Baoyu Image Gen

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

Script Directory

Agent Execution:

1. {baseDir} = this SKILL.md file's directory
2. Script path = {baseDir}/scripts/main.ts
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):

# macOS, Linux, WSL, Git Bash
test -f .baoyu-skills/baoyu-image-gen/EXTEND.md && echo "project"
test -f "${XDG_CONFIG_HOME:-$HOME/.config}/baoyu-skills/baoyu-image-gen/EXTEND.md" && echo "xdg"
test -f "$HOME/.baoyu-skills/baoyu-image-gen/EXTEND.md" && echo "user"

# PowerShell (Windows)
if (Test-Path .baoyu-skills/baoyu-image-gen/EXTEND.md) { "project" }
$xdg = if ($env:XDG_CONFIG_HOME) { $env:XDG_CONFIG_HOME } else { "$HOME/.config" }
if (Test-Path "$xdg/baoyu-skills/baoyu-image-gen/EXTEND.md") { "xdg" }
if (Test-Path "$HOME/.baoyu-skills/baoyu-image-gen/EXTEND.md") { "user" }

| 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](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 | |------|----------| | .baoyu-skills/baoyu-image-gen/EXTEND.md | Project directory | | $HOME/.baoyu-skills/baoyu-image-gen/EXTEND.md | User home |

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

Schema: references/config/preferences-schema.md

Usage

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

# With aspect ratio
${BUN_X} {baseDir}/scripts/main.ts --prompt "A landscape" --image out.png --ar 16:9

# High quality
${BUN_X} {baseDir}/scripts/main.ts --prompt "A cat" --image out.png --quality 2k

# From prompt files
${BUN_X} {baseDir}/scripts/main.ts --promptfiles system.md content.md --image out.png

# With reference images (Google, OpenAI, OpenRouter, Replicate, or Seedream 4.0/4.5/5.0)
${BUN_X} {baseDir}/scripts/main.ts --prompt "Make blue" --image out.png --ref source.png

# With reference images (explicit provider/model)
${BUN_X} {baseDir}/scripts/main.ts --prompt "Make blue" --image out.png --provider google --model gemini-3-pro-image-preview --ref source.png

# OpenRouter (recommended default model)
${BUN_X} {baseDir}/scripts/main.ts --prompt "A cat" --image out.png --provider openrouter

# OpenRouter with reference images
${BUN_X} {baseDir}/scripts/main.ts --prompt "Make blue" --image out.png --provider openrouter --model google/gemini-3.1-flash-image-preview --ref source.png

# Specific provider
${BUN_X} {baseDir}/scripts/main.ts --prompt "A cat" --image out.png --provider openai

# DashScope (阿里通义万象)
${BUN_X} {baseDir}/scripts/main.ts --prompt "一只可爱的猫" --image out.png --provider dashscope

# DashScope Qwen-Image 2.0 Pro (recommended for custom sizes and text rendering)
${BUN_X} {baseDir}/scripts/main.ts --prompt "为咖啡品牌设计一张 21:9 横幅海报，包含清晰中文标题" --image out.png --provider dashscope --model qwen-image-2.0-pro --size 2048x872

# DashScope legacy Qwen fixed-size model
${BUN_X} {baseDir}/scripts/main.ts --prompt "一张电影感海报" --image out.png --provider dashscope --model qwen-image-max --size 1664x928

# Replicate (google/nano-banana-pro)
${BUN_X} {baseDir}/scripts/main.ts --prompt "A cat" --image out.png --provider replicate

# Replicate with specific model
${BUN_X} {baseDir}/scripts/main.ts --prompt "A cat" --image out.png --provider replicate --model google/nano-banana

# Batch mode with saved prompt files
${BUN_X} {baseDir}/scripts/main.ts --batchfile batch.json

# Batch mode with explicit worker count
${BUN_X} {baseDir}/scripts/main.ts --batchfile batch.json --jobs 4 --json

Batch File Format

{
  "jobs": 4,
  "tasks": [
    {
      "id": "hero",
      "promptFiles": ["prompts/hero.md"],
      "image": "out/hero.png",
      "provider": "replicate",
      "model": "google/nano-banana-pro",
      "ar": "16:9",
      "quality": "2k"
    },
    {
      "id": "diagram",
      "promptFiles": ["prompts/diagram.md"],
      "image": "out/diagram.png",
      "ref": ["references/original.png"]
    }
  ]
}

