Google Gemini 2.5 Revolutionizes Image Processing: Beyond Object Recognition, It Can Also Understand Abstract Concepts and Relationships

AIbase基地

Published in AI News · 4 minute read · Jul 23, 2025

Google has recently introduced an innovative feature for its Gemini 2.5 AI model - "Conversational Image Segmentation," allowing users to analyze and highlight image content directly through natural language prompts. This technology goes beyond traditional image segmentation, enabling Gemini to understand and respond to more complex and semantic instructions.

Going Beyond Traditional Methods, Understanding Abstracts and Relationships

Traditional image segmentation usually focuses on identifying fixed categories of objects such as "dogs," "cars," or "chairs." Now, Gemini can understand and apply more complex language to specific parts of an image. It is capable of handling: relational queries, for example, "a person with an umbrella." Logic-based instructions, for example, "all people not sitting." Abstract concepts, even recognizing concepts like "clutter" or "damage" that do not have clear visual outlines.

In addition, thanks to the built-in text recognition feature, Gemini can also identify image elements that require reading on-screen text, such as "cashew candy" in a display case. This feature supports multilingual prompts and can provide object labels in other languages (such as French) as needed.

Wide Applications: From Design to Safety and Insurance

Google states that this technology has broad practical value in multiple fields: image editing: designers no longer need a mouse or selection tools, they can precisely select the desired area by simply giving verbal instructions, such as "select the shadow of the building." Workplace safety: Gemini can scan photos or videos and automatically identify violations, for example, "all people without helmets at the construction site." Insurance industry: claims adjusters can issue commands such as "highlight all buildings damaged by the storm," automatically marking damaged buildings in aerial images, significantly saving manual inspection time.

Developer-Friendly: API Access and Optimization Tips

This powerful feature does not require a special standalone model. Developers can access the "Conversational Image Segmentation" feature directly through the Gemini API, with all requests handled directly by Gemini models equipped with this feature.

The returned results are in JSON format, containing coordinates (box_2d), pixel masks (mask), and descriptive labels (label) of the selected image areas, providing convenience for subsequent development.

To achieve the best results, Google recommends using the gemini-2.5-flash model and setting the thinkingBudget parameter to zero to trigger an immediate response. Developers can conduct preliminary tests through Google AI Studio or Python Colab.

Apple AI Team Turmoil: Self-Developed and Open-Source Dreams Crumble, May Seek Third-Party Large Models!

Apple's AI team faces internal conflict as SVP Federighi rejects open-source plan, citing existing models and iPhone performance concerns. Team questions 'device-first' strategy. Apple delays Siri update, considers partnering with OpenAI, highlighting AI strategy dilemma between privacy and innovation.....

China's First Large Model Approved by Chief Physicians is Now Available on夸克AI Search

The夸克 Health Large Model has passed the written evaluation of 12 core medical disciplines by chief physicians in China, becoming the first large model to reach this level domestically. The model has been integrated into夸克AI Search, allowing users to access chief physician-level AI medical services through deep search. Its core breakthrough lies in establishing a slow-thinking capability, processing complex medical issues through chain reasoning and clinical deduction path modeling.夸克 adopts a dual data production line + dual reward mechanism training system, with a team of thousands of physicians labeling the data, including over 400 deputy chief physicians and above. The platform currently has more than 20 million monthly active medical student users.

Meta Launches Innovative Model AU-Nets to Revolutionize Text Processing

The Meta research team introduced an innovative architecture called AU-Net, breaking through the limitations of traditional tokenization techniques. The model learns directly from bytes, dynamically combining bytes into hierarchical sequence representations through a self-regressive U-Net structure. Its contracting path compresses bytes into semantic units, while the expanding path restores information and fuses details, with skip connections preserving local features. A multi-stage pooling strategy is used to extract semantics at the word and phrase levels, and multi-linear upsampling optimizes information fusion. The self-regressive generation mechanism ensures text coherence and significantly improves inference efficiency. This architecture

Kimi K2 Technical Report Released! How Does This Model Dominate the Global Open-Source Competition?

The Kimi team released the Kimi K2 technical report. This trillion-parameter large model, with 32 billion activated parameters, won the global open-source model competition, with performance comparable to top closed-source models such as GPT4.5. Its innovations include: using the MuonClip optimizer to process 15.5 trillion token data without loss; developing a multi-domain tool data synthesis pipeline; pioneering the restatement method to improve training efficiency, achieving the effect of ten rounds of original data in one round of rewritten data; optimizing model performance through supervised fine-tuning and reinforcement learning. The model is based on N