> ## Documentation Index
> Fetch the complete documentation index at: https://docs.bowerlabs.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Teach Bird your own visual categories

> Build a labeled reference set from your own images, then ask Bird which of your categories a new photo matches. Hands-free in a live session.

**Shipped 2026-08-26 · Bird**

A lot of bench work comes down to a judgment call about what you are looking
at. Which bleaching stage is this coral at? Is this colony morphology type A or
type B? Which of the four wear patterns does this component show? The answer is
usually obvious with the reference images side by side, and much less obvious at
6pm with gloves on.

A **visual reference set** is your own labeled answer key. You pick the
categories, you pick the images that define each one, and Bird compares new
images against that set.

## What's new

* **Build a set in chat.** Tell Bird what the categories are and which images
  define them. It drafts the set, shows you every label with its reference
  images, and waits for you to press **Create set**. Nothing is written until
  you do.
* **Comparative descriptions.** Bower reads the whole set at once and writes
  down the cues that separate each category from the others, so the
  descriptions read as contrasts rather than captions. They stay at the level
  of what is visible.
* **Ask at the bench.** Point Bird at something new and ask *"which bleaching
  stage is this?"* or *"compare this against my wear pattern references"*. In a
  live session it works hands-free, so you can keep both hands on the sample.
* **Your own written criteria.** Give a category the distinction you would
  otherwise have to explain out loud every time, and Bird reads your wording
  from then on.

## The categories are yours

Bird never invents a taxonomy. It only reports which of the categories you
defined a new image most resembles, and it will not answer with a category you
did not define. There is no scientific identification and no diagnosis here:
this is a comparison against your own examples.

Match strength comes back as high, medium, or low, never as a percentage,
because a percentage implies a precision this does not have. If the comparison
is not clean, Bird says the match was inconclusive, names the categories it was
torn between, and asks the question that would settle it. That is the intended
behavior, not a failure. A confident wrong label costs far more than an honest
question.

Every match is an assistive suggestion for you to confirm. Bird will not tag the
image, file it, or change your reference set off the back of its own answer.

## Why

Most reference sets exist because a protocol asks for a judgment it cannot make
for you. A step that says *"score the bleaching stage"* is precise about the
procedure and silent about the call, because the call is visual and the protocol
is text. The judgment is the part that drifts: between people, between shifts,
and between whoever wrote the protocol and whoever is running it at 6pm.

## Notes

A set holds two to eight categories with one to three reference images each. Two
is the minimum because the descriptions are comparative, and there is nothing to
compare against with one.

Reference images must be **JPEG, PNG or WebP**. Bower accepts other image
formats, including HEIC, but they cannot be used as references.

Creating the set **moves** each reference image into the set's collection. If you
also want the set alongside a protocol, link the two with
[linked entities](/organisation/linked-entities) rather than re-filing the
images, because moving a reference image out detaches it from the set.

Bird compares one image per reply. Ask about the next one in your next message.

Full detail in [Visual reference sets](/ai-and-search/visual-reference-sets).