Topaz Video AI Beta v3.2.10.b

I am also interested to know what is that option intended for.
What is Topaz best practice when and when not to use this option (on paper)

So far, I used it to get some detail back when using Iris. It also helps to make upscaled footage look less artificial by blending in some ā€˜dirt/artifacts’ from the original, similar to adding noise. For upscaling DVDs with Iris I have used values between 20%- and 40%.

I really like that feature! It costs some (not too much) performance though.

4 Likes

It seems the ā€œBlend Source Framesā€ only blend in the original (not enhanced) Videoframes. Depending what Slider Setting is used it shows only the enhanced Version or with the Slider set to maxim it only shows the original

1 Like

so in other words if I run the Iris processed video (let’s say I used the value of 50%) on a Video player in slow motion, or frame by frame), I’ll be getting visual shifting between clean fram and dirty unprocessed frame?

Each frame will be a combination of original and processed image. It’s like putting 2 images on top of each other in an image editor where 1 image is semi-transparent so that the resulting image is a combination of both.

2 Likes

I’m currently testing it now!

1 Like

If you tweak the settings themselves you can get some incredibly polished results, I’m testing a 240Ɨ320 video right now, and it captures the face pretty well, (As in it doesn’t turn the face into a gargoyle!) and with the blend addition (Which is bloody brilliant by the way…) really replaces the noise/grain function, while treating them as an augment rather than a substitute for the lost detail with iris.

I’m loving this additional feature and its use with future projects!

3 Likes

so what is the percentage slider? how many frames would have that merger or how translucent (or how much of the source frame data) the source frames overlay would be?

The slider determines the translucency of the source frame. The lower the percentage the less visible will the source frame be in the combined frame.

4 Likes

For example, if 1080/23.976p is set to 2160/59.94p, the Show Export Command would result in the following

-filter_complex tvai_fi=model=chf-3:slowmo=1:rdt=-0.000001:fps=59.94:device=-2:vram=1:instances=1,tvai_up=model=iris-1:scale=0:w=3840:h=2160:preblur=0:noise=0:details=0:halo=0:blur=0:compression=0:blend=0.26:device=-2:vram=1:instances=1

The tvai_fi (frame interpolation) filter is chained with the tvai_up (upscale) filter, and blend is only written in the tvai_up filter.
After increasing fps in chronos, frame blending is performed when upscaling in iris, so the frame blend source is after the intermediate frame is created.

I think it is supposed to blend the source video into a natural-looking image by blending in the source video with the oil painting-like image created by excessive application of the AI filter, but I am skeptical about the effect because I believe it would also bring back artifacts caused by compression.

This is variant I suppose where you increase the FPS and have to choose chronos or fast chronos and then you choose iris and the blending of the source lets say 30 FPS is done on all of the 60 new frames.

which seems a variant on the examples where the source and destination frames are the same

Where is IRIS I am missing it in the bench mark

Yes, it will also bring back artifacts from the source but in my experience, diluted details will usually be emphasized before random artifacts like noise will become visible again (at low source percentage like 20%). That’s probably because the random artifacts will be completely missing from the processed image (when noise reduction is high enough) but real details might only be somewhat diminished. And when blended over, the original details overlapping with the diminished details will emphasize those. There as random noise from the original will not overlap because the noise in the processed frame has been removed and therefore will only be seen when using higher source percentages. I hope that makes sense.

3 Likes

Yes, I understand what you are saying.
In the past, I have tried a similar blending process with a non-AI noise filter, but as the blending ratio increased to recover detail, the artifacts also increased, so I had a hard time finding the right balance.
However, unlike conventional noise filters, AI filters are supposed to learn to remove noise more appropriately, so I felt that they should be adjusted by learning, not by blending detail as in the past.

I believe the artifacts are more predominate visible in Slow motion cases. so if you planning to Slow-motion your video, Iris with blending might not be the best choice.

Can confirm as I’m currently running jobs using the blend slider at 40%

1 Like

Don’t think Iris has been added to the benchmark tool yet.

Thank you for testing the beta and letting us know of the issues. These issues should be fixed in the new build.

4 Likes