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 | |--------|-------------| | --prompt <text>, -p | Prompt text | | --promptfiles <files...> | 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\|openrouter\|dashscope\|jimeng\|seedream\|replicate | Force provider (default: auto-detect) | | --model <id>, -m | Model ID (Google: gemini-3-pro-image-preview; OpenAI: gpt-image-1.5; OpenRouter: google/gemini-3.1-flash-image-preview; DashScope: qwen-image-2.0-pro) | | --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, OpenRouter multimodal models, Replicate, 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 | |----------|-------------| | OPENAI_API_KEY | OpenAI API key | | OPENROUTER_API_KEY | OpenRouter API key | | GOOGLE_API_KEY | Google API key | | DASHSCOPE_API_KEY | DashScope API key (阿里云) | | 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 | | 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) | | REPLICATE_IMAGE_MODEL | Replicate model override (default: google/nano-banana-pro) | | JIMENG_IMAGE_MODEL | Jimeng model override (default: jimeng_t2i_v40) | | SEEDREAM_IMAGE_MODEL | Seedream model override (default: doubao-seedream-5-0-260128) | | OPENAI_BASE_URL | Custom OpenAI endpoint | | 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 | | 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 > ~/.baoyu-skills/.env

Model Resolution

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

1. CLI flag: --model <id>
2. EXTEND.md: default_model.[provider]
3. Env var: <PROVIDER>_IMAGE_MODEL (e.g., GOOGLE_IMAGE_MODEL)
4. Built-in default

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:

1. Show: Using [provider] / [model]
2. Show switch hint: Switch model: --model <id> | EXTEND.md default_model.[provider] | env <PROVIDER>_IMAGE_MODEL

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:

1. qwen-image-2.0-pro, qwen-image-2.0-pro-2026-03-03, qwen-image-2.0, qwen-image-2.0-2026-03-03

- 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

1. qwen-image-max, qwen-image-max-2025-12-30, qwen-image-plus, qwen-image-plus-2026-01-09, qwen-image

- 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

1. Legacy DashScope models such as z-image-turbo, z-image-ultra, wanx-v1

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

When translating CLI args into DashScope behavior:

1. --size wins over --ar
2. For qwen-image-2.0*, prefer explicit --size; otherwise infer from --ar and use the official recommended resolutions below
3. For qwen-image-max/plus/image, only use the five official fixed sizes; if the requested ratio is not covered, switch to qwen-image-2.0-pro
4. --quality is a baoyu-image-gen 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 | normal | 2k | |-------|----------|------| | 1:1 | 1024*1024 | 1536*1536 | | 2:3 | 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-image-gen does not expose them as dedicated CLI flags today.

Official references:

1. [Qwen-Image API](https://help.aliyun.com/zh/model-studio/qwen-image-api)
2. [Text-to-image guide](https://help.aliyun.com/zh/model-studio/text-to-image)
3. [Qwen-Image Edit API](https://help.aliyun.com/zh/model-studio/qwen-image-edit-api)

OpenRouter Models

Use full OpenRouter model IDs, e.g.:

1. google/gemini-3.1-flash-image-preview (recommended, supports image output and reference-image workflows)
2. google/gemini-2.5-flash-image-preview
3. black-forest-labs/flux.2-pro
4. Other OpenRouter image-capable model IDs

Notes:

1. OpenRouter image generation uses /chat/completions, not the OpenAI /images endpoints
2. If --ref is used, choose a multimodal model that supports image input and image output
3. --imageSize 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:

1. owner/name (recommended for official models), e.g. google/nano-banana-pro
2. owner/name:version (community models by version), e.g. stability-ai/sdxl:<version>

Examples:

# Use Replicate default model
${BUN_X} {baseDir}/scripts/main.ts --prompt "A cat" --image out.png --provider replicate

# Override model explicitly
${BUN_X} {baseDir}/scripts/main.ts --prompt "A cat" --image out.png --provider replicate --model google/nano-banana

Provider Selection

1. --ref provided + no --provider → auto-select Google first, then OpenAI, then OpenRouter, then Replicate (Jimeng and Seedream do not support reference images)
2. --provider specified → use it (if --ref, must be google, openai, openrouter, or replicate)
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 | |--------|------------------|-------------|-----------------|----------------------|----------| | normal | 1K | 1024px | 1K | 1K | Quick previews | | 2k (default) | 2K | 2048px | 2K | 2K | Covers, illustrations, infographics |

Google/OpenRouter imageSize: Can be overridden with --imageSize 1K|2K|4K

Aspect Ratios

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

1. Google multimodal: uses imageConfig.aspectRatio
2. OpenAI: maps to closest supported size
3. OpenRouter: sends imageGenerationOptions.aspect_ratio; if only --size <WxH> is given, aspect ratio is inferred automatically
4. Replicate: passes aspect_ratio to model; when --ref is provided without --ar, defaults to match_input_image

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:

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

Parallel behavior:

1. Default worker count is automatic, capped by config, built-in default 10
2. 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
3. You can override worker count with --jobs <count>
4. Each image retries automatically up to 3 attempts
5. Final output includes success count, failure count, and per-image failure reasons

Error Handling

1. Missing API key → error with setup instructions
2. Generation failure → auto-retry up to 3 attempts per image
3. Invalid aspect ratio → warning, proceed with default
4. 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.