Topaz Video AI Beta 3.2.1.0.b

AlbertTopaz · April 7, 2023, 10:21pm

Hello everyone!

First, we have released the new Docs page for both Photo and Video. This was a huge effort from our Support team so please take a look!

Topaz Labs Docs

Here is a new beta available for testing.

Downloads
Windows
Mac (DMG) | Mac (PKG)

Changelog

Toggle in preferences for supporting model downloads when behind a proxy
Presets has option to keep export settings
Create/Edit Preset Dialog responds to “Enter” key input
App Window appearance and scaling fixes
Fixes issues with models not loading/working on RTX 20 series
ProRes standard output added
Adjustments to top menu bar options
Added “Try Topaz Photo AI” to top menu bar
Added additional logging details
Improved synchronization between input and output playback
Increased speed and accuracy for playback

Proteus V4 Beta:

A new model for general enhancement of medium and low-quality video. A potential alternative to Proteus V3 Auto.
The model is trained to work slightly better on human faces than the existing Proteus Auto, especially on small to medium-sized faces.

Known issues:

Currently, Proteus V4 Beta only has a 4x upscaling option. Regardless of the set output resolution or scale, the model will 4x upscale it and rescale it to the desired resolution if needed.
Input/Output greater than 8K when previewed crashes the app

Thanks for testing!

Please upload problem videos and logs here: Submit files

20rushtonj · April 7, 2023, 11:40pm

This should be added:

Proteus V4 Alpha only offers a ONNX model and a MLCore model. Users with RTX GPUs may see lower than expected performance.

Or

Proteus V4 Alpha does not offer a TensorRT modes. Users with RTX GPUs may see lower than expected performance.

Also, it doesn’t matter, but Alpha is used in the GUI and Beta is used in the release notes.

Quick feedback on the new Proteus v4

In some simple tests, 4x scale, full auto, Proteus v4 seems to be better at dealing with noise without introducing as many artifacts as v3.

And it does a much better job at hiding compression artifacts. But it’s not as sharp/oversharpened (it can be argued to be a good or bad thing), and as a result “fine details” may look worse compared to v3.

Also, this is just a personal thing, I feel like the version number should be included in the GUI in all locations.

For example, Proteus should be called Proteus v3 and that should be represented in the model selector and the benchmark results in the GUI.

I specifically bring up the benchmark results as people like to compare previous versions of TVAI to the new one. And if the model version has changed between TVAI versions, but it’s still got the same name between updates. Then people comparing performance between versions might be confused by the results.

Also, there was a point during the alpha that Apollo, as picked from the GUI, would use Apollo 8, while the benchmark used Apollo 6. And that could be confusing when the performance suggested by the benchmark isn’t as close to the real world performance as they thought.

david.123 · April 8, 2023, 12:36am

Thanks, that explains the lowly 2fps I was getting with a 4090 card. Will wait until they have an RTX specific model to test, this is way too slow.

20rushtonj · April 8, 2023, 1:38am

Proteus v4 doesn’t seem to handle blurry/compressed carpet as well as v3. v4 makes it all blobby.

Note: The camera is stationary and carpet are stationary, so the carpet does not move through out the video.

Setttings: 4x upscale, full auto.

Note about photos:

This is at 100% scale.
Original video on the left, upscaled on the right.

Proteus v3

Proteus v4

Sorry, I can not share the original video.

ssbroly · April 8, 2023, 2:30am

this version of proteus looks not bad but on my video of madonna in sd, just with proteus, it’s at 9 spf… extremely slow and after 3 minutes, it puts the preview in error. So if I understood correctly, it increases the resolution to x4 and decreases it to 480p? even if I select x4, it’s still fine at 2880 x 2304, but it’s just as slow. my graphics card is a gtx 1660.

dimar102 · April 8, 2023, 4:52am

I’m having interesting results with animation when using Proteus V4 Alpha, but the manual dials are unpredictable. Things look nice and sharp, but a lot of times it introduces additional tiny drawings that don’t belong to the image.

godefroydemaupeou-68 · April 8, 2023, 6:39am

This version of VAIE crashes on stabilization. this is not the case with the previous one.

