LMArena Nano Banana: How Google’s Mystery Model Topped Every AI Image Leaderboard

LMArena is a blind testing platform where AI models compete under random codenames — and one codename broke every record: nano-banana. After 5 million community votes, Google confirmed that Nano Banana was actually Gemini 2.5 Flash Image, their multimodal image generation and editing model. It held the #1 spot on both the Image Edit Arena and the Text-to-Image leaderboard simultaneously, with a 171-point Elo lead that no other model has matched.

This guide covers the full story behind nano-banana on LMArena — what the model can do, how it compares to Nano Banana Pro and Nano Banana 2, and three free ways to test it yourself.

What Is Nano Banana?

Nano Banana is the internal codename Google used for Gemini 2.5 Flash Image during anonymous pre-release testing on LMArena. Google confirmed this connection in August 2025, ending weeks of community speculation about the mysterious model that was dominating every image benchmark on the platform.

The name itself carries no special meaning. LMArena assigns random two-word codenames to models during blind testing to prevent brand bias — “nano-banana” was simply the label that stuck. The community adopted it so widely that even after Google’s official reveal, most users still search for “nano banana” rather than “Gemini 2.5 Flash Image.” Google leaned into this, making the model available under the identifier Gemini-2.5-Flash-Image Preview (nano-banana).

What makes this AI-powered image editing model stand out is scope. Most image AI tools specialize in either generation (creating images from text) or editing (modifying existing photos). Nano Banana handles both within a single model, using natural language prompting to interpret complex multi-step instructions. A user can upload a photo, describe changes in plain English, and receive an edited result that preserves lighting, perspective, and environmental coherence — capabilities that earned it the community nickname “Photoshop killer.”

How LMArena Tests AI Models

LMArena is where nano-banana’s reputation was built. Understanding how the platform works explains why the model’s dominance was so significant.

What Is LMArena?

LMArena (formerly LMSYS Chatbot Arena) is an open platform where AI models compete in blind head-to-head comparisons. Users see outputs from two anonymous models side by side and vote on which result is better. The platform uses an Elo rating system — the same ranking method used in competitive chess — to produce leaderboards ranked purely by human preference. No corporate marketing, no cherry-picked demos. Just blind votes.

How Nano Banana Dominated the Arena

Google’s newest image model entered the Image Edit Arena without any branding or announcement. The results were immediate. Nano Banana rose to #1 on both the Image Edit and Text-to-Image leaderboards within its first testing period. Over 5 million community votes were cast, with 2.5 million votes specifically evaluating Nano Banana outputs. The model achieved a 171-point Elo lead — the largest margin in Arena history — making it statistically clear this was not an incremental improvement but a generational leap.

Three Ways to Test on LMArena

LMArena offers three testing modes for image models:

1. Battle — anonymous voting where you see results from two unknown models and pick the better output
2. Side-by-Side — comparing two named models on the same prompt to directly evaluate differences
3. Direct Chat — one-on-one interaction with a specific model for iterative refinement and multi-step editing

No account is required for Battle mode. Side-by-Side and Direct Chat may require a free login depending on server load.

Core Capabilities of Nano Banana

Google’s text-to-image generator handles two distinct workflows — creating images from scratch and editing existing photos — with a level of coherence that separated it from every competitor on LMArena.

Text-to-image generation covers multiple visual styles. Photorealistic renders, artistic illustrations, stylized compositions, macro photography, and product mockups all fall within Nano Banana’s range. Multi-step prompts work reliably: instructions like “turn the bottom character into 2B from Nier: Automata and the top character into Master Chief from Halo” execute with clarity where older models produced confused blends.

Natural-language image editing is the signature capability. Upload an existing photo, describe changes in conversational English, and the model modifies the image while maintaining the original scene’s integrity. One-pass scene editing means background replacement, object placement, style transfer, and face completion happen in a single generation step. This is what earned the “Photoshop killer” label from the LMArena community.

Character and scene consistency reaches 95% across edits. Faces, body proportions, and clothing stay recognizable even after significant scene changes. Scene-aware replacements — changing a dog to a dragon while keeping the background untouched, for example — preserve the original composition’s spatial logic. This consistency metric is a major improvement over older models that would distort identity during complex edits.

Nano Banana vs Nano Banana Pro vs Nano Banana 2

Google released three generations of this model in rapid succession. Each targets a different use case and budget level. The naming can be confusing, so here is the breakdown.

