Have you tried out Gaia CG for low quality SD sources?
I did some comparisons here to demonstrate the broken DTV-2 preset. You can see, that DTV-1 is much better than DTV-2, which adds a haze like blur effect and does look really band. The other two are using a 50p pre-deinterlaced file and AHQ-11 as well as GHG-5. Here you can see that 50p-GHQ-5 is also better than 50p-AHQ-11.
i still cannot upload images, donât know why. Can anyone help ?
I found that DTV-2 (and AMQ) only produces hazy results if the source footage got too much in-camera sharpening while upscaling to HD. It seems to try a dehalo of those coring lines around edges but sometimes fails to find the right size and then increases the radius too much. What I do is a first pass with DTV-2 at 100% size which almost always works as intended.
Then I use AMQ to upscale in a second pass which works very well then.
I kept all the templates from version 1.8.1 to version 2.
That way, I can test each model and see which one is best for which type of video.
2 Likes
Hi Bluewave, thanks for mentioning this. Iâve found that out myself when playing around with the settings in depth. When upscaling from SD to HD like 576 to 1080 in height it works as intended because this below 200%
When upscaling from 720 to 1920 it gets blurry because this is more than 200%
So thanks for confirming this.
1 Like
Thanks. Maybe I meant GaiaCG. I recall reading that it had been reduced/worsened even though the version number was the same. So, if it is the same in binary check it canât be a different version?
you need to click on the image to get this visible.
3 Likes
For upscaling DVD sources, I get the best result with Artemis LQ v8. Every version of Artemis after v8 just isnât as good. It appears, that starting with Artemis v9 and above, Artemis LQ v8 was âmovedâ to Artemis HQ for v9, v10, and v11. But those results arenât as good. Some very low quality sources are definitely better with Artemis LQ v10, or Artemis MQ v11 if it has oversharpening halos⊠Artemis LQ v11âs dehalo feature just makes the whole image too blurry/smokey/dreamy looking.
2 Likes
Theyâre models for different AI engines.
ml = Apple Core ML
ox = Open Neural Network Exchange (used for GPU upscaling)
ov = Intel OpenVINO (used for CPU upscaling)
Here is another example how terrible DTV-2 looks like:
which files need to be copied from the previous versions to get these included ?
Uploaded 2 more files, Matters of Honor and Messages from Earth (extracted scenes using Avidemux). Dione Interlaced v1 and v2 (dtd-1 and dtd-2) produce the same issue during horizontal panning scenes. Some sections of the scene appear âdelayedâ during the pan producing some kind of smearing effect.
I had made a backup of my models before each update of the software because at each update, it removes the contents of the folder to place those of the new version.
Normally, they are placed in the folder âC:\Users\user name\AppData\Roaming\Topaz Labs LLC\Video Enhance AI modelsâ.
Me, I put them in âD:\Video Enhance AI modelsâ (an external disk to save space on disk âCâ).
Same question, new user to the software. Is there a way to go back and get previous model versions?
Only if you have made a backup of the models folder of the old versions.
Models from 1.7.0 I think can still work in 2.0.0. But more for 1.6.1 models.
thanks, that worked, great.
Youâre welcome
1- new container MKV ( its the best its lossless quality even better than MP4 )
The container has nothing to do with lossy/lossless. However, using MKV instead of MOV/MP4 makes it possible to recover partial jobs if a crash happens. Currently you can only resume crashed encodes for image sequences.
2- new codec H.265
Forget it. H.265 (HEVC) is stuck in patent and licensing hell and even big companies avoid it. Just set VEAI to encode to an intermediate format like ProRes or 16bit TIFF and use freely available tools to convert to HEVC.
4- Adding 10 Bit Color Depth
VEAI already supports 10 bit colour depth (ProRes) and even 16 bit colour depth (TIFF). The models do a great job debanding 8 bit sources.
1 Like
Test with Dione Interlaced V2: SD to HD.
Dione-TD (the one that exports at half the frame rate, e.g. 25 instead of 50) is only useful for cleaning residual interlacing on otherwise progressive sources. As your example shows, it creates a Kronenburgian amalgam of the top and bottom field with no regard for motion or temporal coherence. It shouldnât be used on true interlaced content.
See what it does on a scene change:
There are random elements of both fields mixed in the final image. Additionally, it seems that at least in the current version, it uses naive linear interpolation for fields instead of proper edge directed or ML-inferred interpolation (see the stairstep artefacts on the railing at the bottom right).
1 Like
Unfortunately VEAI does not allow yet to set the aspect ratio freely.
If you remux your source to MKV using MKVToolNix you can override the aspect ratio using the MKV metadata and VEAI will correctly use the correct aspect ratio. I agree it would be more useful to be able to override it in VEAIâs UI, but itâs not a big deal currently.
Additionally, I donât know if overriding the aspect ratio actually causes a different model scaling to be used (or oversampled), or if VEAI just relies on libavcodec to prescale the aspect ratio prior to upscaling. I intend to do a moirĂ© test with vertical bars to verify this, unless a Topaz employee could tell us which would save me the time of finding out myself.
EDIT:
I decided to do the experiment as it was simple enough. I created a 480x480 video file that contained alternating black and white vertical bars of 1 pixel wide. I then remuxed that file to MKV with an aspect ratio of 4:3, yielding a corrected input resolution of 640x480, which shows moiré as expected.
Upscaling the 1:1 aspect ratio file to 960x720 (200% scale, cropping off the top and bottom) yielded a consistent pattern of doubled pixels, no moiré.
Upscaling the 4:3 aspect ratio file to 960x720 (150% scale) yielded the same pattern and no moiré. This indicates that VEAI correctly scales the AI model to match the output aspect ratio without prescaling the input.
2 Likes
