I would do something like this, should work for any type of image.

Original -> Upscale and Downscale.

Alter the frame sizes ie 4:3, 1080p, 4K

Encode all video formats in all video codex's.

Make stills for timelines which are likely to be interesting.

Use a variety of background filters like those seen in Zoom calls to obscure backgrounds and then use all the stills with and without backgrounds, and also alter the colouring/whitebalance.

Hash Stills and you will have loads of stills!

Train an AI to look for the stills allowing for colour changes like a slight filter to alter the colouring, or maybe do it in black and white.

When still is suspected, hash the AI found stills against known stills

Obviously different websites allow uploading of different formats/codexes which is a filter in itself and frame sizes

Because of editing, an AI needs to scan the film, when a still is found, make a copy and hash it to see if there is a match.

At least thats how I would do it.

I think you can spend too much time looking at other stuff AI & training data and sometimes its the background objects which are more important, like clothing, jewellery & furniture, but these new background filters seen on Zoom calls also makes it harder for the AI stills recognition.

Expensive in terms of cpu processing, but probably a better way to spend money than on human wages going through stuff/moderating.