design-image-generation

Generate images using AI models — Gemini, DALL·E, Stable Diffusion, Midjourney, Flux, Leonardo, Ideogram, and more. MCP bridge tools are primary; the Design Stack media worker is the fallback for adva

SKILL.md

GitHub

AI Image Generation

Use this skill when creating AI-generated images for design projects, social media, brand assets, or content marketing.

MCP Tool Discovery

The mcp-ai-wpoos plugin may register tools under agent-specific prefixes or as bare names, depending on how the WordPress assistant was configured:

Pattern: nv_oos_{AGENT_NAME}_agent_{TOOL_NAME}
Example: nv_oos_sophie_agent_generate_gemini_image_validated
         nv_oos_local_agent_generate_gemini_image_validated
         generate_gemini_image_validated  (bare — may not be registered)

Before using any tool, check what's available in your tool list. The agent name varies per installation — Sophie, Local, or any custom name configured in WordPress. If bare tool names are registered, use those. Otherwise use the agent-prefixed variant.

When an MCP tool covers your task, use it directly. The Design Stack media worker (http://localhost:3100) is a fallback only for providers not yet available through MCP (Midjourney, Flux, Leonardo, Ideogram, Stable Diffusion, etc.).

Key MCP tools (names vary by agent prefix)

Function Tool Name Pattern What it does
Gemini image gen nv_oos_*_agent_generate_gemini_image_validated or bare generate_gemini_image_validated Generate images with Gemini
OpenAI/DALL·E image gen generate_openai_image Generate images with DALL·E 3 / GPT image models
DALL·E variations create_image_variation Create DALL·E 2 variations of an existing image
DALL·E editing edit_openai_image Edit images with DALL·E 2 (inpainting)
Image resize resize_image Resize images to specific dimensions
Background removal remove_background Make image backgrounds transparent
Alt text generation generate_image_alt_text_validated Generate accessible alt text
Image analysis analyze_image AI vision analysis of images
Web search nv_oos_*_agent_web_search_validated Research visual references before generating
Remote site ops remote_wp_connection List/manage connected spoke sites

WP-CLI Commands

Key mcp-ai WP-CLI commands for generation troubleshooting. Run via terminal:

Command Use for
wp mcp-ai provider List configured AI providers (Gemini, OpenAI, etc.)
wp mcp-ai health Check provider connectivity and tool registration
wp mcp-ai cache clear Clear provider caches after configuration changes
wp mcp-ai credential Manage API keys for image generation providers

When to use this skill

Trigger when ANY of the following is true:

  • The task asks to "generate an image", "create a visual", or "make a graphic".
  • The user needs hero images, social graphics, product mockups, or brand illustrations.
  • Selecting the right model/provider for a specific visual task.
  • The task mentions AI image generation with any provider.

Generating images via MCP (Primary)

Gemini (recommended default)

Use the Gemini image generation tool available in your tool list (agent-prefixed or bare). It's fast, produces high-quality images, and requires no external API setup beyond the WordPress plugin configuration.

Parameters (consistent across all variants):
  prompt       — Text description of the desired image
  model        — "gemini-3.1-flash-image" (default, fast) or "gemini-3-pro-image" (premium)
  aspect_ratio — "auto", "1:1", "3:4", "4:3", "9:16", "16:9"
  mime_type    — "image/png" (default), "image/jpeg", "image/webp"
  file_name    — Optional base name for the saved attachment
  output_format — "default" (raster) or "svg" (vectorized)
  timeout      — Override request timeout in seconds (5–300)

Example prompt: "A cozy coffee shop interior with morning sunlight streaming through large windows, warm wood tones, photorealistic, 4:3 aspect ratio"

The generated image is automatically saved to the WordPress Media Library as an attachment.

