chose the faster one based on the output resolution would be an idea. Unless you find a bug that affects rendering speed with the higher resolutions in the new code and are able to speed this up as well, of course :slightly_smiling_face:

11

I couldn’t see any quality differences between setting RAM to 10% and 100% up to now.

But somehow Iris seems to behave quite differently with different output sizes, e.g. in 4k it seems to do things differently (apart from just having the higher resolution) than in 1080p. Same goes for 8k vs. 4k.

But this isn’t specific to this Alpha.

12

Iris uses different models for 1x, 2x and 4x upscale. Seems like the 4x model is running slower.
Since they are different models the quality also varies a bit between them.

3 Likes
13

Understood, but I was on the fence to report it - thinking to myself: “Is it GUI related? Maybe I shouldn’t report it…” then thought I better since you stated nothing on the GUI has changed, so it must be related to processing back end changes.

14

Hoping this pertains to the dupe’d frames from the interpolation issue, I ask because I hate relying on assumptions… Will be enthusiastically testing for this! :smiley:

3 Likes
15

This should be fixed as well, let me know if you still have issues.

1 Like
16

Will do, I’ll also upload the test videos as well as my logs!

1 Like
17

Just report any issue that are not obviously GUI. When in doubt just report it and we will figure it out.

3 Likes
18

There’s a noticeable tiling while using Apollo Fast v2 and…

image
image1920Ă—1046 75.1 KB

…Apollo v8 models:

image
image1920Ă—1047 74.5 KB

image
image1920Ă—1050 83.7 KB

1 Like
19

I wish I had a better answer.
The only solution for this issue is to set vram to max and run Apollo. This might work for 720p input but not work for higher resolution.
The other option is to downscale run Apollo and upscale back.

20

@mikmod1 can you upload the video.

1 Like