Google's Gemini app generated one billion images with its Nano Banana Pro model in under two months, reaching the milestone just 53 days after launch. The image generation tool, officially known as Gemini 3 Pro Image, launched on November 20 and has become a key driver for Google's AI adoption.
Free users get three Nano Banana Pro generations daily, while Google AI Pro subscribers receive 100 images and Ultra tier users get 1,000. The model excels at text rendering across multiple languages and offers studio-quality creative controls including localized editing, camera angle adjustments, and lighting transformations.
Now Chinese startup Z.ai has released an open source challenger that outperforms Google's proprietary model in specific benchmarks. GLM-Image, a 16-billion parameter model from the recently public company, beats Nano Banana Pro on the CVTG-2k benchmark for complex visual text generation.
GLM-Image scored a Word Accuracy average of 0.9116 on the benchmark, which evaluates text rendering across multiple image regions. Google's Nano Banana 2.0 (referred to as Nano Banana Pro) scored 0.7788 on the same test, marking what Z.ai describes as a "generational leap in semantic control."
The open source model achieves this performance by abandoning the industry-standard pure diffusion architecture used by most image generators. Instead, GLM-Image employs a hybrid auto-regressive plus diffusion design that enables superior text-heavy image generation.
Despite the benchmark advantage, early user testing suggests GLM-Image trails Google's offering in practical application. Users report the open source model is "far less accurate at instruction following and text rendering" compared to Nano Banana Pro in real-world usage.
Google's model maintains an edge in single-stream English long-text generation with a score of 0.981 versus GLM-Image's 0.952, according to benchmark data. The proprietary tool also integrates across Google's ecosystem including AI Mode, NotebookLM, Google Slides, Vids, and Flow.
Nano Banana Pro's success comes amid growing scrutiny of AI image generation safety. Competitors like xAI's Grok have faced restrictions and country-level blocks due to unrestricted image generation features. Malaysia and Indonesia became the first countries to block Grok access earlier this month.
Google watermarks all Nano Banana Pro generations and edits, and the Gemini app includes built-in detection for identifying Google AI-generated content. The company's safety-focused approach contrasts with competitors facing regulatory challenges over explicit content generation.
For enterprises, the competition presents new options. GLM-Image offers a cost-effective, customizable alternative with friendly licensing terms. While it may lack the polish of Google's offering, the open source model provides "good enough" performance for specific text-heavy use cases like infographics and technical diagrams. Kie.ai's Z Image API and Nano Banana Pro API represent another front in the battle for developer adoption.
The benchmark results highlight a growing specialization in AI image generation. While Google dominates in aesthetics and ecosystem integration, open source alternatives are emerging with superior performance in niche applications like complex text rendering.
Both models target enterprise applications including collateral creation, training materials, onboarding documents, and stationary design. The divergence in strengths suggests companies may choose different tools based on whether they prioritize text accuracy or visual quality.
Google's Josh Woodward, vice president of Gemini and Google Labs, announced the billion-image milestone on X. The achievement follows Nano Banana's earlier success driving 10 million new users to Gemini in September and editing over 200 million images in its first week.
The rapid adoption underscores growing enterprise demand for AI image generation tools. As companies seek to automate content creation, the competition between proprietary and open source models is accelerating innovation in text rendering capabilities.