DALL·E variations

Use create_image_variation to generate alternative versions of an existing image in your Media Library:

Parameters:
  image_id       — WordPress attachment ID of the source image
  model          — "dall-e-2" (only option)
  n              — Number of variations (1–10, default 1)
  size           — "256x256", "512x512", "1024x1024"
  output_format  — "default" (raster) or "svg"

DALL·E image editing (inpainting)

Use edit_openai_image to edit parts of an image:

Parameters:
  image_id          — WordPress attachment ID of the image to edit
  prompt            — Description of the desired edits
  mask_id           — Optional mask image (transparent areas get edited)
  request_user_mask — Set true to have the user paint a mask in chat
  model             — "dall-e-2"
  n                 — Number of results (1–10)
  size              — "256x256", "512x512", "1024x1024"

Provider comparison (full ecosystem)

Quality & Style

Provider Access Best For Style
Gemini ✅ MCP Fast creative images, text+image output Native multimodal
DALL·E 3 ⚠️ Media Worker Marketing graphics, realistic scenes Natural language, photorealism
DALL·E 2 ✅ MCP (variations/edit) Variations, inpainting Basic edits
Midjourney v6 ⚠️ Media Worker Artistic, moody, cinematic Best aesthetics
Flux (Replicate) ⚠️ Media Worker Photorealistic, high-quality renders State-of-the-art detail
Leonardo XL ⚠️ Media Worker Game assets, character design Flexible presets
Firefly v3 ⚠️ Media Worker Commercial-safe, brand-compatible Adobe ecosystem
Ideogram v2 ⚠️ Media Worker Text-in-image, logo concepts Best text rendering
Stable Diffusion XL ⚠️ Media Worker Artistic styles, custom fine-tuning Open-source
SD3 / SD3 Turbo ⚠️ Media Worker Fast generation, latest quality Stability's newest
Getimg.ai ⚠️ Media Worker SDXL with fine-tuned models Realistic Vision, Dreamshaper
Clipdrop ⚠️ Media Worker Quick one-off images Simple API
DeepAI ⚠️ Media Worker Budget-friendly, niche styles Fantasy, cyberpunk

✅ = Available via MCP bridge ⚠️ = Requires media worker fallback

Speed & Cost

Provider Speed Free Tier Paid Starts At
Gemini 5–15s Generous free tier Pay-as-you-go
DALL·E 3 10–30s None $0.040/image (1024)
Midjourney 30–90s None $10/mo (200 fast hours)
Flux Schnell 2–5s None $0.003/image (Replicate)
SD3 Turbo 2–5s 25 credits Pay-as-you-go
Leonardo 5–15s 150 credits/day $12/mo
Ideogram 5–15s 25/day (slow) $7/mo
Getimg.ai 3–10s 100/month $12/mo
Firefly 10–20s 25/month $5/mo

Media worker fallback

For providers not covered by MCP tools, use the media worker API at http://localhost:3100:

DALL·E (OpenAI)

POST /api/image/generate
{ "model": "dall-e-3", "prompt": "...", "size": "1024x1024", "style": "vivid" }
  • Requires: OPENAI_API_KEY
  • Models: dall-e-3, dall-e-2
  • Sizes: 1024x1024, 1792x1024, 1024x1792
  • Style: natural or vivid
  • DALL·E 3 cannot create variations or edits — use DALL·E 2 for those (available via MCP)

Midjourney

POST /api/image/generate
{ "provider": "midjourney", "model": "midjourney-v6", "prompt": "...", "aspect_ratio": "16:9" }
  • Requires: MIDJOURNEY_API_KEY
  • Models: midjourney-v6, midjourney-v6.1, midjourney-v5.2, niji-v6
  • Use parameters: --ar 16:9 --style raw --v 6.1

Stable Diffusion (Stability AI)

POST /api/image/generate
{ "provider": "stability", "model": "sd3", "prompt": "...", "negative_prompt": "ugly, blurry" }
  • Requires: STABILITY_API_KEY
  • Models: sd3, sd3-turbo, stable-diffusion-xl
  • Keywords with commas: cinematic lighting, 8k, highly detailed
  • Negative prompts: ugly, deformed, blurry, low quality

