Image Generation (AI SDK)

Official API-based image generation. Supports OpenAI, Azure OpenAI, Google, OpenRouter, DashScope (阿里通义万象), MiniMax, Jimeng (即梦), Seedream (豆包) and Replicate providers.

Script Directory

Agent Execution:

1. 1. {baseDir} = this SKILL.md file's directory
2. Script path = INLINECODE1
3. Resolve ${BUN_X} runtime: if bun installed → bun; if npx available → npx -y bun; else suggest installing bun

Step 0: Load Preferences ⛔ BLOCKING

CRITICAL: This step MUST complete BEFORE any image generation. Do NOT skip or defer.

Check EXTEND.md existence (priority: project → user):

CODEBLOCK0

CODEBLOCK1

Result	Action
Found	Load, parse, apply settings. If default_model.[provider] is null → ask model only (Flow 2)
Not found

⛔ Run first-time setup (references/config/first-time-setup.md) → Save EXTEND.md → Then continue |

CRITICAL: If not found, complete the full setup (provider + model + quality + save location) using AskUserQuestion BEFORE generating any images. Generation is BLOCKED until EXTEND.md is created.

Path	Location
INLINECODE8	Project directory
INLINECODE9

User home |

Legacy compatibility: if .baoyu-skills/baoyu-image-gen/EXTEND.md exists and the new path does not, runtime renames it to baoyu-imagine. If both files exist, runtime leaves them unchanged and uses the new path.

EXTEND.md Supports: Default provider | Default quality | Default aspect ratio | Default image size | Default models | Batch worker cap | Provider-specific batch limits

Schema: INLINECODE12

Usage

CODEBLOCK2

Batch File Format

CODEBLOCK3

Paths in promptFiles, image, and ref are resolved relative to the batch file's directory. jobs is optional (overridden by CLI --jobs). Top-level array format (without jobs wrapper) is also accepted.

Options

Option	Description
INLINECODE19, INLINECODE20	Prompt text
INLINECODE21

Read prompt from files (concatenated) | | --image <path> | Output image path (required in single-image mode) | | --batchfile <path> | JSON batch file for multi-image generation | | --jobs <count> | Worker count for batch mode (default: auto, max from config, built-in default 10) | | --provider google\|openai\|azure\|openrouter\|dashscope\|minimax\|jimeng\|seedream\|replicate | Force provider (default: auto-detect) | | --model <id>, -m | Model ID (Google: gemini-3-pro-image-preview; OpenAI: gpt-image-1.5; Azure: deployment name such as gpt-image-1.5 or image-prod; OpenRouter: google/gemini-3.1-flash-image-preview; DashScope: qwen-image-2.0-pro; MiniMax: image-01) | | --ar <ratio> | Aspect ratio (e.g., 16:9, 1:1, 4:3) | | --size <WxH> | Size (e.g., 1024x1024) | | --quality normal\|2k | Quality preset (default: 2k) | | --imageSize 1K\|2K\|4K | Image size for Google/OpenRouter (default: from quality) | | --ref <files...> | Reference images. Supported by Google multimodal, OpenAI GPT Image edits, Azure OpenAI edits (PNG/JPG only), OpenRouter multimodal models, Replicate, MiniMax subject-reference, and Seedream 5.0/4.5/4.0. Not supported by Jimeng, Seedream 3.0, or removed SeedEdit 3.0 | | --n <count> | Number of images | | --json | JSON output |

