Image Generation (AI SDK)

Multi-provider image generation. Default provider: Tuzi (兔子API, api.tu-zi.com).

Script Directory

Agent Execution:

1. SKILL_DIR = this SKILL.md file's directory
2. Script path = ${SKILL_DIR}/scripts/main.ts

Step 0: Load Preferences ⛔ BLOCKING

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

0.1 Check API Key

Check if the selected provider's API key is available. For Tuzi (default):

# Check env, then .tuzi-skills/.env files
echo "${TUZI_API_KEY:-not_set}"
grep -s TUZI_API_KEY .tuzi-skills/.env "$HOME/.tuzi-skills/.env"

Result	Action
Key found	Continue to Step 0.2
Key NOT found	⛔ Run API key setup (see references/config/first-time-setup.md → "API Key Setup") → Store key → Then continue

CRITICAL: If API key is missing, MUST guide user to obtain and store it BEFORE any generation. Generation is BLOCKED until key is configured.

0.2 Check EXTEND.md

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

test -f .tuzi-skills/tuzi-image-gen/EXTEND.md && echo "project"
test -f "$HOME/.tuzi-skills/tuzi-image-gen/EXTEND.md" && echo "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) → 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
.tuzi-skills/tuzi-image-gen/EXTEND.md	Project directory
$HOME/.tuzi-skills/tuzi-image-gen/EXTEND.md	User home

EXTEND.md Supports: Default provider | Default quality | Default aspect ratio | Default image size | OpenAI image API dialect | Default models

Schema: references/config/preferences-schema.md

Usage

# Basic (uses Tuzi provider by default)
npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "A cat" --image cat.png

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

# With quality (Tuzi: 1k/2k/4k)
npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "A cat" --image out.png --quality 2k

# 4K VIP model
npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "A cat" --image out.png --model gemini-3-pro-image-preview-4k-vip

# With reference images
npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "Make it blue" --image out.png --ref source.png

# From prompt files
npx -y bun ${SKILL_DIR}/scripts/main.ts --promptfiles system.md content.md --image out.png

# Async model (auto-polls)
npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "A cat" --image out.png --model gemini-3-pro-image-preview-2k-async

# Other providers
npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "A cat" --image out.png --provider google
npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "A cat" --image out.png --provider openai
npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "一只可爱的猫" --image out.png --provider dashscope
npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "A cat" --image out.png --provider replicate

# OpenAI-compatible gateway dialect
npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "A cat" --image out.png --provider openai --imageApiDialect ratio-metadata

Options

Option	Description
--prompt <text>, -p	Prompt text
--promptfiles <files...>	Read prompt from files (concatenated)
--image <path>	Output image path (required)
--provider tuzi|google|openai|dashscope|replicate	Force provider (default: auto-detect, Tuzi first)
--model <id>, -m	Model ID (see Tuzi Models section for full list)
--ar <ratio>	Aspect ratio (e.g., 16:9, 1:1, 4:3). Tuzi converts to NxN format
--size <WxH>	Size override (e.g., 1024x1024, 16x9)
--quality normal|2k	Quality preset. Tuzi: maps to 1k/2k. Google: maps to 1K/2K
--imageSize 1K|2K|4K	Image size (Tuzi and Google). Overrides --quality
--imageApiDialect openai-native|ratio-metadata	OpenAI-compatible endpoint dialect. ratio-metadata is for gateways that expect aspect-ratio size plus metadata.resolution
--ref <files...>	Reference images. Tuzi: base64 in JSON body. Google: multimodal. OpenAI: edits API
--n <count>	Number of images
--json	JSON output

Tuzi Models

Tuzi API (api.tu-zi.com) is the default provider. Models differ in quality, speed, and supported parameters.

