Topaz Video AI Beta v3.5.0.0.b

[jacob.muchow]

Hello everyone,

We have another beta available for your testing.
3.5.0.0.b

• Windows
• Mac - DMG | PKG

Changelog from 3.4.4:

• Final updates for Iris v2 and Nyx v1 before release.
• Adds support for rotating videos 90, 180, and 270 degrees.
• Adds crash logs to “logs for support” and changes format to zip file.
• Fixes bug causing logs to be inaccessible on Windows.
• Fixes special characters in file path causing export failure when using Export As.
• Fixes crash on app open for some machines.
• Fixes trim accuracy issues for ProRes.
• Fixes bug causing inaccuracy with video duration after resetting trim range.

Attention: The format for support logs is being changed to include license information (auth.tpz). While this file is encrypted and only useful on your machine, please try to avoid posting it or your support logs here by mistake. This would not likely lead to a problem, but better to be safe than sorry. Please submit any logs directly to our support team or devs.

Nyx Beta:

• A new model for denoising high-quality video.
• The focus of the model is on footage from high-quality cameras captured in less-than-ideal conditions.

Nyx Beta Updates:

• Improved quality.
• TensorRT models for more tile sizes.

Iris v2 Beta:

• Iris-v2 offers more details and sharpness.
• It uses TensorRT models for Nvidia GPUs and could result in a 50% speed-up.
• It also reduces the “previous frame’s ghost” artifacts.
• It reduces “marching ant” (diagonal movement of texture) artifacts, especially, for 1x and 2x.
• Try to play with the parameters if the “Auto” doesn’t give you the desired result.

Known Issues:

• Intel ARC is currently unsupported.
• Videos with mismatched metadata and streams will display incorrect duration
• Frame number preview length may shorten on app restart

---

3.5.0.1.b

• Windows
• Mac - DMG | PKG

v3.5.0.1.b Changelog

• Fixes preview mismatch / preview not generating at correct playhead location.
• Fixes regression to 2nd Enhancement causing worse results.
• Fixes beta auto-update notification.
• Adds some helpful messages to UI for Nyx/Iris when enabled.
• Fixes presets to respect Iris version.
• Fixes Iris version selector appearing a second time for Interlaced/Interlaced-Progressive videos.
• Fixes Iris V2 label for Interlaced videos.
• Moves legacy models preference to Advanced section of prefs.
• Renames “Second Pass Enhancement” to just “Second Enhancement”.
• Fixes changing “input fps” for any video accidentally updating your preference.

---

Please upload problem videos and logs here: Submit Files.

[20rushtonj (Alaska)]

Nyx still only supports 1x enhancement (as expected), yet when you select 2x upscale, then select Nyx, the video will be denoised with Nyx, then upscale with a simple upscaler providing less than ideal quality.

I feel like something should be done about this. Maybe a warning in the UI letting users know that Nyx only supports 1x? Maybe when Nyx is selected, and the scale is above 1x, a second pass is automatically added?

Maybe a new “Denoise” panel should be added and Nyx should be shifted out of the Enhance section and into there. That way people can denoise then upscale without having knowledge about how Nyx operates and or need to enable the two pass system.

---

I also wanted to confirm. Does Nyx “Auto” settings still not work well? It was mentioned as a known issue for a while, but has since been removed, but none of the release notes mention that Nyx “Auto” improved. So I’m just double checking.

[20rushtonj (Alaska)]

Two pass doesn’t work properly or recent changes are detrimental.

In older versions of TVAI the two pass system worked like this:

1. Process first pass.
2. Process second pass.
3. Use Lanczos scaler to bring video to final resolution.

Now the process works like this.

1. Use Lanczos scaler to bring video to final resolution.
2. Process first pass.
3. Process second pass.
4. Use Lanczos scaler to bring video to final resolution.

This can result in a reduction in quality and significant increase in processing time and memory usage.

Note: I was using the macOS version of the app.

[suraj]

Interesting, let me check on that. If your observation is correct this needs to be fixed by release.

[20rushtonj (Alaska)]

These observations are based on the commands in the “Show Export Commands” option.

