baoyu-image-gen

Image Generation (AI SDK)

Official API-based image generation. Supports OpenAI GPT Image 2.5, Azure OpenAI, Google, OpenRouter, DashScope (阿里通义万象), Z.AI GLM-Image, MiniMax, Jimeng (即梦), Seedream (豆包), Replicate and Agnes.

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. All scripts/... paths below are relative to {baseDir}. 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

This step MUST complete before any image generation — generation is blocked until EXTEND.md exists.

Check these paths in order; first hit wins:

• Found → load, parse, apply. If default_model.[provider] is null → ask model only.
• Not found → run first-time setup (references/config/first-time-setup.md) using AskUserQuestion to collect provider + model + quality + save location. Save EXTEND.md, then continue. Do not generate images before this completes.

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

Usage

Minimum working examples — see references/usage-examples.md for the full set including per-provider invocations and batch mode.

Identity-preserving reference prompts

When the user wants a real person/character/object preserved from reference images, do not replace the reference with a long generic description. Prefer short, hard identity-preservation language:

• "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."
• If using multiple references, state that they are the same subject and should jointly define identity.

Pitfall: long descriptions like "young East Asian woman, oval face, clear eyes..." can cause the model to synthesize a new person matching the description instead of preserving the referenced person.

# Basic
${BUN_X} {baseDir}/scripts/main.ts --prompt "A cat"

Reference-Image Identity Preservation

When the user wants a person/object preserved from reference images:

• Prefer a small curated set of existing source references (usually 2–4) over many images; large multi-megabyte refs can destabilize streaming providers.
• Make the prompt say the references are the same subject and the output must use that identity. Avoid long generic facial-feature descriptions that can cause the model to synthesize a new similar-looking person.
• Do not use newly generated outputs as references unless the user explicitly asks; generated refs compound drift.
• If results become too polished or influencer-like, reduce stylized refs and add explicit anti-beautification constraints (no face slimming, eye enlargement, heavy makeup, commercial travel shoot, over-smoothing).
• If the subject should look younger/older, preserve the face and express age through clothing, posture, scene, and styling; do not ask the model 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 --model gpt-image-2.5-flare uses the standard OpenAI Images API (/v1/images/generations or /v1/images/edits) and requires OPENAI_API_KEY. A Codex or ChatGPT desktop login is a different entitlement and is not a drop-in replacement for OPENAI_API_KEY; do not paste a Codex OAuth token into OPENAI_API_KEY or only set OPENAI_BASE_URL to a Codex backend.

If the user wants to use their Codex subscription / GPT Image 2 entitlement without an OpenAI API key, route through a Codex-native backend instead of this skill's openai provider:

• In Codex runtime: use the native imagegen skill/tool.
• In non-Codex runtimes with codex CLI installed and logged in: use baoyu-image-gen --provider codex-cli (preferred — it gives you the same retry / cache / batch flow as every other provider). The provider spawns the bundled scripts/codex-imagegen/main.ts; the same code lives upstream at packages/baoyu-codex-imagegen/src/main.ts for standalone callers.
• In Hermes runtimes with a native image_generate tool: use that tool as a fallback, and state whether reference images were passed directly or reconstructed from extracted traits.

Do not modify the existing openai provider to silently consume Codex OAuth. The first-class Codex-CLI path is the dedicated codex-cli provider, which has its own auth (Codex login), route (codex exec), request shape, and tests. See references/codex-oauth-vs-openai-api-key.md.

Model Resolution

Priority (highest → lowest) applies to every provider:

1. CLI flag --model <id>
2. EXTEND.md default_model.[provider]
3. Env var <PROVIDER>_IMAGE_MODEL
4. Built-in default

For OpenAI, the built-in default is gpt-image-2.5-flare (fast, lowest latency). gpt-image-2.5-sunburst is the most capable variant for complex scenes and precise edits; gpt-image-2, gpt-image-1.5, gpt-image-1, and dated GPT Image snapshots (e.g. gpt-image-2.5-flare-2026-09-08, gpt-image-2-2026-04-21) remain selectable with --model or OPENAI_IMAGE_MODEL.

