New Problems. While attempting to de-interlace an SD video the results was a fairly well de-interlaced export imafe. The only real problem with it was the right edge of the imasge has a sawtooth edge. - Does mean the interlaced odd/even interlace lines are different widths, or is ther some kind of mismatch or timing problem when the odd/even lines are reshuffeled during interlace? Secondly, VEAI 3.0 appears to be naming the output image like it’s a prefiew file snd ignoring the. It writes into the origin folder.and ignores the file name entered on the Export As screen.
1 Like
I don’t disagree with some of your points, but the negative and unkind demeanor in your posts turns me off, and I suspect Topaz reacts the same way.
1 Like
Same.
They have said clearly, that you can and should keep 2.6.4 around and it will work for anyone that has bought it in the last year. True it’s annoying and rash that the installer for 3.0.0 removes it, but it was not hard for me to get 2.6.4 back and running.
Should they have kept it in early access for a few more months? Yes. Does management always judge the state of things correctly? No. I blame the manager that demanded there be an Octobar release of the new version.
1 Like
Great version - However I would wish for better options for setting the Nvidia Codecs in a future version - Like a deeper layer of “expert settings” that would be able to stick within the GUI. I can currently only get what I want from command-line utilization.
But it would be nice [and much faster to work with for batch encoding] if the GUI featured more advanced codec NVENC codec settings available where one could specify the codec settings away from the “medium” and “-b:v” settings, so it would be possible to specify something like “-c:v” “hevc_nvenc” “-profile:v” “main10” “-preset” “p7” “-pix_fmt” “p010le” “-cq” “1” “-qmin” “1” “-qmax” “1” or even “-c:v” “hevc_nvenc” “-preset” “lossless” for use with encoding in the GUI (and likewise for the upcoming AV1 encode support)
Thanks 
3 Likes
Version 3.0.0 didn’t remove 2.6.4 for me (Windows 10). It gave me an option to remove version 2 but I declined it.
3 Likes
I think I have figured it out, but am waiting on confirmation. 3.0.0 always sets the scale of the model to 0. This results in a model for a bigger resolution being selected.
I’m also seeing much slower performance on my Apple M1 Mac Mini compared to 2.6.4
To render a 5 second preview it was taking 5-7 min, was nowhere near that slow previously. Looks like .5 frames per second. I was seeing half that speed or faster in 2.6.4
I’ll check the encoder settings as ida suggested and see if that improves.
Same with Windows 11. Version 2.6.4 gets only removed when that option is explicitly selected.
1 Like
Where do I find the direct frame number input on the timeline as in 2.6.4 and earlier versions, that was promised for the release-version of v3?
There is also no option to set a specific input pixel aspect ratio for all formats, only original pixel type and square pixels, and only for formats that already carry pixel aspect ratio metadata within the files.
Please add an option to manually interpret the footage (framerate, field order, pixel aspect ratio, etc.) at one single place, like in any other professional video software… this can also be used throughout the program (so there is no need to define the input field order again at the enhance page, etc.).
Or even better: Some presets for common aspect ratios and the option to enter a custom pixel aspect ratio manually if needed.
I already posted the most common pixel aspect ratios in the beta thread.
So there will be no longer a need to resize input videos with non-square pixel aspect ratios in another software from rectangular to square pixels beforehand to get a correct scaling result.
1 Like
Hi,
The problem isn’t with Avisynth’s license itself, but rather with FFmpeg. Back in version 3 the FFmpeg developers made the decision that enabling Avisynth support would require enabling GPL code within FFmpeg, as they didn’t wish to use the linking exception in Avisynth’s license.
Unfortunately we cannot distribute a version of FFmpeg with GPL codecs enabled. We’ve reached out to some of the more popular FFmpeg build scripts to see if they’d be willing to have support for our filters as a solution, but it may be a while before that’s ready. Once they’re done, however, you should be able to replace the bundled copy with one with your desired features.
In the meantime, for anyone feeling adventurous, the redistributables for our filters are available on our GitHub.
5 Likes
nowhere. 
I do understand that, but unfortunately this remains a dealbreaker for me moving to 3.0. So if I’m understanding you correctly, it’s as simple as recompiling ffmpeg with the patch and replacing the one in the install directory? Because I can absolutely do that if that’s the case.
