Which local model through Ollama is actually decent at reading spines off a bookshelf photo?

Replies

1. K

   Kat H. @kat-hLinguistics grad student · uses Python and Ollama

   The general vision models are the wrong tool for this. Run a real OCR model first (I use one off Hugging Face, the PaddleOCR-style ones handle rotated text) to get the raw strings, then hand the strings to the language model to clean up into titles. Two steps, way fewer hallucinations.

   Sep 2Isaac B., lars and Gia F. agree
2. L

   lars @larsuses Ollama

   Also rotate the image 90 degrees and run it again. Spines are mostly vertical text and half the models only read horizontal. Take the union of both passes.

   Sep 3Isaac B. and Kat H. agree
3. G

   Gia F. @gia-fPhotographer · uses Python

   The bigger vision models do fine if you crop tight. Split the shelf into strips of five or six spines and send each one instead of the whole shelf.

   Sep 3Isaac B. agrees
4. I

   Isaac B. @isaac-buses Python and Ollama

   Trying all three. Rotate-and-union alone took it from one in five to about one in eight. The OCR-first pipeline is next.

   Sep 5Kat H. agrees