Skip to content

Nano Banana 2 — Gemini Image Generation

STATUS: active OWNER: felice (primary visual assets) ALSO_USED_BY: christel, ilian, alexander, valentijn CATEGORY: third-party-integration COST: ~$0.09/image (Flash 1K), higher for Pro/4K ADDED: 2026-04-01 UPDATED: 2026-04-01


WHAT IS NANO BANANA 2

Nano Banana 2 is a CLI tool for AI image generation powered by Google Gemini 3.1 Flash Image model. Generates high-quality images from text prompts with support for multiple resolutions, aspect ratios, transparent backgrounds, and style transfer via reference images. Excellent text rendering in generated images. Search grounding for factual accuracy. Outputs JPEG or PNG files to the working directory.


INSTALLATION

Already installed at ~/tools/nano-banana-2/. Claude Code skill installed at .claude/skills/nano-banana/.

No further installation steps required. The tool is available to all agents with CLI access.


API KEY

The Google API key is stored in Vault at secret/admin-ui/api-keys (key: google). Configured locally at ~/.nano-banana/.env.

This is the same Gemini API key used for LLM inference. Nano Banana 2 uses it for the image generation endpoint.


CLI USAGE

nano-banana "prompt describing the image" [options]

KEY OPTIONS

Option Values Default Purpose
--resolution 0.5k, 1k, 2k, 4k 1k Output resolution
--aspect 16:9, 9:16, 1:1, 4:3, 3:4, 3:2, 2:3, 21:9 16:9 Aspect ratio
--model flash, pro flash Gemini model tier
--greenscreen (flag) off Transparent background (PNG output)
--ref path to image none Reference image for style transfer

EXAMPLES

Basic image generation:

nano-banana "a modern SaaS dashboard with dark sidebar, blue accent color, data table with 5 rows"

High-resolution hero image:

nano-banana "aerial view of Amsterdam canals at golden hour, editorial photography" --resolution 4k --aspect 21:9

Transparent background product shot:

nano-banana "a minimal white smartphone mockup showing a fitness app" --greenscreen

Style transfer from reference:

nano-banana "a team collaboration workspace, warm lighting" --ref ./brand-mood-reference.jpg

Pro model for higher quality:

nano-banana "photorealistic close-up of artisan coffee being poured, shallow depth of field" --model pro --resolution 2k


COST

Model Resolution Approximate Cost
Flash 0.5k ~$0.04/image
Flash 1k ~$0.09/image
Flash 2k ~$0.15/image
Flash 4k ~$0.25/image
Pro 1k ~$0.30/image
Pro 2k ~$0.50/image
Pro 4k ~$0.80/image

COST OPTIMIZATION: - Use Flash for iteration and exploration (covers 90% of use cases) - Use Pro only for final deliverables requiring maximum quality - Start at 1k resolution, upscale to 2k/4k only for print or hero images - Use 0.5k for rapid concept validation before committing to higher resolution


CAPABILITIES

Use Case Suitability Notes
Product mockups Excellent Use --greenscreen for compositing
UI previews Excellent Describe layout precisely for best results
Hero images Excellent Use --resolution 2k or 4k with --aspect 21:9
Marketing assets Excellent Style transfer via --ref for brand consistency
Brand mood boards Excellent Generate multiple variants at Flash/1k for speed
Storyboard frames Good Sequential prompts with consistent style via --ref
Text in images Excellent Superior text rendering compared to DALL-E 3 and Midjourney
Photorealistic content Good (Flash) / Excellent (Pro) Pro model for editorial-quality photography

AGENT USAGE

Agent Role Typical Use
Felice Visual Production Lead Primary user. Product shots, hero images, marketing assets, all visual deliverables
Christel Brand Strategist Brand exploration, mood boards, color palette visualization
Ilian Motion Producer Storyboard frames, reference images before animation
Alexander Design System Engineer Quick mockup reference before detailed Stitch screen generation
Valentijn Strategy Competitive visual analysis, market positioning visuals

OUTPUT

Images are saved to the current working directory as JPEG (default) or PNG (when using --greenscreen). Filenames are auto-generated from the prompt.

RULE: always save generated images to the project directory structure, never leave them in temporary locations. RULE: log all generations with prompt, cost, and accepted/rejected status for budget tracking.


LIMITATIONS

Limitation Workaround
Cannot generate logos Use SVG code generation (Alexander/Floris)
Cannot edit existing images inline Use --ref for style transfer from existing image
No video generation Use Runway (image-to-video) or Remotion (programmatic video)
No seamless tiling Use Midjourney with --tile for patterns
No character reference system Use --ref for approximate style consistency

EU COMPLIANCE

Google has EU data processing agreements (DPA) and operates EU data centers (Belgium, Netherlands, Finland, Germany). Image generation requests can be routed through EU infrastructure. Acceptable under GE EU service provider policy as infrastructure provider.

Google is not an EU-founded company, but qualifies under the "infrastructure provider" exception in the EU service provider policy due to lack of viable EU-native alternatives for multimodal AI generation at this capability level.


READ_ALSO: domains/visual-production/image-generation.md (image generation domain knowledge) READ_ALSO: domains/design/design-tool-workflow.md (design tool selection decision tree) READ_ALSO: development/integrations/google-stitch.md (Stitch for structured UI design) READ_ALSO: development/integrations/ui-ux-pro-max.md (design intelligence skill)