Visual recognition · the Antilibrary

How well AI reads bookshelf images

Three experiments on turning a photo of book spines into a catalogue — and what we learned trying.

Small experiments, shared to show the method. A few shelves, run locally and in the cloud. Treat the directions as real and the decimals as indicative — not a verdict on any model.

Building an antilibrary starts with knowing what you own. The fastest way to catalogue a shelf is to photograph it — so we asked whether a model can read the books back out of the picture. We came at it three ways: many small local models reading one messy bookcase; a frontier model tested with and without an OCR pre-step; and, on the opposite extreme, the best models handed the easy case of neat shelves and sharp photos. Different setups, so read them as three probes pointing the same way, not one scoreboard.

What we found

  1. It works — but the photograph matters more than the books. A tidy, straight-on, legible shelf reads cleanly, even with repeated titles or look-alike covers. A crowded or upside-down shot breaks even good models.
  2. Among local models, bigger isn’t better. A 7B model beat an 11B and a 12B; a 2B was effectively blind. What separates a safe model from a risky one is its fabrication rate — invented books — not its size or its accuracy score.
  3. A frontier model is reliable and cheap. It read a clean 13-book shelf perfectly for about a cent.
  4. Skip the OCR step. Showing the model the photo directly was cheaper, faster, and kept the shelf’s order — reading the spines with OCR first didn’t help.
  5. On easy shelves the tiers converge. Given neat shelves and sharp photos, four capable models — closed and open-weight — all read at 93–98%. The gap shrinks to a handful of books, and the only signal left is fabrication: the rare invented book, which is the error you can’t catch by eye. An open-weight model matched the frontier for a hundredth of the cost.

The studies

Study 1 · local models
How five AI models read shelf pictures
Five open-weight models over eight photos of one real, messy bookcase — 189 books. Does the model decide the result, or the photograph?
Best local reader (Qwen-7B): 0.73 recall, 1.00 on clean shelves. The 2B model: near zero.
Read the study →
Study 2 · frontier & OCR
OCR first, or just look?
One shelf, taken two ways: OCR the spine text then hand it to an LLM, versus showing the model the raw photo. Which is better — and can a cheap model do it?
Direct photo → Sonnet: 13/13 for ~1¢ — 4.4× cheaper than the OCR route, and it kept shelf order.
Read the study →
Study 3 · the best case
How well AI reads a tidy shelf
Six neat, sharp shelves — 183 books — read by four frontier models and two OCR pipelines. On the easy case, do the tiers still separate?
The capable readers converge at 93–98%; an open-weight model matched the frontier for ~$0.01. The residual signal is fabrication.
Read the study →

More studies will be added here as they are run; this page is the front door to all of them.