Nanobanana 2 is here!

NanoBanana 2: A Deep Dive into Google’s Latest Image Generation Model

Key Concepts:

• NanoBanana 2: Google’s latest image generation and editing model, based on Gemini 3.1 Flash.
• Gemini 3.1 Flash/Pro: Underlying large language models powering NanoBanana 2 and its predecessor, NanoBanana Pro, respectively. Flash prioritizes speed, while Pro prioritizes quality.
• Prompt Adherence: The model’s ability to accurately interpret and execute user instructions.
• Hallucination: The tendency of AI models to generate incorrect or nonsensical information.
• ELO Score: A rating system used to compare the performance of different AI models.
• Higsfield: A platform integrating various AI models, including NanoBanana 2 and Soul 2.0.
• Soul 2.0: Higsfield’s foundation image model designed for aesthetically pleasing image generation.
• Grounding: Utilizing web search to ensure AI-generated content is accurate and relevant.

I. Introduction & Context

The video focuses on a detailed evaluation of NanoBanana 2, Google’s newest image generation and editing model. The presenter highlights the current landscape of image models capable of basic tasks like colorization and pose swapping, but emphasizes NanoBanana 2’s potential for handling extremely challenging prompts, particularly those requiring strong “world understanding.” The video aims to compare NanoBanana 2 with its predecessor, NanoBanana Pro, analyzing its performance, specifications, and accessibility.

II. World Understanding & Complex Prompts: Celebrity & Character Generation

A key strength of NanoBanana 2 is its ability to accurately depict a large number of recognizable figures. The presenter demonstrates this by prompting the model to create a group selfie featuring 20 celebrities (Johnny Depp, Jackie Chan, Taylor Swift, etc.). NanoBanana 2 successfully renders all individuals, though the high density results in some noise in facial details. In contrast, NanoBanana Pro struggles with coherence and misrepresents Blackpink. This demonstrates NanoBanana 2’s improved realism and overall image quality.

The model also excels at generating cartoon, anime, and 2D characters, accurately depicting a diverse group including Levi Ackerman, Ana Forger, Rick and Morty, and Snow White. Both models perform well in this test, with NanoBanana 2 accurately rendering details like Kenny and Bart Simpson’s four fingers, reflecting their original cartoon designs.

III. Emotional Expression & Nuance

NanoBanana 2 is tested on its ability to generate a 4x4 grid of a young female displaying 16 different emotions (happiness, awe, sadness, etc.). The presenter notes that accurately portraying subtle emotions is a challenge for many image generators. NanoBanana 2 performs admirably, rendering most emotions convincingly, with slight ambiguity regarding “love.” NanoBanana Pro shows minor weaknesses in portraying jealousy and awe. The presenter concludes NanoBanana 2 is slightly more realistic overall.

IV. Challenging Tests: Pokémon, Where’s Waldo, Endangered Frogs & More

The video presents several demanding tests:

• Pokémon Grid: NanoBanana 2 successfully generates a 4x4 grid of Pokémon based on their Pokédex numbers, while NanoBanana Pro hallucinates a 3x5 grid with repeated and incorrect characters. NanoBanana 2 makes two errors (Unknown A and Cascoon), while NanoBanana Pro’s errors are more significant.
• Where’s Waldo: Both models generate complex “Where’s Waldo” images filled with meme characters, successfully incorporating Waldo. NanoBanana Pro’s image is deemed more detailed and closely resembles a traditional “Where’s Waldo” illustration.
• Endangered Frogs: The model is prompted to generate images of six critically endangered frogs with their scientific names and descriptions. NanoBanana 2 accurately identifies the species and provides correct descriptions, but the generated images are not always perfect matches. NanoBanana Pro shows similar performance. This test highlights the difficulty of generating accurate images of rare species with limited available data.
• Homework Assistance: NanoBanana 2 is tested on its ability to fill in blanks on a biology worksheet and a math assignment with handwritten-style text. It performs better than NanoBanana Pro on the math assignment, demonstrating improved accuracy.
• Floor Plan Generation: NanoBanana 2 struggles to accurately generate a photo from a floor plan, making errors in room layout. NanoBanana Pro performs slightly better, but still doesn’t perfectly capture the intended viewpoint. Conversely, NanoBanana 2 excels at generating a 2D floor plan from an image.
• Data Table to Chart Conversion: NanoBanana 2 successfully converts a data table into a chart, but makes errors in scaling the bars for certain data points. NanoBanana Pro performs more accurately in this test.

V. Higsfield Integration & Soul 2.0

The video introduces Higsfield, a platform that integrates various AI models, including NanoBanana 2 and its own foundation model, Soul 2.0. Soul 2.0 is specifically designed for generating aesthetically pleasing images, offering curated presets (Y2K, editorial, etc.) and features like style transfer and improved physics. The presenter demonstrates how to combine Soul 2.0 for initial image generation with NanoBanana Pro for editing, emphasizing the benefit of a strong base image.

VI. NanoBanana 2: Specifications & Performance

NanoBanana 2 is based on Gemini 3.1 Flash, prioritizing speed over the quality focus of NanoBanana Pro (Gemini 3 Pro). Key specifications include:

• Character Resemblance: Up to five characters.
• Object Fidelity: Up to 14 objects.
• Aspect Ratios: Includes new 4:1 and 8:1 panoramic ratios.
• Resolution: Up to 4K.

NanoBanana 2 is significantly faster than NanoBanana Pro while maintaining comparable quality. Independent benchmarks (artificial analysis) currently rank NanoBanana 2 as the leading text-to-image model, with a lower API cost. However, it ranks lower than NanoBanana Pro in image editing tasks.

VII. Accessibility & Conclusion

NanoBanana 2 is currently available for free on:

• Gemini App: Replaces NanoBanana Pro in both Fast and Pro models.
• Google Search (AI Mode): Integrated into Google’s AI-powered search experience.
• Google AI Studio: Requires a paid API key for advanced control over settings like temperature, resolution, and grounding with Google Search.

The presenter concludes that NanoBanana 2 is a powerful and versatile image generation model, offering significant improvements in speed, cost, and overall performance. While it may not surpass NanoBanana Pro in all editing tasks, its strengths in world understanding, prompt adherence, and accessibility make it a leading contender in the rapidly evolving AI landscape. The video ends with a promotion for an Nvidia RTX 5090 GPU giveaway in conjunction with GTC 2026.