Image Generation

Setup#

On first use, read setup.md.

When to Use#

User needs AI-generated visuals, edits, or consistent image sets. Use this skill to pick the right model, write stronger prompts, and avoid outdated model choices.

Architecture#

User preferences persist in ~/image-generation/. See memory-template.md for setup.

text

~/image-generation/
├── memory.md      # Preferred providers, project context, winning recipes
└── history.md     # Optional generation log

Quick Reference#

TopicFileInitial setupsetup.mdMemory templatememory-template.mdMigration guidemigration.mdBenchmark snapshotsbenchmarks-2026.mdPrompt techniquesprompting.mdAPI handlingapi-patterns.mdGPT Image (OpenAI)gpt-image.mdGemini and Imagen (Google)gemini.mdFLUX (Black Forest Labs)flux.mdMidjourneymidjourney.mdLeonardoleonardo.mdIdeogramideogram.mdReplicatereplicate.mdStable Diffusionstable-diffusion.md

Core Rules#

1. Resolve aliases to official model IDs first#

Community names shift quickly. Before calling an API, map the nickname to the provider model ID.

Community labelOfficial model ID to try firstNotesNano Bananagemini-2.5-flash-image-previewCommon nickname, not an official Google model IDNano Banana 2 / ProVerify provider docsUsually a provider preset over Gemini image modelsGPT Image 1.5gpt-image-1.5Current OpenAI high-tier image modelGPT Image mini / iMinigpt-image-1-miniBudget/faster OpenAI variantFLUX 2 Pro / Maxflux-pro / flux-ultraMany platforms rename these SKUs

2. Pick models by task, not by hype#

TaskFirst choiceBackupExact text in imagegpt-image-1.5IdeogramMulti-turn editsgemini-2.5-flash-image-preview``flux-kontext-proPhotoreal hero shotsimagen-4.0-ultra-generate-001``flux-ultraFast low-cost draftsgpt-image-1-mini``imagen-4.0-fast-generate-001Character/product consistencyflux-kontext-max``gpt-image-1.5 with referencesLocal no-API workflowsflux-schnellSDXL

3. Use benchmark tables as dated snapshots#

Benchmarks drift weekly. Use benchmarks-2026.md as a starting point, then recheck current rankings when quality is critical.

4. Draft cheap, finish expensive#

Start with 1-4 low-cost drafts, pick one, then upscale or rerender only the winner.

5. Keep a fallback chain#

If the preferred model is unavailable, fallback by tier:

1. same provider lower tier, 2) cross-provider equivalent, 3) local/open model.

6. Treat DALL-E as legacy#

OpenAI lists DALL-E 2/3 as legacy. Do not use them as default for new projects.

Common Traps#

• Using vendor nicknames as model IDs -> API errors and wasted retries
• Assuming "Nano Banana Pro" or "FLUX 2" are universal IDs -> provider mismatch
• Copying old DALL-E prompt habits -> weaker output vs modern GPT/Gemini image models
• Comparing text-to-image and image-editing scores as if they were the same benchmark
• Optimizing every draft at max quality -> cost spikes without quality gain

Security & Privacy#

Data that leaves your machine:

• Prompt text
• Reference images when editing or style matching

Data that stays local:

• Provider preferences in ~/image-generation/memory.md
• Optional local history file

This skill does NOT:

• Store API keys
• Upload files outside chosen provider requests
• Persist generated images unless user asks to save them

External Endpoints#

ProviderEndpointData SentPurposeOpenAIapi.openai.comPrompt text, optional input imagesGPT Image generation/editingGoogle Gemini APIgenerativelanguage.googleapis.comPrompt text, optional input imagesGemini image generation/editingGoogle Vertex AIaiplatform.googleapis.comPrompt text, optional input imagesImagen 4 generationBlack Forest Labsapi.bfl.aiPrompt text, optional input imagesFLUX generation/editingReplicateapi.replicate.comPrompt text, optional input imagesHosted third-party image modelsMidjourneydiscord.comPrompt textMidjourney generation via Discord workflowsLeonardocloud.leonardo.aiPrompt text, optional input imagesLeonardo generation/editingIdeogramapi.ideogram.aiPrompt textTypography-focused image generation

No other data is sent externally.

Migration#

If upgrading from a previous version, read migration.md before updating local memory structure.

Trust#

This skill may send prompts and reference images to third-party AI providers. Only install if you trust those providers with your content.

Related Skills#

Install with clawhub install <slug> if user confirms:

• image-edit - Specialized inpainting, outpainting, and mask workflows
• video-generation - Convert image concepts into video pipelines
• colors - Build palettes for visual consistency across assets
• ffmpeg - Post-process image sequences and exports

Feedback#

• If useful: clawhub star image-generation
• Stay updated: clawhub sync