Image generation

Skill hyperfx-ai/marketing-skills/skills/image-generation

Generate images through the Hyper MCP with the unified `images_generate` tool — text-to-image, image-to-image, and branded ad creatives — choosing the model (gpt-image-2, nano-banana, nano-banana-pro, seedream-4.5) per task. Use when the user asks to generate an image, create an ad creative, do an image-to-image edit, render text inside an image, or produce a print-quality poster.From its SKILL.md

Install

npx -y skills add hyperfx-ai/marketing-skills --skill image-generation

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

SKILL.md

4.1 KB, 888 tokens by cl100k_base, as published. Nobody here has run it

Image Generation

Generate images with the images_generate tool. It handles text-to-image, image-to-image (pass reference_images), and multi-image composition. By default (model="auto") it picks the best model for the request; set model to choose one.

Requirements

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

Call shape

images_generate(
    requests=[{"id": "ad1", "prompt": "A polished SaaS ad, clean composition"}],
    aspect_ratio="16:9",     # "1:1" (default), "9:16", "16:9", "4:5", "2:3", "3:2", "3:4", "4:3", "21:9", ...
    quality="standard",      # "draft" | "standard" | "high"
    n=1,                      # 1-4 images per request
    model="auto",            # see "Choosing a model" below
)

Image-to-image / brand references: put files in the request: requests=[{"prompt": "Compose into a gift basket", "reference_images": ["file1", "file2"]}].
Reproducible output: pass seed=....
Ground in real-world search: pass use_search=True.
Do not display image URLs — they render automatically in chat.

Choosing a model

model="auto" is the right default. Override only when the task clearly calls for a specific model:

Task	`model`
First-pass concepts / quick ad ideation	`gpt-image-2`
Image-to-image with references, high-resolution refinement, broad aspect ratios	`nano-banana`
Readable text inside the image (posters, labels, infographics) or search-grounded scenes	`nano-banana-pro`
Product photography, material/fabric fidelity, accurate spatial depth	`seedream-4.5`

See references/image-prompting.md for per-model prompt-writing tips.

Branded / website ad creatives — extract branding first

If the user gives a website URL and wants on-brand creatives:

Call firecrawl_branding_extract with the URL → returns brand colors, fonts, personality/tone, and saved image files (logo, favicon, og_image).
Optionally firecrawl_urls_scrape with formats=["screenshot"] for visual context.
Write the prompt using the actual hex colors, font names, and tone, and pass the logo file_id in reference_images.

The branding result's file field is a JSON data file, NOT an image — never pass it as a reference. Only logo.file_id and images.*.file_id are usable images.

Higher-level workflows

For multi-shot product or marketplace work, prefer the workflow tools — they preserve product identity and return structured results:

images_product_photoshoots_create — multi-shot product photography (studio, lifestyle, hero, carousel, ad pack). See references/product-photoshoot.md.
images_marketplace_cards_create — marketplace listing image sets (Amazon main + secondary, A+ modules, Shopify). See references/marketplace-cards.md.

Reminders

Do NOT display image URLs to the user — they show automatically in chat.
Refine vague prompts unless the user wants verbatim generation.
Match aspect_ratio to intent (social, print, web).
Use quality="high" for production, "draft"/"standard" while iterating.
Generated file_ids can be reused as reference_images in later calls.
For website brand work, call firecrawl_branding_extract before generating.

What ships with it: 3 files

7.7 KB alongside SKILL.md

references/

Gives 0 of the 12 instructions most images graphics skills give in 888 tokens

Counted across 341 of the 354 authors here whose files we hold, read 2026-09-06

Generate images via gemini_generate_imagein 14 of 341, across 10 files
Set the aspect ratio via set_aspect_ratioin 14 of 341, across 10 files
Verify MCP image tools are available before generatingin 13 of 341, across 9 files
Show estimated cost before generatingin 13 of 341, across 9 files
Run the post-generation SEO checklistin 13 of 341, across 9 files
Structure prompts as subject, setting, style, lighting, compositionin 10 of 341, across 5 files
Infer the brand strategy before generatingin 9 of 341, across 5 files
Use at most two logo concept methodsin 9 of 341, across 5 files
Use one dominant palette with repeating accentsin 9 of 341, across 5 files
Extract fileKey and nodeId from the Figma URLin 9 of 341
Read product marketing context before asking questionsin 8 of 341, across 3 files
Keep the logo simple, symbolic, and ownablein 8 of 341, across 4 files

Said here and by no other author read

Use model='auto' by default, overriding only for clear model-specific tasks
Use nano-banana-pro for readable text or search-grounded scenes
Use seedream-4.5 for product photography and material fidelity
Set use_search=True to ground output in real-world search
Call firecrawl_branding_extract before generating branded website creatives
Write prompts using extracted hex colors, fonts, and tone

Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.