Gemini 2.5 can now segment images from conversational prompts

For years, getting an AI to cut out an object from a photo meant thinking like a database. You had to provide clear nouns, like "car" or "dog." Even recent open-vocabulary tools needed explicit labels. But if you asked an AI to grab "the car that is farthest away," it would usually give up.

Google just rolled out conversational image segmentation for Gemini 2.5. The model can now parse complex phrases, spatial relationships, and conditional logic, and then hand back a precise pixel-level mask of exactly what you meant.

Slashing through vague descriptions

Previous visual models essentially matched pixels to isolated nouns. Gemini 2.5 tackles referring expression segmentation, which requires real reading comprehension alongside visual perception.

Google breaks this down into several query types. It can handle spatial and relational context, like finding "the person holding the umbrella" or counting to pinpoint "the third book from the left."

It also handles relative attributes and negative conditions. You can ask for "the most wilted flower in the bouquet," or filter for "the people who are not sitting." The model evaluates the physical posture and the negation together before drawing the mask.

Where it really shows off is with abstract concepts that lack fixed visual definitions. If an insurance adjuster asks to segment "homes with weather damage," Gemini uses its background knowledge to isolate storm-related dents and cracks while ignoring regular rust or light reflections.

Thanks to built-in OCR and multilingual support, you can also pinpoint objects using text written directly on them across various languages.

One API to replace the pipeline

For builders, this replaces what used to be a frustrating multi-model pipeline. You normally had to stitch together an object detector, a text reader, and something like SAM, then host and maintain all of them yourself.

Now it is available directly through the Gemini API. Google recommends using the gemini-2.5-flash model with the thinking budget turned off, asking for structured JSON that returns the 2D bounding box, the mask, and descriptive labels.

Developers can already try it inside Google AI Studio or through a dedicated Colab notebook provided by the Gemini team.

Practical use cases range from creative editing, like telling Photoshop to select "the shadow cast by the building," to workplace safety monitoring that flags "any employees on the factory floor not wearing a hard hat."

Why it matters

We think this is one of the most practical vision upgrades Google has shipped recently.

Frontier models have been great at talking about pictures for a while, but real-world enterprise tasks require precise, pixel-level action. By connecting high-level language reasoning directly to segmentation masks, Google bridges the gap between seeing an image and actually manipulating it.

For independent developers and small startups, skipping the headache of training and hosting dedicated computer vision models is a massive win. You can now build insurance inspection tools or creative editing features in an afternoon using a single lightweight flash API.

The clever part is leveraging world knowledge for abstract ideas like "mess" or "damage." But that is also where the risk lives. Abstract reasoning introduces ambiguity. If a safety system misinterprets what counts as an improper uniform or a missed helmet, subtle false negatives could create real liability. It is fantastic for human-in-the-loop workflows, but fully autonomous deployment will still need a tight leash.