Being the author of slidestr.net AND posting some NSFW content on my photography profile, I know this issue from both angles.

Public web clients need some kind of default filter. Using ML seems to be a good approach but we must also accept that this will never be solved perfectly. In a year from now we might discuss what is acceptable or find users try to game these systems. With more video content the classification will also become more complicated and costly.

IMHO filter choices should be part of an onboarding process where a user selects their interests and could also decide on a NSFW filter explicitly. It would be better to mark/blur content than to hide accounts completely without the user knowing. If there ever is a „suggested accounts for #photography“ list, I‘d like to be on it and not be omitted by some default behavior.