Latest / Big Iron Radio / Homelab Brief - Batch Photo Captioning with Vision LLMs

Transcript

Homelab Brief - Batch Photo Captioning with Vision LLMs · Big Iron Radio

• 0:02
  Welcome to the Homelab Brief from Big Iron.
• 0:06
  Today, captioning a photo library with a local vision LLM and writing those captions where they'll last.
• 0:14
  Quick note, this is an AI voice drawn from a Big Iron guide.
• 0:18
  Links below.
• 0:20
  Why caption at all if you already have ClipSearch?
• 0:23
  Because Clip finds photos by visual similarity, it doesn't explain them.
• 0:28
  Captioning fills that gap.
• 0:29
  It's what gives you accessibility, richer text search, and honestly, succession planning your errors being able to understand the archive.
• 0:38
  And the permanence rule is the same as always.
• 0:41
  Write captions to XMPDC, description for the prose, DC, subject for keywords with ExifTool, not into some app's private database.
• 0:51
  Metadata that travels with the file survives every app migration and outlives you.
• 0:56
  A caption locked in a database your family can't open helps no one.
• 1:01
  On the model, QN2, 5VL7B, is the SANE 2026 default.
• 1:08
  It runs on a 12-gigabyte card and produces solid, factual captions.
• 1:13
  There's a whole Vision LLM picking guide if you want to compare alternatives, but that's the safe starting point.
• 1:19
  Set your time expectations correctly.
• 1:23
  Budget seconds per image, not milliseconds.
• 1:26
  A hundred thousand photo library is three to four days of dedicated GPU time, and longer once it's sharing the card with a Mitch's own machine learning jobs.
• 1:37
  So caption albums you care about, not the entire fire hose, in one pass.
• 1:41
  Last thing, prompt discipline.
• 1:44
  Forbid flowery vocabulary and demand facts, not inferences.
• 1:48
  You want two people at a beach at sunset, not a poem about longing.
• 1:53
  Precise prompts are what make the captions actually searchable.
• 1:57
  The full setups in the show notes.
• 1:59
  That's your brief.
• 2:00
  Ooh.
• 2:00
  It's actually dangerous.
• 2:01
  It's great.
• 2:02
  Thanks for listening.
• 2:03
  There we go.