"-filter_complex" "scale=w=3840:h=2160:flags=lanczos:threads=0,tvai_up=model=ahq-12:scale=1:blend=0.2:device=0:vram=1:instances=1,tvai_up=model=ghq-5:scale=0:w=3840:h=2160:device=0:vram=1:instances=1,scale=w=3840:h=2160:flags=lanczos:threads=0"

With these settings:

[20rushtonj (Alaska)]

TVAI also seems to save both the temp and final versions of the video at the same time if the “Include Live Preview” option is ticked in the export settings.

Here’s a short video showing it off in the export commands: https://youtu.be/M8K7EMHSRzM

[alanhusband-677420]

In this version the preview generated is always from somewhere before the position of the marker on the timeline, which then leads to the preview window mismatching the original window. Obviously this doesn’t occur if you preview from the beginning of the video.

Will upload logs to OP Dropbox.

[Imo]

The difference between the input and ouput previews got even worse with this beta. Now the input and output preview frame are not even close anymore.

[Geekoo]

I subscribe, that’s exactly what you’re saying. I believe it’s a bug and not a working version at all. No matter what resolution option you choose for input or how you scale it, only the 1x model files are being downloaded, so the output lacks quality.

[TPX (Thomas D.)]

@jacob.muchow

Nyx Beta Updates:

• Improved quality.
• TensorRT models for more tile sizes.

It does load only the “nyx-v1-fgnet-fp16-1088x1024-1x-rt705-8500” tile size.

There is no speedup with a Quadro RTX 5000.

The Quality is better.

---

---

@Imo

How much does the performance decrease with the 4090 when you process three or four 4K videos with Nyx at the same time?

My Quadro RTX 5000 was beaten by my Radeon although they are equally fast in single videos.

---

---

Radeon Pro W6800

Quadro RTX 5000

The RTX 5000 does need much less memory, maybe due to fp16 model, but i dont know it the model before was fp32.

[topaz257 (Michael Karg)]

Impossible to state without CPU, RAM, tvai model, resolutions info. 1 task might completele utilize the 4090 (no upscaling, model with good Nvidia support, very fast CPU + RAM), or just by 25% (x4 upscaling with slower CPU+RAM, model with not so great Nvidia support)

[TPX (Thomas D.)]

Its about GPU multitasking.

How good is the 4090 at handling 2 - 3 - 4 4K processes at the same time.

As a Professional Cinematographer i would cut up to 6 short videos and process all of them with TVAI.

But not sequential but in Parallel, so the GPU needs to keep up.

[Imo]

@TPX

processing 1 * 4 k video @ 2,7 fps
processing 2 * 4 k video @ 1,8 fps and 1,0 fps
processing 3 * 4 k video failes with errors on all three.

all test done with Nyx on Auto.

[TPX (Thomas D.)]

Wohoo, thank you so much, i need to buy you a coffee.

What does happen if you set the settings manual?

I think i read anywhere that the CPU does the estimation for Automatic settings.

[TPX (Thomas D.)]

Starting up 4 processes.

Full Load

The 4090 would still be faster even with less processes.

Sys Specs

Topaz Video AI Beta  v3.5.0.0.b
System Information
OS: Windows v11.22
CPU: AMD Ryzen 9 7950X 16-Core Processor              63.727 GB
GPU: NVIDIA Quadro RTX 5000  14.795 GB

[Imo]

Here are the results with “relative to auto” and your settings. I once noticed that turning off “recover original detail” may speed up the processing, but not always.

one processing

two processings

[TPX (Thomas D.)]

Interesting how it speeds up.

I assume that the system is still perfectly usable and doesn’t start to stutter.

---

---

Is there anyone here who could test Nyx with a Radeon 7900 XTX, with two three four 4K processes.

I would be interested to see how this GPU handles it.

---

---

I let my system test run with 4 processes for 24 hours.

[TicoRodriguez]

Test image with FFMPEG’s built-in testsrc filter.

Nyx has less color alteration.
For example, the red color changes from R:G:B=255:0:0 to R:G:B=254:0:0, which is within the margin of error.

Iris v2 has some strange color transformations.
For example, red goes from R:G:B=255:0:0 to R:G:B=238:14:11, which is fundamentally wrong.

If the color balance is made absurd, it will affect the grading process, so please consider an option that does not change the colors.
Other models can have less color alteration, so why can’t Iris do it as well?

FFMPEG testsrc
FFMPEG testsrc & nyx v1

FFMPEG testsrc & FFMPEG scale x4
FFMPEG testsrc & Iris v2 x4
FFMPEG testsrc & GaiaCG(GCG) x4