image-generation

Image Generation Prompting Skill

This skill teaches you to be an effective art director when calling the Nano Banana MCP server (mcp__nanobanana__generate_image). It covers model selection, prompt engineering, visual grounding, cost optimization, and editing workflows. The MCP tools are session-available — this skill provides the expertise to use them well.

When this activates: Any request involving image generation, editing, upscaling, visual grounding, or prompt crafting for Nano Banana / Gemini image models.

---

Model Selection

Use model_tier to choose the right engine. Default to nb2 explicitly — avoid auto which may escalate to Pro unnecessarily, increasing cost.

Decision Tree

Is this a new project or prompt?
├── YES → Use nb2 (Nano Banana 2 / Gemini 3.1 Flash Image)
│         95% of Pro quality at a fraction of the cost
│
└── NO, maintaining existing NB1 pipeline?
    └── Use flash (Nano Banana 1 / legacy)
        Cheapest, fastest, no thinking mode

Is nb2 producing poor results after prompt refinement?
├── Try enabling thinking (if on Pro) or refining the prompt first
└── Still failing? → Use pro (Nano Banana Pro)
    Best for: extreme logical constraints, multi-layered compositions,
    complex spatial reasoning, visual grounding

Do you need visual grounding (Google Search image lookup)?
└── Use pro — the MCP tool marks enable_grounding as "Pro model only"
    (See Visual Grounding section for details)

Quick Reference

Tier	model_tier value	Best for	Thinking	Grounding	Max resolution
NB1	flash	Legacy pipelines, cheapest option	No	No	1024px
NB2	nb2	Default for all new work	See note	See note — set enable_grounding: false explicitly	4K
Pro	pro	Edge cases, grounding, complex reasoning	Yes	Yes (default)	4K

---

Golden Workflow

For any non-trivial image request, follow this cost-optimized workflow:

1. Generate 2-4 variations at low resolution — use n: 4 with resolution: "1k" to produce quick, cheap drafts
2. Review with the user — present the options, discuss what works
3. Upscale the winner — take the best result and re-generate at resolution: "high" or higher

Cost tip: The Batch API provides a 50% discount. For bulk workflows (brand assets, content calendars), generate dozens of low-res variations via batch, pick the best, then upscale only the winners.

Cost tip: The official guide notes that NB2 at 512px costs roughly the same as NB1 at standard resolution. The MCP tool's lowest resolution value is "1k" — the 512px option may be available via the direct API or a future MCP update. Use "1k" as the current cost-optimized draft resolution.

Always use this workflow by default unless the user explicitly asks for a single high-res image or time is more important than cost.

---

Prompt Writing Principles

Be Specific About the Medium

Don't just describe the subject — describe how it should look as if it were a real piece of art or photography:

• Photography: "cinematic golden-hour photograph", "macro lens close-up with shallow depth of field"
• Illustration: "Franco-Belgian comic book style with ink outlines", "watercolor botanical illustration"
• 3D: "Pixar-style 3D render with subsurface scattering", "isometric low-poly game asset"
• Traditional: "chunky wax-crayon strokes on lined notebook paper", "oil painting with visible brushwork"

Composition and Scene

Include explicit compositional direction:

• Camera angle: "shot from below", "bird's eye view", "Dutch angle"
• Lighting: "dramatic rim lighting", "soft diffused overcast", "neon-lit cyberpunk alley"
• Background: "clean gray-blue studio background", "blurred bokeh city lights"
• Spatial relationships: "standing next to", "reflected in a puddle", "silhouetted against"

Consistency Across Multiple Generations

When generating a series (e.g., comic panels, product shots), anchor consistency:

• Describe the character/subject identically each time
• Reference the same style and lighting conditions
• Use system_instruction to set a persistent style (max 512 chars)

Text Rendering

NB2 handles text well. When you need text in the image:

• Put the exact text in quotes within the prompt
• Specify font style, size relative to the image, and placement
• Keep text short — single words or short phrases work best

Negative Prompts

Use negative_prompt (max 1024 chars) to steer away from unwanted elements:

• Style avoidance: "photorealistic" when you want illustration
• Artifact prevention: "blurry, low quality, watermark, text artifacts"
• Content exclusion: specific objects or styles to avoid

System Instructions

Use system_instruction (max 512 chars) for persistent style guidance across a session:

• "Professional product photography with clean white background"
• "Dark fantasy illustration style with muted earth tones"
• "Minimalist line art, black ink on white, no shading"

---

Visual Grounding

Visual grounding uses Google Search to look up real-world imagery before generating — ensuring accuracy for specific locations, species, monuments, and other real subjects.

Discrepancy Note

The official Nano Banana guide attributes Image Grounding to NB2. However, the MCP tool schema marks enable_grounding as "Pro model only". Until this is clarified:

• Operational guidance: Use model_tier: "pro" when grounding is needed for accuracy
• The parameter defaults to true in the tool, so when using Pro it's enabled automatically
• When using NB2, explicitly set enable_grounding: false to avoid ambiguity — the tool accepts the parameter but the schema says it only applies to Pro
• If you're already on nb2 and results look factually accurate, the model may be using internal knowledge — but for guaranteed grounding, use Pro

When to Use Grounding

• Specific locations: Churches, bridges, city squares, niche buildings — "the main historical church in Voiron, France"
• Nature/species: Exact animal species, breeds, insects — "a machaon butterfly"
• Monuments and landmarks: Accurate architectural details

Limitations

• Cannot search for people — grounding will not retrieve images of specific individuals
• Adds latency and cost — only enable when factual accuracy matters
• Requires Pro tier in the current MCP server configuration

---

Thinking Mode

The thinking_level parameter controls reasoning depth.

