K2G-D-2

Nano Banana Pro (Gemini 3 Pro Image)

Created 2026-09-05Updated 2026-09-05ai / generative
  • Description: Study note on "Nano Banana Pro", Google's Gemini 3 Pro Image (preview 2025-11-20, general availability 2026-05-28). The successor to Nano Banana, built on the Gemini 3 Pro reasoning model: best-in-class in-image text rendering (multilingual), blend up to 14 images, identity of up to 5 people, 1K/2K output (4K in preview), fine-grained camera/lighting/focus controls, SynthID + visible watermark. Product release, no paper; note contrasts it with the original Nano Banana.
  • My Notion Note ID: K2G-D-2
  • Created: 2026-09-05
  • Updated: 2026-09-05
  • License: Free to share: please credit Yu Zhang and link back to yuzhang.io

Table of Contents

1. Overview

Nano Banana Pro = Google's Gemini 3 Pro Image, the successor to Nano Banana (Gemini 2.5 Flash Image). Timeline: preview 2025-11-20, general availability 2026-05-28.

It is built on Gemini 3 Pro, Google's flagship reasoning model, and positioned as their most capable image generation + editing model: sharper multimodal reasoning, better real-world grounding, higher-fidelity synthesis. Like the original, it is a product documented via Google's blog and model card, with no paper.

2. What changed from Nano Banana

Same natively-multimodal idea (image generation as a modality of a Gemini model), scaled up on the stronger Gemini 3 Pro base. The headline gains:

  • In-image text rendering goes from a stated weakness of the original to a headline strength: Google calls it "the best model for creating images with correctly rendered and legible text directly in the image", including long passages and multilingual layouts, varied fonts, textures, and calligraphy.
  • More inputs blended: up to 14 images combined while keeping visual consistency (vs the original's more modest multi-image blending).
  • Stronger identity preservation: keep the likeness of up to 5 people across a composition.
  • Higher resolution: native 1K/2K output (GA), with 4K in preview, plus flexible aspect ratios.

3. Capabilities

Google-stated:

  • Text in images: long, legible, multilingual text rendering (the standout feature).
  • Multi-image composition: blend up to 14 images with consistent style/geometry.
  • Identity: preserve up to 5 distinct people's likenesses.
  • Fine-grained creative controls: localized edits; camera-angle changes; focus/depth-of-field (bokeh); scene relighting (e.g. day-to-night); color grading.
  • Resolution / ratios: 1K/2K (GA) and 4K (preview), multiple aspect ratios.

4. Serving, pricing, and watermarking

  • Where: rolling out across the Gemini app, Google Ads, Google Workspace (Slides, Vids), the Gemini API, Google AI Studio, Vertex AI, and Flow. Free tier has limited quota; paid tiers get more.
  • Pricing (third-party trackers, not Google's own text): about 0.134per1K/2Kimage∗∗and∗∗0.134 per 1K/2K image** and **0.24 per 4K image; end-to-end latency around ~24s P50 (OpenRouter tracker). Treat as reported, third-party figures.
  • Watermarking: all outputs carry the invisible SynthID watermark. Free and Pro tiers also get a visible Gemini "sparkle" watermark; that visible mark is removed for Google AI Ultra subscribers and in Google AI Studio.

5. Why "built on Gemini 3 Pro" matters

  • The image model inherits the base LLM's reasoning and world knowledge. A stronger base is the plausible reason text rendering and prompt-faithful composition improved so much: correctly spelling long/multilingual text and laying out a coherent multi-image scene is as much a language-and-reasoning problem as a pixel problem.
  • Same black-box caveat as the original: Google has not published how the image head works internally (diffusion, autoregressive image tokens, or hybrid). Only interface and behavior are documented.
  • For the underlying "native multimodal generation vs a standalone diffusion model" primer, see the original note: Nano Banana section 6.

6. Limitations and what is not disclosed

  • Not disclosed: model size, architecture internals (diffusion / autoregressive / hybrid), and training data. (Benchmark numbers are published: the Gemini 3 Pro Image model card reports per-language text-rendering error rates and Elo scores vs GPT-Image 1 / Flux Pro / Seedream v4.)
  • Provenance caveat: no paper. Capability claims are from Google's Nano Banana Pro blog post and model card; pricing/latency figures are from third-party trackers and are flagged as such.
  • Watermarking and quota policy vary by tier and were still rolling out at GA, so specifics may drift.

Personal comment: the jump in text rendering is the most credible signal that "put image generation inside a strong reasoning model" pays off, since legible multilingual text is exactly the failure mode pure diffusion models struggle with. But the "best model" claims rest on Google's own published numbers (text-rendering error rates, Elo scores), not independently audited results, so this note records the vendor's positioning rather than an independently verified ranking. This is my own reading, not a claim by Google.

References