Nano Banana Pro: raw intelligence with tool use (quesma.com)

Google quietly rolled out gemini-3-pro-image-preview — popularly dubbed Nano Banana Pro — an image-generation model that pairs improved “raw” visual intelligence with built-in tool use (Google Search, Maps) and a novel internal chaining of “thought images” before final rendering. Reviewers highlight much stronger instruction following, longer consistent text in images, fewer hallucinations, and an ability to synthesize data into polished infographics and maps (examples include a 30‑year economic growth infographic for Poland and a cartoon-style SF→Yosemite route map). In short, Nano Banana Pro moves image models from creative demo art toward fact-driven, production-ready outputs by querying and integrating external information during generation. That leap has concrete technical implications and clear caveats. Tool integration enables data-driven visuals and complex scene composition, but intelligence is uneven: specialized skills are strong, yet critical domain errors remain (notably failing to produce a safe, correct electrical schematic). This combination raises both opportunity and risk — better educational and research visuals on one hand, and more convincing-but-wrong charts on the other. The release also exposes a measurement gap: current benchmarks (Text-to-Image Arena, GenAI Image Showdown) and manual checks won’t scale to task-specific factuality and instruction-following needs. Nano Banana Pro is a game changer for AI image generation, but practitioners must pair it with domain validation and push for more rigorous, automated evaluation.