Search for something inside an image
Instead of choosing from a fixed list of object categories, Search the Image lets you describe what you are looking for in English. ImageSignals then searches the image for regions that best match your query.
What can you search for?
Queries can describe objects, visual attributes or concepts that may appear in the image. These are examples; results vary by image and query.
- “red backpack”
- “person wearing glasses”
- “blue car”
- “wooden chair”
- “dog”
Common uses
Finding a specific object
Locate a particular item in a busy image.
Searching by description
Search using characteristics such as colour, appearance or context rather than only a fixed object label.
Visual exploration
Explore an image by asking where particular concepts or items may appear.
Image review
Quickly focus attention on regions that may be relevant to a specific question.
How does image search work?
The system compares the text query with visual regions in the image and highlights regions that appear most relevant.
Unlike conventional object detection, this approach is open-vocabulary: the query is not limited to a small predefined menu of labels.
The result is a similarity or relevance signal, not proof that the requested object is present.
Limitations
Queries currently work best in English
Use concise English descriptions for the most predictable results.
Ambiguous descriptions
Broad or subjective queries may produce uncertain matches.
Small objects
Very small or partially hidden targets can be difficult to localise.
Similarity is not certainty
A highlighted region represents the strongest visual match found by the model and may occasionally be incorrect.
Frequently Asked Questions
How is this different from Find Objects?
Find Objects uses a predefined set of object categories. Search the Image lets you enter a free-text query and look for visual regions matching that description.
Why do queries need to be in English?
The current model is optimised for English text queries. Other languages may produce less predictable results.
Can I search for colours or attributes?
Yes, descriptive queries such as “red bag” or “person wearing glasses” may work, although performance varies by image and query.
What does a match mean?
A match means the model found a region visually related to the query. It should be treated as a relevance signal, not definitive identification.
What happens if the object is not present?
The model may return no strong match, or the closest visual region may still have some similarity. Always interpret weak results cautiously.