Image Generation

Guide for selecting the right image generation tool through the Hyper MCP based on what the user is trying to make.

Requirements

This skill assumes the Hyper MCP is connected to your agent so the tools below are available. For brand-consistent ad creative work, Firecrawl must also be configured under your Hyper integrations.

Tool surface

Out of scope

• Full ad creative workflows (brand extraction → copy → image) — use ad-creative-generation.
• Video generation, transcription, captions, voiceover — use video-generation.

Tool Selection Framework

Rule 0: Website / brand ad creative requests → extract branding first (Firecrawl)

If the user provides a website URL and asks for ad creatives, brand-consistent visuals, or marketing assets, always extract branding before generating images.

1. Call firecrawl_extract_branding with the URL. The result returns brand colors, fonts, personality / tone, and saved image files (logo, favicon, og_image).
2. Optionally call firecrawl_scrape_url with formats=["screenshot"] to capture the landing page for visual context.
3. Compose your image generation prompt using the extracted colors (hex codes), font names, and personality tone. Pass the logo file_id from the result as a reference_image_file_id.

Critical: The result includes a file field — this is a JSON data file, NOT an image. Never pass it to image gen tools. Only logo.file_id and images.*.file_id are actual images you can use as references.

Rule 1: Explicit text-heavy assets → Nano Banana Pro (Gemini)

If the primary requirement is readable embedded text inside the image itself, such as headlines, labels, UI text, or poster copy:

nano_banana_image_generation(
    model="pro",
    requests=[{"id": "ad1", "prompt": "Create an ad with readable text '50% OFF'"}],
    aspect_ratio="16:9",
    image_size="2K"
)

Rule 2: Reference images or brand assets → OpenAI Edit by default

If the user provides images, branding assets, screenshots, or wants an initial branded composition, prefer openai_image_edit.

openai_image_edit(
    requests=[{"prompt": "Compose into gift basket", "reference_images": ["file1", "file2"]}]
)

Use Nano Banana only when the user explicitly wants Nano Banana or the image is text-heavy enough that text rendering is the main concern.

Rule 3: Initial ad creative or first-pass concepts → OpenAI

For initial creative exploration, first-pass ad concepts, or early branded directions, prefer OpenAI tools:

openai_image_generation(
    requests=[{"prompt": "A polished SaaS ad concept with clean composition"}],
    size="1024x1024"
)

For reference-based branded ads, prefer:

openai_image_edit(
    requests=[{"prompt": "Create a polished branded ad concept", "reference_images": ["file1", "file2"]}]
)

Rule 4: Nano Banana Pro for production text rendering

Use nano_banana_image_generation(model="pro") when the user explicitly wants Nano Banana or when the output needs strong text rendering inside the image. Do not default to Nano Banana for the first creative pass when OpenAI tools can handle the concept.

Available Tools

OpenAI Image Generation

Best for: Quick ideation, first-pass concepts, general creative requests without reference images.

• Model: gpt-image-2
• Standard sizes: 1024x1024, 1024x1536, 1536x1024
• Quality: low, medium, high
• Fast results for general creative requests

OpenAI Image Edit

Best for: Initial branded compositions, website-based ad creatives, image-to-image generation with OpenAI quality.

• Editing or composing images using references
• Multi-image composition

Nano Banana Image Generation (Gemini)

Best for: Text-heavy production assets, explicit Nano Banana requests, and high-resolution refinements.

Models:

• model="pro" (default): Gemini 3 Pro — best quality, supports Google Search grounding
• model="flash": Gemini 2.5 Flash — faster and cheaper

Key Capabilities:

• Text in images (headlines, labels, UI elements)
• High-resolution output (1K, 2K, 4K)
• Various aspect ratios (1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9)
• Image-to-image with reference_images
• Real-world grounding via use_google_search=True

Seedream 4.5 Image Generation (ByteDance via fal.ai)

Best for: Product photography, marketing assets, text rendering in images, material / fabric preservation, scenes with accurate spatial depth.

Key Capabilities:

• Fast generation (~2-3 seconds)
• Resolutions up to 4K (auto_2K, auto_4K, or preset sizes)
• Photorealistic quality with strong material / texture detail
• Size presets: square_hd, square, portrait_4_3, portrait_16_9, landscape_4_3, landscape_16_9, auto_2K, auto_4K

seedream_image_generation(
    requests=[{"prompt": "Product shot of leather handbag on marble surface"}],
    image_size="auto_2K"
)

Nano Banana Image Edit

Best for: Modifying single existing images.

• Targeted changes to uploaded images
• Same quality options as generation

Nano Banana Multi-Turn

ONLY use when the user explicitly says "use multi turn image editing".

• mode="sync": Chain mode — output from turn N becomes input for turn N+1
• mode="parallel": Independent mode — all turns processed independently

Request Format

All tools use the requests parameter:

requests=[
    {"id": "optional_id", "prompt": "Description", "reference_images": ["optional"]}
]

Common Patterns

Pattern H: Brand-Consistent Ad Creative

1. Call firecrawl_extract_branding with the website URL
2. Read the returned colors, fonts, personality, and logo file_id
3. Write a prompt that includes the actual hex colors, font names, and tone from the result
4. Pass the logo file_id as a reference image
5. Generate the first pass with openai_image_edit when using brand references, or openai_image_generation for loose ideation
6. Use nano_banana_image_generation(model="pro") only if the image itself needs strong readable text rendering

Important Reminders

• DO NOT display image URLs to the user — they are automatically shown in chat
• Refine vague prompts — unless the user explicitly wants verbatim generation
• Match aspect ratio to intent — social media, print, web, etc.
• Use appropriate resolution — 4K for production, 1K / 2K for iteration
• Remember file_ids — generated images can be reused in subsequent operations
• For website-based brand work — call firecrawl_extract_branding before generating ad creatives
• For initial ad creatives — prefer OpenAI tools first; reserve Nano Banana Pro for explicit text-heavy assets or explicit user preference