2023-04-08-08-33-3-Main.tzlog (60.9 KB)

osol · April 8, 2023, 6:10pm

Version 4 is way too aggressive. Worst cartoon effect yet. (It might be good for cartoons thought)

mikmod1 · April 8, 2023, 6:39pm

For cartoons, it is great, this destroyed D&D animation looks amazing:

Now, to test it on SG-1 series (480p).

mikmod1 · April 8, 2023, 6:51pm

SG-1 shows a little degradation on details on the walls, but that’s just Auto, I’ll try increasing details next. But, wow:

Yeah, wall details get destroyed a bit:

Wow, the faces here, it’s like Face Restoration applied to a movie:

…and since first episodes were badly blurred for some reason, this is going to be fun!

Imo · April 8, 2023, 7:45pm

I had a quick look into the preview and the new model looks quite good. On some videos it lightens up the video, colors look a bit washed out and minor details get lost. For a longer test this version is too slow yet. at normal speed this movie is done in around 18 hours:

Rayburner · April 8, 2023, 8:07pm

Hi,

for me the model downloading issue with my RTX 2080 Super is mostly solved (for the 3.2.0 release version it doesn’t download models at all).

Models that still don’t download are for Chronos, Apollo, Artemis HQ and Proteus V4

On the other hand on my RTX 3070 TI on a different system all models work fine.
I am pretty impressed with the result of the Proteus V4 model. The faces look great compared to V3 even at a resolution of 4k.

looking forward to find out what else Proteus V4 can do.

Imo · April 9, 2023, 9:45am

may I ask why the app was not compiled with the option to make use of the tensor cores? testing the product at that “speed” - well lets say slowness - is almost impossible. your free beta testers spend a lot of their rare spare time to help you inproving the product, but at speeds so slow that you can’t do anything much beside getting 0,1 sec long preview images which makes a proper testing impossible. and this is not an alpha version anymore.

robin.koehler · April 9, 2023, 10:24am

V4 produces very noticeable banding in my test footage independent from the output codec chosen. Much worse than V3 in the same scenes.

V4 scrubs out a LOT of detail, and it does not seem to handle depth of field consistently.

The overall image quality is very high, it does a much better job restoring a cohesive image from smaller detail (and faces as advertised).

The colour space also doesn’t seem to be correct. Red and blue especially shift substantially, and near-black colour gets crushed completely.

meimeiriver · April 9, 2023, 11:40pm

Odd. RTX 3080 Ti here, but seeing no noticeable slowdown (and certainly no drop to 2 fps). It’s around 6-7 fps on ‘Auto’, and close to 10 fps on manual (upscale to 4K).

UHD4KUser · April 10, 2023, 5:56am

Noticed with v4, it can create these strange “blotches”. I saw similar with v3 when sized to x4, where x2 had no such issues. The “blotched” areas aren’t with compression artifacts, so I don’t understand why it would create such a strange effect.

Original left/v4 right

nipun.nath · April 10, 2023, 5:38pm

Hello everyone, thanks for testing the Proteus v4 model despite its slowness. The model is significantly slower than v3 because:

On NVIDIA, it doesn’t use TensorRT.
On Windows, it operates on FP32 instead of FP16.
It has only one block size which might not be optimal for the given input resolution.
It upscales to 4x regardless of the given output resolution. Then resizes it back to the final size if required.

All of these will be optimized before the final release. Right now, we are focusing on improving the quality, especially the details that some of you mentioned.

Thanks for your valuable feedback.

Imo · April 10, 2023, 6:32pm

Hello nipun,

thanks for the info. The incredible low processing speed makes it almost impossible to test this beta properly. Please offer an updated version soon if possible.

Thanks, Imo

david.counts · April 11, 2023, 2:51am

What settings are you using? Auto? Manual? And how are you viewing the results? Thanks!