Flux / SDXL / Kandinsky (Replicate)

POST /api/image/generate
{ "provider": "replicate", "model": "flux-schnell", "prompt": "..." }
  • Requires: REPLICATE_API_KEY
  • Models: flux-schnell, flux-pro, flux-dev, sdxl, playground-v2.5, kandinsky-3

Other providers

Provider API Key Env Var Models
Leonardo LEONARDO_API_KEY leonardo-xl, leonardo-vision, leonardo-phoenix
Ideogram IDEOGRAM_API_KEY ideogram-v2, ideogram-v1, ideogram-v1-turbo
Getimg.ai GETIMG_API_KEY realistic-vision, dreamshaper, juggernaut-xl, flux-schnell
Firefly FIREFLY_CLIENT_ID + FIREFLY_CLIENT_SECRET firefly-v3
DeepAI DEEPAI_API_KEY deepai-text2img
Clipdrop STABILITY_API_KEY (shared) clipdrop-sdxl

Listing available providers

GET http://localhost:3100/api/image/providers

Returns which providers are configured and their available models.

Prompt engineering

Structure

[Subject] + [Style/Medium] + [Lighting] + [Composition] + [Quality boosters]

Platform-specific dimensions

Platform Aspect Prompt Style
Instagram feed 1:1 Product-focused, clean background
Instagram story 9:16 Vertical, space for text overlay
Twitter/X 16:9 Horizontal, bold, shareable
LinkedIn 1.91:1 Professional, clean
Pinterest 2:3 Vertical, inspirational

Gemini specific (MCP primary)

  • Be explicit: "Generate an image of..."
  • Describe visual style clearly: "photorealistic", "vector illustration", "watercolor"
  • Works well with structured descriptions including colors, materials, lighting
  • The generated image is auto-saved to the WordPress Media Library
  • Use output_format: "svg" to vectorize the result for logos/icons

DALL·E specific (media worker)

  • Natural language works best
  • Use style: vivid for saturated, dramatic results
  • Text rendering is strong — can generate images with legible words

Stable Diffusion specific (media worker)

  • Keywords with commas: cinematic lighting, 8k, highly detailed, trending on artstation
  • Negative prompts: ugly, deformed, blurry, low quality, watermark, text, bad anatomy
  • CFG scale: 7–9

Midjourney specific (media worker)

  • Use parameters: --ar 16:9 --style raw --v 6.1
  • Quality: --q 2 (max), --q 0.5 (fast)
  • Creative: --weird 500 or --stylize 1000

Multi-Site: Generating for spoke sites

When working with multiple WordPress sites (Hub + spokes), use remote_wp_connection to discover which spoke sites are connected and available before generating images for them:

1. Discover connections:
   remote_wp_connection { action: "list_connections" }
   → Returns: [{ id: "conn_8srcad8zylhe", name: "Parfumerie Website", url: "..." }]

2. Verify site is active and has media:
   remote_wp_connection { connection_id: "conn_8srcad8zylhe", action: "get_media", per_page: 3 }

3. Generate images using your available Gemini tool (agent-prefixed or bare) —
   saved to the current site's Media Library. For cross-site generation, create
   the image on the Hub and then push to the spoke via remote_wp_connection
   create_post with the attachment URL.

4. (Future Gateway feature): Direct cross-site tool calls via site prefix:
   parfum::generate_gemini_image_validated { prompt: "..." }

Research-First Generation

Use the web search tool available in your tool list (e.g., nv_oos_*_agent_web_search_validated) to find reference images and style inspiration before generating. This dramatically improves prompt quality:

1. Search for visual references:
   web_search { query: "minimalist coffee shop interior design 2026 trends" }

2. Use search results to enrich your prompt with real-world details

3. Generate with the Gemini tool available in your tool list

WordPress integration

Generated images are automatically saved to the WordPress Media Library via MCP tools. The attachment ID is returned and can be used immediately for featured images, post content, or social sharing.

