I wonder how viable this is for video and image generation models.

Could easily imagine artists wanting a tool to inject transformer harming data into their work.