Currently disappointed in Video AI - Photo AI does a massive better job at cleaning up faces

I’ve taken a screenshot showing Video AI result (your left) and photo AI (right)
I didn’t upscale the processed the image of video AI, the Video AI seems to only upscale to 4k but doesn’t offer much in terms of clarity.

What happened? why can’t I get the same frame clarity & upscaling with Video AI as Photo AI does.

VideoAI does several frames per second while PhotoAI takes some seconds per frame. So, yes, of course it can give better results.

Besides while it would be possible to just process a video as a sequence of images this would yield better results for the single freeze frames but lose temporal data, so have worse outcoming with moving luminance/chroma noise and in general moving objects.

It’s always some kind of trade-off…

Which AI model and setting were you using in TVAI? :thinking:

Video is not the same as still image.
You can try extracting frames from the video to images, then batch processing all the frames in TPAI, and finally joining them back into a video afterward.

I think I’d prefer to use more time for quality since that’s what everyone is using the product for in some way. At least if you have both products installed and activated that maybe an option in Video AI be created that processes the frames for higher quality and returns them to Video AI. I’d rather not have to break the video down to still images.

I went through all Enhancement model combinations just to see the result. My video output is also set to high. I’ve tried upscaling with enhancement and even work with 4k video resolution (without upscale) that has a low quality capture and tried the models to remove blur and keep details.

It either keeps the same low quality or makes it worse by removing noise (detail) on a person’s face.

I’ve used photos and CG and the result is the same.

As already said such an approach might yield higher quality fro freeze frames, but overall it’s better to process videos as what they are: videos - in order to not loose temporal information which can be worthful for processing noise.
And just breaking a video down to single frames, enhance them one by one without “knowledge” of the previous/next frame would likely often end up in a choppy result, like a flip book.

P.S.: For missing details due to noise removal: Did you try setting the recover Original Details slider to a high value (>80)?