Discrepancy Note

The official Nano Banana guide states that NB2 has a thinking mode that can be toggled on/off. However, the MCP tool schema restricts thinking_level to Pro model only. Until this is clarified:

• Operational guidance: Only set thinking_level when using model_tier: "pro"
• If a future MCP server update enables thinking on NB2, this section will apply to both tiers

Default: OFF (don't set it)

For most generation tasks, thinking adds latency without improving results. Leave it unset.

When to Enable (Pro tier in current MCP server)

Set thinking_level to "low" or "high" when using Pro and:

• The model generates nonsensical results despite clear prompts
• You're creating complex infographics with spatial layout requirements
• You're combining grounding with spatial reasoning (e.g., "accurate church with specific foreground elements")

Value	Use case
"low"	Faster Pro generation with light reasoning
"high"	Maximum quality, complex multi-element compositions

---

Aspect Ratios and Resolution

Available Aspect Ratios

Only these values are accepted by the MCP tool:

Ratio	Orientation	Good for
1:1	Square	Profile pics, icons, social media
2:3	Portrait	Phone wallpapers, book covers
3:2	Landscape	Photography, web hero images
3:4	Portrait	Instagram, product shots
4:3	Landscape	Presentations, traditional photo
4:5	Portrait	Instagram portrait
5:4	Landscape	Print photos
9:16	Tall portrait	Stories, mobile full-screen
16:9	Wide landscape	Desktop wallpaper, video thumbnail
21:9	Ultra-wide	Cinematic, web banners

Note: The official Nano Banana guide mentions extreme ratios (1:8, 1:4) for comic strips and banners. These are not currently in the MCP server's accepted values and will be rejected. Use 21:9 as the widest available option, or compose multiple images side-by-side. These extreme ratios may be added in a future MCP server update.

Resolution Options

Value	Notes
"high"	Default. Standard high resolution
"1k"	Good for drafts and the golden workflow
"2k"	Pro model only
"4k"	Pro model only. Maximum detail

---

Editing and Multi-Image Workflows

Single Image Editing

Provide an existing image and describe the edit:

• Use input_image_path_1 with a local file path
• Or use file_id with a Files API ID (for large or reused images)
• Set mode: "edit" explicitly, or leave as "auto" to auto-detect
• The prompt describes what to change: "Change the sky to sunset colors" or "Remove the person in the background"

Multi-Image Conditioning

Provide up to 3 input images to guide composition:

• input_image_path_1: Primary subject/reference
• input_image_path_2: Style reference or secondary subject
• input_image_path_3: Additional context

The prompt should describe how to combine them: "Place the person from image 1 in the setting from image 2, using the color palette from image 3."

File ID for Efficiency

When working with the same image repeatedly (iterative edits):

1. Upload once via mcp__nanobanana__upload_file — returns a URI and metadata
2. Use the returned URI as the file_id value in subsequent generate_image calls (e.g., files/abc123)
3. Avoids re-reading the file from disk each time

---

Parameter Quick Reference

Parameter	Type	Default	Guidance
prompt	string (1–8192 chars)	required	Be specific: subject, composition, style, lighting, medium
model_tier	flash / nb2 / pro / auto	auto	Use nb2 explicitly for new work
n	integer 1–4	1	Use 2–4 for the golden workflow
aspect_ratio	enum (see table)	null (1:1)	Match to output use case
resolution	high / 4k / 2k / 1k	high	Use 1k for drafts; 2k/4k Pro only
negative_prompt	string (max 1024)	null	Style/artifact avoidance
system_instruction	string (max 512)	null	Persistent style across generations
enable_grounding	boolean	true	Pro only. For real-world accuracy
thinking_level	low / high / null	null	Pro only. For complex compositions
mode	generate / edit / auto	auto	Explicit when intent is clear
input_image_path_1/2/3	file path	null	Up to 3 input images
file_id	string	null	Files API ID for reused images
output_path	file path	null	Custom save location
return_full_image	boolean	null	Return full-res in MCP response (large!)

---

References

For worked prompt examples organized by technique, read: references/prompt-patterns.md

For deep model comparison, cost optimization, and troubleshooting, read: references/model-selection.md

---

Behavioral Guidance

When the user asks you to generate an image:

1. Use the golden workflow by default — generate multiple low-res drafts, then upscale the best one. Skip this only if the user explicitly wants a single image or speed matters.

2. Ask clarifying questions for vague requests. "Make me a cool picture" needs refinement. Ask about:

   • Subject and scene
   • Style/medium (photo, illustration, 3D, etc.)
   • Mood and lighting
   • Intended use (determines aspect ratio and resolution)

3. Set model_tier: "nb2" explicitly rather than relying on auto. Only escalate to pro when NB2 consistently fails or grounding is needed.

4. Write detailed prompts even if the user gives a short description. Transform "a cat on a roof" into a full art-directed prompt with medium, lighting, composition, and style.

5. Offer refinement after generation. After showing results, suggest specific adjustments: "I can adjust the lighting, change the style, or try a different composition."

6. Use system_instruction when generating multiple related images to maintain visual consistency.

7. Match aspect ratio to use case. Don't generate 1:1 when the user needs a desktop wallpaper (16:9) or phone background (9:16).

8. For editing requests, confirm which image to edit and what changes are needed before calling the tool. Describe edits precisely in the prompt.

9. When using NB2, set enable_grounding: false explicitly since the parameter defaults to true but is documented as Pro-only.

10. Other available tools: mcp__nanobanana__show_output_stats shows recent generation history and disk usage. mcp__nanobanana__maintenance handles cleanup of expired files and local storage (use dry_run: true first). mcp__nanobanana__upload_file uploads large or reused images to the Files API.