Image Generation (AI SDK)

User Input Tools

When this skill prompts the user, follow this tool-selection rule (priority order):

1. Prefer built-in user-input tools exposed by the current agent runtime — e.g., AskUserQuestion, request_user_input, clarify, ask_user, or any equivalent.
2. Fallback: if no such tool exists, emit a numbered plain-text message and ask the user to reply with the chosen number/answer for each question.
3. Batching: if the tool supports multiple questions per call, combine all applicable questions into a single call; if only single-question, ask them one at a time in priority order.

Concrete AskUserQuestion references below are examples — substitute the local equivalent in other runtimes.

Script Directory

{baseDir} = this SKILL.md's directory; scripts/... paths are relative to it. Main script: {baseDir}/scripts/main.ts. Batch payload helper: {baseDir}/scripts/build-batch.ts. Resolve ${BUN_X}: prefer bun; else npx -y bun; else suggest brew install oven-sh/bun/bun.

Step 0: Load Preferences (blocking)

No image generation until EXTEND.md exists. Check in order; first hit wins:

• Found → load and apply. If default_model.[provider] is null → ask model only.
• Not found → run first-time setup (references/config/first-time-setup.md) via AskUserQuestion to collect provider + model + quality + save location. Save EXTEND.md, then continue.

Legacy: if .baoyu-skills/baoyu-imagine/EXTEND.md exists and the new path doesn't, the runtime renames it to baoyu-image-gen. If both exist, the runtime leaves them alone and uses the new path.

EXTEND.md keys: default provider, quality, aspect ratio, image size, OpenAI image API dialect, default models, batch worker cap, provider-specific batch limits. Schema: references/config/preferences-schema.md.

Usage

More examples (per provider, batch file format): references/usage-examples.md.

${BUN_X} {baseDir}/scripts/main.ts --prompt "A cat" --image cat.png
${BUN_X} {baseDir}/scripts/main.ts --prompt "A landscape" --image out.png --ar 16:9 --quality 2k
${BUN_X} {baseDir}/scripts/main.ts --promptfiles system.md content.md --image out.png
${BUN_X} {baseDir}/scripts/main.ts --prompt "Make blue" --image out.png --ref source.png
${BUN_X} {baseDir}/scripts/main.ts --prompt "A cat" --image out.png --provider dashscope --model qwen-image-2.0-pro
${BUN_X} {baseDir}/scripts/main.ts --prompt "A cat" --image out.png --provider codex-cli --ar 16:9   # Codex login, no OPENAI_API_KEY; needs `codex` on PATH

# Batch; build the payload from outline.md + prompts/ (e.g. baoyu-article-illustrator output)
${BUN_X} {baseDir}/scripts/build-batch.ts --outline outline.md --prompts prompts --output batch.json --images-dir attachments
${BUN_X} {baseDir}/scripts/main.ts --batchfile batch.json --jobs 4

Reference-Image Identity Preservation

When a real person/character/object must be preserved from reference images:

• Use short, hard identity language instead of a long generic description: "Use the person/object in the reference image(s) as the same identity. Do not redesign it or create a similar-looking new subject. Only change scene, clothing, pose, lighting, rendering style, and composition. Keep the face/proportions/hair/key accessories/overall identity from the references." With multiple refs, say they are the same subject and jointly define identity.
• Long feature descriptions ("young East Asian woman, oval face, clear eyes...") make the model synthesize a new matching person instead.
• Prefer 2–4 curated existing source refs; many or multi-megabyte refs can destabilize streaming providers.
• Don't use newly generated outputs as refs unless the user asks (drift compounds).
• Too polished/influencer-like → reduce stylized refs, add anti-beautification constraints (no face slimming, eye enlargement, heavy makeup, commercial travel shoot, over-smoothing).
• Younger/older look → keep the face; express age via clothing, posture, scene, styling. Never ask to change facial identity.

Options

Environment Variables

Load priority: CLI args > EXTEND.md > env vars > <cwd>/.baoyu-skills/.env > ~/.baoyu-skills/.env

Codex/ChatGPT OAuth Is Not an OpenAI API Key

--provider openai calls the standard OpenAI Images API (/v1/images/generations, /v1/images/edits) and requires OPENAI_API_KEY. A Codex/ChatGPT login is a different entitlement: never paste a Codex OAuth token into OPENAI_API_KEY or just point OPENAI_BASE_URL at a Codex backend. Do not modify the openai provider to silently consume Codex OAuth.

To use a Codex subscription without an API key:

• Codex runtime: native imagegen skill/tool.
• Non-Codex runtime with codex CLI installed and logged in: --provider codex-cli (preferred; same retry/cache/batch flow). It spawns the bundled scripts/codex-imagegen/main.ts (upstream: packages/baoyu-codex-imagegen/src/main.ts).
• Hermes runtime with native image_generate: use as fallback; state whether refs were passed directly or reconstructed from extracted traits.

