home/image-gen/nano-banana-2-prompting-guide

Nano Banana 2 Prompting Guide

GPTClaudeGemini··▲ 791 copies·updated 2026-07-14

nano-banana-2-prompting-guide.prompt

# Nano Banana 2 Prompting Guide

Comprehensive reference for prompting Google's Nano Banana 2 (Gemini 3.1 Flash Image) for high-quality image generation. Covers JSON prompt structure, professional output techniques, icon/logo design, and platform-specific tips. Compiled from community research across Reddit, X, YouTube, TikTok, Instagram, and web sources (April 2026).

---

## Model Overview

Nano Banana 2 launched February 26, 2026, built on Gemini 3.1 Flash. It combines the speed of Nano Banana (fast) with the quality of Nano Banana Pro:

- **Near-Pro quality at 2-5x speed** — generations typically take 20-30 seconds
- **4K output** — controllable aspect ratios and resolution
- **Precise text rendering** — legible text in images (infographics, greeting cards, marketing)
- **Search grounding** — pulls from Google Search during generation for real-world subject accuracy
- **Character consistency** — maintains identity across multiple generations from reference photos
- **Style templates** — built-in UI presets (cinematic, moody, steampunk, sketch, etc.)

Available in Gemini web app, mobile, Google AI Studio, and via API. Web interface defaults to Nano Banana 2; Pro/Ultra subscribers can regenerate with Nano Banana Pro via the three-dot menu.

---

## JSON vs Natural Language

**Use JSON. Always.** The community consensus is overwhelming:

| Metric | JSON | Natural Language |
|--------|------|-----------------|
| Color/lighting/composition precision | 92% | 68% |
| Processing speed | 200-700ms | Slower |
| Memory consumption | 25-30% less | Higher |
| Reproducibility | High (swap single fields) | Low (rewrite everything) |

JSON forces the model to categorize information, isolating variables like lighting, composition, and subject details. This prevents "concept bleeding" — where environment colors leak onto the subject, or background elements get held by a character.

**Source:** God of Prompt, Miraflow, Atlabs AI

---

## JSON Schema: The 6-Part Core Structure

The community has converged on a standard structure. Start with these six blocks:

when to use it

Community prompt sourced from the open-source GitHub repo rolandtolnay/llm-toolkit (MIT). A "Nano Banana 2 Prompting Guide" style prompt — adapt the placeholders and specifics to your task. Imported as-is and not independently retested here, so check the output before relying on it.