Try a simple example
What can AI see in a picture?
Crop the sign. A small exception appears, while the table disappears. What an image supports depends on what is actually in the view.
Illustrative response
The large sign shows 8–16. A smaller note is not readable here, so I cannot rule out an exception.
Evidence in this view
- Large hours signVisible at this detail
- Small closure noteToo little detail in this teaching view
- Left cupVisible at this detail
- Middle cupVisible at this detail
- Partly hidden objectPartly hidden or cut off
Authored illustration and example responses. No image is sent to a model. The visibility rules are teaching choices, not measured AI performance.
For reduced detail, small text is replaced with strokes in this authored drawing. Crops use the original scene, not an enlargement of the reduced version.
Read the picture experiment
This invented café scene contains an 8–16 sign and a smaller “Closed Tue” note. In the reduced full view, the hours remain readable but the note does not. A crop from the original reveals both and excludes the table. The table contains two fully visible cups and a partly hidden cup-shaped object. More resolution cannot reveal what is behind the plant or outside the frame. These are authored examples, not responses from a vision model.
More detail, less context
The smaller full view hides the Tuesday exception. Cropping the original makes it readable, but removes the cups. Enlarging a missing or covered detail cannot supply evidence that was never visible.
Use a full view and a useful close-up
For a label, receipt or screenshot, keep the original and add a clear crop. Ask what is directly visible, what is uncertain and what would need another source. Verify important text against the original before acting.
Go deeper
Pixels become features
Some vision encoders divide an image into patches and transform them into learned representations. Our grid illustrates this idea; it does not run an encoder.
An encoder is not an answer generator
A vision-language system can connect an image encoder to a language model. LLaVA is one published example. The language model then generates a response conditioned on visual information and the prompt.
A representation is not complete evidence
This authored scene separates three limits: lost detail, missing context and occlusion. A crop can help with the first while worsening the second. Nothing in this activity measures a real model’s recognition accuracy.
Explore the original sources
Primary references. These explain the mechanisms; the examples here are authored teaching simulations.
What would you like to try next?
Choose another small example, or return to this one whenever you need it.
See all the examples