For Google, the built-in default is gemini-3-pro-image. gemini-3.1-flash-image is the faster low-cost option, and gemini-3.1-flash-lite-image is the cheapest — it only produces 1K output, so --quality 2k / --imageSize 2K|4K is clamped to 1K with a warning.

For DashScope, the built-in default is qwen-image-2.0-pro; qwen-image-3.0-pro is the newest flagship and uses the same sizing rules.

For Azure, --model / default_model.azure is the Azure deployment name. AZURE_OPENAI_DEPLOYMENT is the preferred env var; AZURE_OPENAI_IMAGE_MODEL is kept as a backward-compatible alias. If your Azure deployment is named after the underlying model, use gpt-image-2.5-flare; otherwise use the exact custom deployment name.

EXTEND.md overrides env vars: if EXTEND.md sets default_model.google: "gemini-3-pro-image" and the env var sets GOOGLE_IMAGE_MODEL=gemini-3.1-flash-image, EXTEND.md wins.

Display model info before each generation:

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

OpenAI-Compatible Gateway Dialects

provider=openai means the auth and routing entrypoint is OpenAI-compatible. It does not guarantee the upstream image API uses OpenAI native semantics. When a gateway expects a different wire format, set default_image_api_dialect in EXTEND.md, OPENAI_IMAGE_API_DIALECT, or --imageApiDialect:

• openai-native: pixel size (1536x1024) and native OpenAI quality fields
• ratio-metadata: aspect-ratio size (16:9) plus metadata.resolution (1K|2K|4K) and metadata.orientation

Use openai-native for the OpenAI native API or strict clones; try ratio-metadata for compatibility gateways in front of Gemini or similar models. Current limitation: ratio-metadata applies only to text-to-image; reference-image edits still need openai-native or a provider with first-class edit support.

Provider-Specific Guides

Each provider has its own quirks (model families, size rules, ref support, limits). Read these 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's subject reference is more specialized toward character/portrait consistency)
2. --provider specified → use it (if --ref, must be google/openai/azure/openrouter/replicate/seedream/minimax/codex-cli/agnes)
3. Only one API key present → use that provider
4. Multiple keys → default priority: Google → OpenAI → Azure → OpenRouter → DashScope → Z.AI → MiniMax → Replicate → Jimeng → Seedream → Agnes
5. codex-cli is never auto-selected — set default_provider: codex-cli in EXTEND.md or pass --provider codex-cli. It spawns codex exec via the bundled scripts/codex-imagegen/main.ts TS entrypoint (run with bun) and uses the user's Codex subscription (no OPENAI_API_KEY). Requires codex on PATH with an active codex login.

Quality Presets

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

For OpenAI native gpt-image-2.5-* and gpt-image-2, normal maps to quality=medium and a low-latency valid size near the requested aspect ratio; 2k maps to quality=high and 2048px-class sizes such as 2048x2048, 2048x1152, or 1152x2048. Use explicit --size for valid custom or 4K outputs, 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-* and gpt-image-2 use the closest valid custom size for the requested ratio; older GPT Image and DALL·E models use their closest supported fixed size
• OpenRouter: imageGenerationOptions.aspect_ratio; if only --size <WxH> is given, the ratio is inferred
• Replicate: behavior is 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; if --size <WxH> is given without --ar, sends width/height for image-01

Generation Mode

Default: sequential. Batch parallel: enabled automatically when --batchfile contains 2+ pending tasks.

Rule of thumb: once prompt files are saved and the task is "generate all of these", prefer batch over subagents. Use subagents only when generation is coupled with per-image thinking or divergent creative exploration.

Parallel behavior:

• Default worker count is automatic, capped by config, built-in default 10
• Provider-specific throttling applies only in batch mode; defaults are tuned for throughput while avoiding RPM bursts
• Override with --jobs <count>
• Each image retries up to 3 attempts
• Final output includes success count, failure count, and per-image failure reasons

Error Handling

• Missing API key → error with setup instructions
• 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

Codex image2 fallback

If --provider openai --model gpt-image-2.5-flare fails because OPENAI_API_KEY is missing but the current runtime has a native image-generation backend or the repo-level codex-imagegen wrapper is available, use that path rather than leaving the user waiting. Be explicit about whether the fallback is true reference-image generation or only a text-prompt reconstruction from extracted visual traits. See references/codex-image2-fallback.md.

References

Extension Support

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