Nano Banana (Original)

Powered by Gemini 2.5 Flash Image. This is the version that dominated LMArena. Free users get 3 generations per day at 1MP (approximately 1024×1024 resolution). It is optimized for speed — generating edits in roughly 0.8 seconds — making it ideal for quick concept exploration and casual use. The trade-off is resolution: outputs max out at 1MP, and complex text rendering inside images often produces garbled letters.

Nano Banana Pro

Powered by Gemini 3 Pro Image, Google DeepMind’s flagship image model. Pro represents a substantial jump in capability: 2K and 4K resolution output, improved texture and lighting realism, multilingual text rendering inside images, connection to Google Search for real-time knowledge grounding, and support for up to 14 image inputs with 5 consistent characters in a single scene. Access requires a Gemini Pro subscription at $19.99/month or Ultra at $124.99/month.

Nano Banana 2 (Latest)

Officially named Gemini 3 Flash Image (also known as GEMPIX2 in internal testing), Nano Banana 2 replaces both the original and Pro as the default image generator in the Gemini App. It combines faster generation speeds with enhanced 4K output and robust subject consistency for multiple objects. The “Redo with Pro” workflow lets users generate with Nano Banana 2 first, then selectively upgrade specific outputs to Pro-level quality.

How to Try Nano Banana for Free

Three paths give you access to Google’s AI image model without paying. Each works differently and has its own limitations.

1. Open the Gemini App on web or mobile and select “Create images” or “Thinking Model”
2. Type your prompt or upload an image you want edited
3. Review the output — free users receive 3 generations per day at 1MP resolution
4. After the free quota, the system reverts to a standard model until the next day
5. For developer access, visit Google AI Studio and select Gemini 2.5 Flash Image under “Image Models” — a free $300 trial credit is included for new accounts

LMArena itself offers unlimited free testing. Visit lmarena.ai, select Image modality, and use Battle mode for anonymous comparisons or Direct Chat to interact with Nano Banana one-on-one. No account or payment required for basic testing.

Pricing and Subscription Tiers

Google structures Nano Banana access across four tiers, each with different resolution caps, daily limits, and watermark policies.

• Free tier — 3 images per day at 1MP. Visible Gemini sparkle watermark plus invisible SynthID digital signature. Sufficient for testing and casual exploration
• Pro subscription ($19.99/month) — approximately 100 images per day at up to 4K resolution. Visible watermark retained but higher quality output and priority processing
• Ultra subscription ($124.99/month) — approximately 1,000 images per day at full 4K. No visible watermark, only invisible SynthID metadata. Highest priority and speed
• API / Developer access (~$0.15 per 4K image) — available through Google AI Studio or Vertex AI. Pay-per-generation billing, up to 16MP output, no visible watermark. Best for production integration

Known Limitations

Every generative AI model has failure modes, and Nano Banana is no exception. Knowing these limitations upfront saves time when the model produces unexpected results.

Visual artifacts appear as inconsistencies in reflections, lighting logic, or object placement — particularly in complex multi-element scenes. A person standing near water might have a reflection that does not match their pose, or shadows might point in conflicting directions. These errors are less frequent than in older models but they still surface in roughly 1 out of 10 complex compositions.

Text rendering remains a persistent challenge. Like most image AI models, Nano Banana struggles to produce fully legible text within generated images. Letters may appear mirrored, blurred, or missing. Nano Banana Pro (Gemini 3 Pro Image) improved this significantly with multilingual text support, but the original model and Nano Banana 2 still produce noticeable errors with numbers and letterforms.

Anatomical errors — particularly with hands and fingers — can still occur. Incorrect digit counts, fused fingers, and awkward joint angles appear most often in full-body compositions. Google DeepMind has acknowledged this as an active area of improvement across the Gemini model family.

SynthID: How Google Watermarks AI Images

SynthID embeds a digital watermark directly into the pixels of AI-generated images, making it imperceptible to the human eye while remaining detectable by automated tools.

Google DeepMind — SynthID

SynthID is Google DeepMind’s imperceptible digital watermarking technology, applied to every image Nano Banana generates. The watermark survives common transformations including cropping, resizing, JPEG compression, and screenshot capture. Users can verify any image by uploading it to the Gemini App and asking “Was this generated by Google AI?” — the system checks for SynthID signatures and returns a confidence score.

Watermark visibility depends on the subscription tier. Free and Pro outputs include both a visible Gemini sparkle logo in the corner and the invisible SynthID layer. Ultra and API outputs retain only the invisible SynthID — no visible branding. The watermark does not affect visual quality and cannot be removed without degrading the image beyond recognition, providing a reliable chain of provenance for AI-generated content.

FAQ