If youtube is to be believed, thumbnails make a dramatic difference to first-glance engagement. A model trained to find great(by what metrics?) screenshots from the film sounds like a fun project.

That'd take a human with taste a terrifyingly large 5 minutes, whereas AI trash can be entirely programmatic, with maybe an intern picking the best of 4 in a couple seconds. Think of the savings, and with only a moderate impact on the reputation of your service.

This is the baffling part. Maybe it's like a "soft launch" thing to see if it works at all?

Good point. I wonder if Amazon is experimenting with this with the idea that they will be able to make more "eye grabbing" thumbnails with genAI as opposed to pure stills from the given film. The click through rate may be a key metric internally, and since prime video is an add-on to prime the risk of customers dropping the service over this is low (compared to if Netflix did this)

That would require paying a contractor for ~5 minutes of time to skip through the movie until they find a good scene, take the screenshot, trim it, and upload it. Ugh. We've already got all these GPUs over in AWS, just spin up an image generation model and prompt it "Make a thumbnail preview image for the film {{film.name}}", good enough.

Licensing fees. Why pay for individual images when you can just generate them and pay once for the tool.