All labs / Perception Checking model…
01.4 — Perception / Intermediate

Smart Vision

Perception Intermediate Hosted inference

Bridges pixels and words: an image encoder and a language decoder trained together, so the model can write a caption or answer a question about what it can see.

Model
BLIP (large)
Parameters
470M · ViT-L image encoder + text decoder
Input
Image + question
Output
Text
Typical time
2–6s
Last run
Input Awaiting an image

Nothing analysed yet

Add an image, choose what you want the model to produce, and run it.

Output

BLIP pairs an image encoder with a language decoder. It writes one word at a time, looking back at the image patches at every step to decide what to say next.

On alt text Machine-generated alt text is a useful starting point and a poor finishing point. It describes what is visually present; it cannot know why the image is on the page. A photograph of a chart needs its figures, not "a graph on a white background". Treat every result here as a first draft to be edited by someone who knows the context.
How it works

The image is cut into patches and encoded into vectors. A language decoder then generates words one at a time, attending back to those patches at every step to decide what to say next.

Full identifier
Salesforce/blip-image-captioning-large
Where it runs
Hosted inference API
Parameters
470M

Trained largely on web image–caption pairs, so it inherits their biases in who and what gets described, and how.

What to notice
  • Captions drift toward the generic. 'A man riding a horse' is a safe, high-probability sentence; specific details are riskier and so get produced less often.
  • It will confidently describe things that aren't there if the scene resembles a common training example. This is hallucination, and it looks exactly like a correct answer.
  • Ask a question the image cannot answer and it will still answer. There is no abstention.
A note on language

This page says “the model predicts”, not “the model knows”. That is deliberate. None of these systems understand the images or sentences you give them; they map inputs to outputs using patterns fixed at training time. The difference matters most exactly when the output is impressive.