Environment Variables

Variable	Description
INLINECODE47	OpenAI API key
INLINECODE48

Azure OpenAI API key | | OPENROUTER_API_KEY | OpenRouter API key | | GOOGLE_API_KEY | Google API key | | DASHSCOPE_API_KEY | DashScope API key (阿里云) | | MINIMAX_API_KEY | MiniMax API key | | REPLICATE_API_TOKEN | Replicate API token | | JIMENG_ACCESS_KEY_ID | Jimeng (即梦) Volcengine access key | | JIMENG_SECRET_ACCESS_KEY | Jimeng (即梦) Volcengine secret key | | ARK_API_KEY | Seedream (豆包) Volcengine ARK API key | | OPENAI_IMAGE_MODEL | OpenAI model override | | AZURE_OPENAI_DEPLOYMENT | Azure default deployment name | | AZURE_OPENAI_IMAGE_MODEL | Backward-compatible alias for Azure default deployment/model name | | OPENROUTER_IMAGE_MODEL | OpenRouter model override (default: google/gemini-3.1-flash-image-preview) | | GOOGLE_IMAGE_MODEL | Google model override | | DASHSCOPE_IMAGE_MODEL | DashScope model override (default: qwen-image-2.0-pro) | | MINIMAX_IMAGE_MODEL | MiniMax model override (default: image-01) | | REPLICATE_IMAGE_MODEL | Replicate model override (default: google/nano-banana-pro) | | JIMENG_IMAGE_MODEL | Jimeng model override (default: jimengt2iv40) | | SEEDREAM_IMAGE_MODEL | Seedream model override (default: doubao-seedream-5-0-260128) | | OPENAI_BASE_URL | Custom OpenAI endpoint | | AZURE_OPENAI_BASE_URL | Azure resource endpoint or deployment endpoint | | AZURE_API_VERSION | Azure image API version (default: 2025-04-01-preview) | | OPENROUTER_BASE_URL | Custom OpenRouter endpoint (default: https://openrouter.ai/api/v1) | | OPENROUTER_HTTP_REFERER | Optional app/site URL for OpenRouter attribution | | OPENROUTER_TITLE | Optional app name for OpenRouter attribution | | GOOGLE_BASE_URL | Custom Google endpoint | | DASHSCOPE_BASE_URL | Custom DashScope endpoint | | MINIMAX_BASE_URL | Custom MiniMax endpoint (default: https://api.minimax.io) | | REPLICATE_BASE_URL | Custom Replicate endpoint | | JIMENG_BASE_URL | Custom Jimeng endpoint (default: https://visual.volcengineapi.com) | | JIMENG_REGION | Jimeng region (default: cn-north-1) | | SEEDREAM_BASE_URL | Custom Seedream endpoint (default: https://ark.cn-beijing.volces.com/api/v3) | | BAOYU_IMAGE_GEN_MAX_WORKERS | Override batch worker cap | | BAOYU_IMAGE_GEN_<PROVIDER>_CONCURRENCY | Override provider concurrency, e.g. BAOYU_IMAGE_GEN_REPLICATE_CONCURRENCY | | BAOYU_IMAGE_GEN_<PROVIDER>_START_INTERVAL_MS | Override provider start gap, e.g. BAOYU_IMAGE_GEN_REPLICATE_START_INTERVAL_MS |

Load Priority: CLI args > EXTEND.md > env vars > <cwd>/.baoyu-skills/.env > INLINECODE95

Model Resolution

Model priority (highest → lowest), applies to all providers:

1. 1. CLI flag: INLINECODE96
2. EXTEND.md: INLINECODE97
3. Env var: <PROVIDER>_IMAGE_MODEL (e.g., GOOGLE_IMAGE_MODEL)
4. Built-in default

For Azure, --model / default_model.azure should be the Azure deployment name. AZURE_OPENAI_DEPLOYMENT is the preferred env var, and AZURE_OPENAI_IMAGE_MODEL remains as a backward-compatible alias.

EXTEND.md overrides env vars. If both EXTEND.md default_model.google: "gemini-3-pro-image-preview" and env var GOOGLE_IMAGE_MODEL=gemini-3.1-flash-image-preview exist, EXTEND.md wins.

Agent MUST display model info before each generation:

• - Show: INLINECODE106
• Show switch hint: INLINECODE107

DashScope Models

Use --model qwen-image-2.0-pro or set default_model.dashscope / DASHSCOPE_IMAGE_MODEL when the user wants official Qwen-Image behavior.

Official DashScope model families:

• - qwen-image-2.0-pro, qwen-image-2.0-pro-2026-03-03, qwen-image-2.0, INLINECODE114

- Free-form size in 宽*高 format - Total pixels must stay between 512*512 and 2048*2048 - Default size is approximately 1024*1024 - Best choice for custom ratios such as 21:9 and text-heavy Chinese/English layouts

• - qwen-image-max, qwen-image-max-2025-12-30, qwen-image-plus, qwen-image-plus-2026-01-09, INLINECODE125

- Fixed sizes only: 1664*928, 1472*1104, 1328*1328, 1104*1472, 928*1664 - Default size is 1664*928 - qwen-image currently has the same capability as qwen-image-plus

• - Legacy DashScope models such as z-image-turbo, z-image-ultra, INLINECODE136

- Keep using them only when the user explicitly asks for legacy behavior or compatibility

When translating CLI args into DashScope behavior:

• - --size wins over INLINECODE138
• For qwen-image-2.0*, prefer explicit --size; otherwise infer from --ar and use the official recommended resolutions below
• For qwen-image-max/plus/image, only use the five official fixed sizes; if the requested ratio is not covered, switch to INLINECODE143
• INLINECODE144 is a baoyu-imagine compatibility preset, not a native DashScope API field. Mapping normal / 2k onto the qwen-image-2.0* table below is an implementation inference, not an official API guarantee

Recommended qwen-image-2.0* sizes for common aspect ratios:

Ratio	INLINECODE149	INLINECODE150
INLINECODE151	INLINECODE152	INLINECODE153
INLINECODE154

768*1152 | 1024*1536 |
| 3:2 | 1152*768 | 1536*1024 |
| 3:4 | 960*1280 | 1080*1440 |
| 4:3 | 1280*960 | 1440*1080 |
| 9:16 | 720*1280 | 1080*1920 |
| 16:9 | 1280*720 | 1920*1080 |
| 21:9 | 1344*576 | 2048*872 |

DashScope official APIs also expose negative_prompt, prompt_extend, and watermark, but baoyu-imagine does not expose them as dedicated CLI flags today.