Perhaps you could close the feature requests that are now part of this release like video stabilization.
You should still think about this:
this version 3 is a step forward in many respects, but 2 steps backward in many others.
Using FFMpeg as a base is definitely the right way to go, but I agree that adding support for Avisynth should already be one of your priorities for future updates (and maybe adding bitrate selection for converting audio tracks and adding the QTGMC deinterlacing method).
I also add a few points for urgent improvement:
- The handling of deinterlaced video streams.
- The conversion of final files (when I open a cut video inside TVAI 3 and play it with VLC example, the video track is completely inaccurate, a symptom of a badly done conversion).
- The quality of upscaling models (worsened slightly compared to version 2.6.4) and especially the speed of video conversion: although now the program seems to take much better advantage of high-end GPUs such as the Nvidia RTX 3000 series, the difference compared to TVEAI 2.6.4 is still minimal… a symptom that there is something wrong at the base of the PC’s resource exploitation.
Hopefully you can fix these critical issues as soon as possible, but of course I am hopeful!
5 Likes
There is no inverse telecine function and I would like support for this.
Interlaced material is broadly classified into two types.
One is material that was originally shot in interlace.
The other is progressive (24p) converted to interlaced (30i) by telecine.
In the former case, each field of 30i can be mixed to produce 30p, or each field of 30i can be expanded to produce 60p.
I think Dione’s method is to make it 60p.
In the second case, however, it is best to use inverse telecine to convert 30i back to 24p in terms of image quality.
If using ffmpeg, inverse telecine can be done by using yadif or fieldmatch.
ex.
fieldmatch=order=auto:combmatch=none,yadif=0:-1:1,decimate,setpts=N/(24000/1001)/TB
1080i(30i) → inverse telecice → 1080p(24p) → Proteus 2x → 2160p(24p)
-filter_complex fieldmatch=order=auto:combmatch=none,yadif=0:-1:1,decimate,setpts=N/(24000/1001)/TB,veai_up=device=-2:model=prob-3:scale=2:compression=0.24:details=0.24:blur=0:noise=0.12:halo=0:preblur=0,scale=out_color_matrix=bt709
2 Likes
Since the internal colorspace of VAI is RGB, the alpha channel is lost when converting from YUVA to RGB.
I think it would be possible to temporarily separate the alpha channel using the FFMPEG functions alphaextract and alphamerge, process it with the VAI filter, and then rejoin the alpha channels.
4 Likes
Just noticed that they removed not only Avisynth support. Now you are unable to open files with UNC path. For example if you keep all files on NAS, 3.0 don’t want to open files directly from path like
\\Share\Public\Video\something.avi
now you need to copy file to HDD before opening
[macOS Monterey 12.6, iMac 2020]
I have tried numerous combinations of denoise, sharpen, and upscale to 4K on an old 250 lines of horizontal resolution VCR videotape recorded using a vidicon tube camera, but there were always parts that the AI handled poorly.
So I tried this two-program / two-step process using 2.6.4 Video Enhance AI:
- Video Quality = Low
- Video Type = Progressive
- Video Artifact Type = High Compression
- Output Size = 100% (Denoise / DeBlock)
- Video Format = MOV - ProRes 422 HQ)
…then upscale the result to 4K using Video AI 3.0.0:
Model: Upscale to 4K
Video In 1280x720; Video Out 3840x2160 (4K)
Enhancement: Video Type - Progressive; Parameters = Auto
Encoder: H265 Main
Results: I get the best possible denoised, sharpened, upscaled 4K without weird artifacts. It’s literally the best process for these old 250 lines of horizontal resolution that were previously deinterlaced and saved as progressive SDR.
I previously tried every possible combination using v3.0.0.0, and although some modes are amazing, I can never get through the entire video without unwanted glaring artifacts from the AI process. The new v3.0.0 should be able to do this so I don’t have to use the old and new versions in a two-step process.
How can I help Topaz make the new program better?
1 Like
I second the motion. I would use multiple start/end markers to extract segments from a longer video. The inability to do this with 3.0.0 makes it less useful for many of my applications.
2 Likes
