Google has rolled out a new feature called “conversational image segmentation” to interact with images using natural language inside its Gemini 2.5 model. Unlike earlier methods that required specific labels or bounding boxes, this feature allows you to describe parts of an image in your own words, and Gemini will understand what to highlight.
Instead of just saying “a car,” you can now prompt the model with phrases like “the car that is farthest away” or “the person holding the umbrella.” Gemini can now interpret these instructions and return a segmented result, understanding relationships, ordering, comparative attributes, conditions, and even abstract ideas like “damage” or “a mess.”
This is a big change from object-level tagging to contextual understanding. It allows for more fluid interaction with images, which is especially useful in creative workflows, compliance monitoring, and insurance assessments. For example, designers can now isolate “the shadow cast by the building” without fiddling with manual selection tools, and safety officers can ask Gemini to find “employees not wearing hard hats.”
Apart from that, the model also recognises in-image text through OCR, supports multi-language prompts, and handles complex logic like negations, for example, “people who are not sitting.”
This new feature is part of Google’s efforts to make the Gemini API easier for developers to use. You can try image segmentation right away in Google AI Studio or use it in Python with a ready-made Colab. Google says the goal is to help developers build visual tools without needing to train or run their own models.
However, that’s not the only update from Google in recent times. For example, a couple of days ago, Google rolled out Veo 3 integrated into the Gemini API with audio support and ready-made starter apps.
While that’s all about Google and Gemini for now, if you want to stay updated with all the upcoming developments, then make sure you join us on WhatsApp where we share all the latest tech and AI news along with in-depth reviews, analysis, and more.
