I've been noodling with running a preprocessing step to "tag" pictures and videos with a set of tags with richer spatial and temporal features with off the shelf models and then just let a local AI model pick one based one what might match today's theme.
Are you using any models to make your curation step easier/better UX wise? e.g." Compose all Christmas pictures with Grandma and the kids on vacation" and it would give you a collection to curate from my library