Google DeepMind、高精度画像生成モデル「Nano Banana Pro」を発表 Introducing Nano Banana Pro

Google DeepMind has introduced Nano Banana Pro, a new image generation and editing model built on top of Gemini 3 Pro. It is positioned as a higher-end successor to the original Nano Banana (Gemini 2.5 Flash Image) released in August, aimed at users who need stronger control, fidelity and reasoning in visual output.

The headline improvement is text rendering inside images. Diffusion-style models have historically struggled with legible long-form text, multilingual characters and annotated diagrams. Nano Banana Pro is presented as substantially better at producing posters, infographics, diagrams, recipe cards, mockups and educational visuals where typography and layout matter as much as the imagery itself. By inheriting reasoning from Gemini 3 Pro, the model can also turn structured prompts or knowledge-heavy briefs into coherent visual explanations rather than purely decorative outputs.

Editing capabilities have likewise been expanded. The model accepts up to 14 reference images and can keep up to five people consistent across a composition, which is meaningful for brand work, character-driven storytelling and product scenes. Users can adjust lighting, depth of field, color grading, camera angle and focus through natural language, edging closer to a conversational Photoshop-style workflow. Outputs are available at 2K and 4K resolution, making the model more practical for print and large-format use cases that previous Gemini image models could not comfortably serve.

Distribution is broad from day one. On the consumer side, Nano Banana Pro is rolling out in the Gemini app, Google Search's AI Mode, NotebookLM and Google Ads' Asset Studio. For developers and enterprises, it is available through the Gemini API, Google AI Studio, Vertex AI and the Antigravity agent platform. Google says all generated images carry an invisible SynthID watermark, and that users of the Gemini app can additionally inspect C2PA provenance metadata to verify origin.

In the wider landscape, image generation has become a competitive arena where OpenAI's GPT Image 1, Black Forest Labs' FLUX family, Midjourney and Adobe Firefly are each pushing on text fidelity, editability and commercial safety. Google's distinct angle appears to be tight coupling between a frontier reasoning LLM and the image model, which may help with prompts that require world knowledge, charts or factual layouts rather than pure aesthetics. The combination of SynthID and C2PA is also notable: as concerns around synthetic media intensify ahead of further elections and enterprise deployments, provenance signals are becoming a baseline expectation rather than a differentiator, and Google seems to be aligning its rollout accordingly.

It remains to be seen how Nano Banana Pro will be priced relative to the standard Nano Banana tier and how its quality compares in independent evaluations, particularly for non-Latin scripts and dense infographics. Still, the release suggests that Google intends to make image generation a first-class capability of the Gemini 3 generation rather than a side feature, and to push it into surfaces — Search, Ads, NotebookLM — where reach is arguably as important as raw model quality.