// Example: using a generated image as a featured image
$attachment_id = /* result from Gemini image generation */;
set_post_thumbnail( $post_id, $attachment_id );
update_post_meta( $attachment_id, '_ai_prompt', $prompt );
update_post_meta( $attachment_id, '_ai_model', $model );

Common Mistakes

WRONG:
  ● Using generate_gemini_image_validated without first checking your tool list for the correct agent-prefixed name
  ● Generating images at full resolution without resizing — uploads 4096px images to WordPress with no optimization
  ● Using "make it look good" as a prompt instead of structured [Subject]+[Style]+[Lighting]+[Composition] descriptions
  ● Skipping provider comparison — using DALL·E for photorealism when Flux would produce better results at lower cost
  ● Generating images without checking existing Media Library — duplicates assets already available
  ● Forgetting to verify spoke site connectivity via remote_wp_connection before attempting cross-site image generation

RIGHT:
  ✅ Search your tool list for "generate_gemini" to find the correct agent-prefixed tool name before generating
  ✅ Always follow generation with design-image-optimization — resize to target dimensions, convert to WebP
  ✅ Use structured prompts: "Product photography of [item], [lighting], [background], [style], luxury catalog"
  ✅ Match provider to purpose: Gemini for fast creative, Flux for photorealism, Ideogram for text-in-image
  ✅ Run semantic_content_search or search_attachments first to check for existing usable assets
  ✅ Call remote_wp_connection { action: "list_connections" } before any cross-site image operations

Critical rules

  • Check your available tools first — tool names vary by agent configuration. Search your tool list for generate_gemini_image or web_search to find the correct names.
  • MCP first, media worker second — always check if an MCP tool exists before reaching for the media worker.
  • Always optimize after generation — use resize_image (MCP) or the media worker to convert raw AI images to WebP.
  • Store prompts as metadata — for reproducibility and attribution.
  • Respect rate limits — Gemini: generous free tier; DALL·E 3: 5/min (Tier 1); Midjourney: 3 concurrent.
  • Auto-detect provider from model name — when using the media worker, just pass model, the worker finds the right provider.
  • Firefly is commercial-safe — trained on licensed Adobe Stock; ideal for client work.
  • Ideogram for text — if your image needs legible text, use Ideogram or DALL·E 3 via media worker.
  • Flux for photorealism — best quality-to-speed ratio via media worker.
  • Verify spoke sites before cross-site ops — use remote_wp_connection to list connections and test connectivity before targeting remote sites.
  • Research before prompting — use the web search tool available in your tool list to gather visual references for more accurate prompts.

What This Skill Does NOT Cover

  • Image optimization — use design-image-optimization for resizing, format conversion (WebP/AVIF), compression, and alt text generation.
  • Product photography standards — use design-product-photography for e-commerce shot types, composition rules, and catalog consistency.
  • Video creation — use design-video-creation for AI video generation (Sora, Veo, Runway) and FFmpeg processing.
  • Social media publishing — use design-social-publishing to schedule and post generated images to social platforms.
  • Brand identity design — use design-brand-kit for logos, color palettes, and typography that should guide image styling.
  • OCR and text extraction — use extract_image_text or design-document-generation for extracting text from images.

Cross-references

  • Run design-image-optimization after generation — use resize_image (MCP) to optimize.
  • Run design-social-content to generate captions matching visuals.
  • Run design-brand-kit for consistent visual identity.
  • Run design-media-workflow for pipeline automation including cross-site pipelines.
  • Run mcp-ai-wpoos-plugin for remote_wp_connection documentation and multi-site architecture.
  • Use generate_design_ai as an alternative generation path through the Design Stack unified pipeline.

References