Official references:

• - Qwen-Image API
• Text-to-image guide
• Qwen-Image Edit API

MiniMax Models

Use --model image-01 or set default_model.minimax / MINIMAX_IMAGE_MODEL when the user wants MiniMax image generation.

Official MiniMax image model options currently documented in the API reference:

• - image-01 (recommended default)

- Supports text-to-image and subject-reference image generation - Supports official aspect_ratio values: 1:1, 16:9, 4:3, 3:2, 2:3, 3:4, 9:16, 21:9 - Supports documented custom width / height output sizes when using --size <WxH> - width and height must both be between 512 and 2048, and both must be divisible by 8

• - INLINECODE200

- Lower-latency variant - Use --ar for sizing; MiniMax documents custom width / height as only effective for INLINECODE204

MiniMax subject reference notes:

• - --ref files are sent as MiniMax INLINECODE206
• MiniMax docs currently describe subject_reference[].type as INLINECODE208
• Official docs say image_file supports public URLs or Base64 Data URLs; baoyu-imagine sends local refs as Data URLs
• Official docs recommend front-facing portrait references in JPG/JPEG/PNG under 10MB

Official references:

• - MiniMax Image Generation Guide
• MiniMax Text-to-Image API
• MiniMax Image-to-Image API

OpenRouter Models

Use full OpenRouter model IDs, e.g.:

• - google/gemini-3.1-flash-image-preview (recommended, supports image output and reference-image workflows)
• INLINECODE212
• INLINECODE213
• Other OpenRouter image-capable model IDs

Notes:

• - OpenRouter image generation uses /chat/completions, not the OpenAI /images endpoints
• If --ref is used, choose a multimodal model that supports image input and image output
• INLINECODE217 maps to OpenRouter imageGenerationOptions.size; --size <WxH> is converted to the nearest OpenRouter size and inferred aspect ratio when possible

Replicate Models

Supported model formats:

• - owner/name (recommended for official models), e.g. INLINECODE221
• INLINECODE222 (community models by version), e.g. INLINECODE223

Examples:

CODEBLOCK4

Provider Selection

1. 1. --ref provided + no --provider → auto-select Google first, then OpenAI, then Azure, then OpenRouter, then Replicate, then Seedream, then MiniMax (MiniMax subject reference is more specialized toward character/portrait consistency)
2. INLINECODE226 specified → use it (if --ref, must be google, openai, azure, openrouter, replicate, seedream, or minimax)
3. Only one API key available → use that provider
4. Multiple available → default to Google

Quality Presets

Preset	Google imageSize	OpenAI Size	OpenRouter size	Replicate resolution	Use Case
INLINECODE235	1K	1024px	1K	1K	Quick previews
INLINECODE236 (default)

2K | 2048px | 2K | 2K | Covers, illustrations, infographics |

Google/OpenRouter imageSize: Can be overridden with INLINECODE237

Aspect Ratios

Supported: 1:1, 16:9, 9:16, 4:3, 3:4, INLINECODE243

• - Google multimodal: uses INLINECODE244
• OpenAI: maps to closest supported size
• OpenRouter: sends imageGenerationOptions.aspect_ratio; if only --size <WxH> is given, aspect ratio is inferred automatically
• Replicate: passes aspect_ratio to model; when --ref is provided without --ar, defaults to INLINECODE250
• MiniMax: sends official aspect_ratio values directly; if --size <WxH> is given without --ar, width / height are sent for INLINECODE256

Generation Mode

Default: Sequential generation.

Batch Parallel Generation: When --batchfile contains 2 or more pending tasks, the script automatically enables parallel generation.

Mode	When to Use
Sequential (default)	Normal usage, single images, small batches
Parallel batch

Batch mode with 2+ tasks |

Execution choice:

Situation	Preferred approach	Why
One image, or 1-2 simple images	Sequential	Lower coordination overhead and easier debugging
Multiple images already have saved prompt files

Batch (--batchfile) | Reuses finalized prompts, applies shared throttling/retries, and gives predictable throughput |
| Each image still needs separate reasoning, prompt writing, or style exploration | Subagents | The work is still exploratory, so each image may need independent analysis before generation |
| Output comes from baoyu-article-illustrator with outline.md + prompts/ | Batch (build-batch.ts -> --batchfile) | That workflow already produces prompt files, so direct batch execution is the intended path |

Rule of thumb:

• - Prefer batch over subagents once prompt files are already saved and the task is "generate all of these"
• Use subagents only when generation is coupled with per-image thinking, rewriting, or divergent creative exploration

Parallel behavior:

• - Default worker count is automatic, capped by config, built-in default 10
• Provider-specific throttling is applied only in batch mode, and the built-in defaults are tuned for faster throughput while still avoiding obvious RPM bursts
• You can override worker count with INLINECODE264
• Each image retries automatically 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

Extension Support

Custom configurations via EXTEND.md. See Preferences section for paths and supported options.