Fallback: if --provider openai fails only because OPENAI_API_KEY is missing and one of the backends above is available, use it rather than leaving the user waiting, and say whether it was true reference-image generation or a text reconstruction from visual traits. Details: references/codex-oauth-vs-openai-api-key.md, references/codex-image2-fallback.md.

Model Resolution

Priority for every provider: CLI --model <id> > EXTEND.md default_model.[provider] > env <PROVIDER>_IMAGE_MODEL > built-in default.

• OpenAI: default gpt-image-2.5-flare (fastest). gpt-image-2.5-sunburst is most capable for complex scenes and precise edits; gpt-image-2, gpt-image-1.5, gpt-image-1 and dated snapshots (e.g. gpt-image-2.5-flare-2026-09-08, gpt-image-2-2026-04-21) stay selectable.
• Google: default gemini-3-pro-image. gemini-3.1-flash-image is faster/cheaper; gemini-3.1-flash-lite-image is cheapest and 1K-only (--quality 2k / --imageSize 2K|4K clamp to 1K with a warning).
• DashScope: default qwen-image-2.0-pro; qwen-image-3.0-pro is the newest flagship with the same sizing rules.
• Azure: --model / default_model.azure is the deployment name (AZURE_OPENAI_DEPLOYMENT, alias AZURE_OPENAI_IMAGE_MODEL). Use gpt-image-2.5-flare if the deployment is named after the model, else the exact custom name.

Before each generation, display:

• Using [provider] / [model]
• Switch model: --model <id> | EXTEND.md default_model.[provider] | env <PROVIDER>_IMAGE_MODEL

OpenAI-Compatible Gateway Dialects

provider=openai only means OpenAI-compatible auth/routing, not native image semantics. Set default_image_api_dialect (EXTEND.md), OPENAI_IMAGE_API_DIALECT, or --imageApiDialect:

• openai-native: pixel size (1536x1024) and native quality fields — OpenAI and strict clones
• ratio-metadata: aspect-ratio size (16:9) + metadata.resolution (1K|2K|4K) + metadata.orientation — gateways fronting Gemini or similar. Text-to-image only; reference edits need openai-native or a provider with first-class edit support.

Provider-Specific Guides

Read when the user picks that provider or asks for non-default behavior:

Provider Selection

1. --ref provided + no --provider → auto-select Google → OpenAI → Azure → OpenRouter → Replicate → Seedream → MiniMax → Agnes (MiniMax subject reference is specialized for character/portrait consistency)
2. --provider specified → use it (with --ref, must be google/openai/azure/openrouter/replicate/seedream/minimax/codex-cli/agnes)
3. Only one API key present → that provider
4. Multiple keys → Google → OpenAI → Azure → OpenRouter → DashScope → Z.AI → MiniMax → Replicate → Jimeng → Seedream → Agnes
5. codex-cli is never auto-selected — pin via default_provider: codex-cli or --provider codex-cli. Runs codex exec through the bundled scripts/codex-imagegen/main.ts (with bun) on the user's Codex subscription; requires codex on PATH with an active codex login.

Quality Presets

Google/OpenRouter imageSize can be overridden with --imageSize 1K|2K|4K.

OpenAI native gpt-image-2.5-* / gpt-image-2: normal → quality=medium + low-latency valid size near the ratio; 2k → quality=high + 2048px-class sizes (2048x2048, 2048x1152, 1152x2048). Use explicit --size for custom or 4K (e.g. 3840x2160).

Aspect Ratios

Supported: 1:1, 16:9, 9:16, 4:3, 3:4, 2.35:1.

• Google multimodal: imageConfig.aspectRatio
• OpenAI: gpt-image-2.5-* / gpt-image-2 use the closest valid custom size; older GPT Image and DALL·E use the closest fixed size
• OpenRouter: imageGenerationOptions.aspect_ratio; with only --size <WxH>, the ratio is inferred
• Replicate: model-specific — google/nano-banana* uses aspect_ratio, bytedance/seedream-* uses documented Replicate ratios, Wan 2.7 maps --ar to a concrete size
• MiniMax: official aspect_ratio values; --size <WxH> without --ar sends width/height for image-01

Generation Mode

Default: sequential. Batch parallel: automatic when --batchfile has 2+ pending tasks.

Once prompt files are saved and the task is "generate all of these", prefer batch over subagents.

Parallel behavior: worker count auto, capped by config (built-in default 10), override with --jobs <count>; provider-specific throttling applies only in batch mode; each image retries up to 3 attempts; final output reports success count, failure count, and per-image failure reasons.

Error Handling

• Missing API key → error with setup instructions (for OpenAI see the Codex fallback above)
• Generation failure → auto-retry up to 3 attempts per image
• Invalid aspect ratio → warning, proceed with default
• Reference images with unsupported provider/model → error with fix hint

Extension Support

Custom configurations via EXTEND.md. See Step 0 for paths and schema.