Recommended

Model ID	Alias	Quality	Notes
gemini-3-pro-image-preview	nano-banana-pro	--quality 1k/2k/4k	Default. High quality, supports quality param
gemini-3.1-flash-image-preview	nano-banana-2	--quality 1k/2k/4k	Fast, supports extended aspect ratios
gemini-3-pro-image-preview-vip	nano-banana-pro-vip	1k built-in	High quality, VIP
gemini-3-pro-image-preview-2k-vip	nano-banana-pro-2k-vip	2k built-in	High quality 2K, VIP
gemini-3-pro-image-preview-4k-vip	nano-banana-pro-4k-vip	4k built-in	High quality 4K, VIP
gemini-2.5-flash-image-vip	nano-banana-vip	1k built-in	Fastest, VIP

More Models

Model ID	Alias	Notes
gemini-3-pro-image-preview	nano-banana-pro	--quality 1k/2k/4k
gemini-2.5-flash-image	nano-banana	Fast
gemini-3-pro-image-preview-hd	nano-banana-pro-hd	HD built-in
gemini-3-pro-image-preview-2k	nano-banana-pro-2k	2K built-in
gemini-3-pro-image-preview-4k	nano-banana-pro-4k	4K built-in
gpt-image-2	—	OpenAI latest image model, supports modern size rules
gpt-image-1.5	—	Size: 1:1, 3:2, 2:3 only
bfl-flux-2-pro	flux-2-pro	Flux
bfl-flux-2-max	flux-2-max	Flux highest quality
flux-kontext-pro	kontext-pro	Multi-ref editing
flux-kontext-max	kontext-max	Multi-ref editing (max)
doubao-seedream-4-0-250828	Seedream 4.0	2K/4K
doubao-seedream-4-5-251128	Seedream 4.5	2K/4K
doubao-seedream-5-0-260128	Seedream 5.0 lite	2K/3K

Async Models

Auto-detected. Script submits task and polls until complete (5s interval, max 30min).

Model ID	Notes
gemini-3-pro-image-preview-async	1K async
gemini-3-pro-image-preview-2k-async	2K async
gemini-3-pro-image-preview-4k-async	4K async
mj-imagine	Midjourney, MJ params in prompt

Model-Specific Parameters

Quality (--quality or --imageSize 1K|2K|4K):

Applies to	Values	Notes
gemini-3.1-flash-image-preview	1k / 2k / 4k	Default model, quality adjustable
gemini-3-pro-image-preview	1k / 2k / 4k	Quality adjustable
*-2k-vip, *-4k-vip, *-hd	—	Quality built into model name, param ignored
Other models	—	Param ignored

Aspect ratio (--ar):

Applies to	Supported ratios
Gemini models (default)	1:1, 16:9, 9:16, 3:2, 2:3, 4:3, 3:4, 5:4, 4:5, 21:9
gemini-3.1-flash-image-preview	Above + 1:4, 4:1, 1:8, 8:1 (extreme ratios)
gpt-image-2	Flexible ratio via generated valid pixel size; custom --size must satisfy OpenAI constraints
gpt-image-1.5	1:1, 3:2, 2:3
doubao-seedream-*	Use explicit --size / --imageSize; --ar alone is not supported
Omitted	Model auto-decides

Reference images (--ref):

• gpt-image-*: OpenAI-style /images/edits multipart upload
• Other sync models: base64 data URL in JSON image field
• Async models: input_reference in FormData
• Most Tuzi sync models support reference images; doubao-seedream-3-0-* does not

Seedream (doubao-seedream-*):

Applies to	Size rules	Reference images
doubao-seedream-5-0-*	2K, 3K, or explicit WxH	Up to 14
doubao-seedream-4-5-*	2K, 4K, or explicit WxH	Up to 14
doubao-seedream-4-0-*	1K, 2K, 4K, or explicit WxH	Up to 14
doubao-seedream-3-0-*	Explicit WxH only	Not supported

Environment Variables

Variable	Description
TUZI_API_KEY	Tuzi API key (https://api.tu-zi.com)
TUZI_IMAGE_MODEL	Tuzi default model (default: gemini-3-pro-image-preview)
TUZI_BASE_URL	Custom Tuzi endpoint (default: https://api.tu-zi.com/v1)
GOOGLE_API_KEY	Google API key
OPENAI_API_KEY	OpenAI API key
DASHSCOPE_API_KEY	DashScope API key (阿里云)
REPLICATE_API_TOKEN	Replicate API token
GOOGLE_IMAGE_MODEL	Google model override
OPENAI_IMAGE_MODEL	OpenAI model override (default: gpt-image-2)
DASHSCOPE_IMAGE_MODEL	DashScope model override
REPLICATE_IMAGE_MODEL	Replicate model override
GOOGLE_BASE_URL	Custom Google endpoint
OPENAI_BASE_URL	Custom OpenAI endpoint
OPENAI_IMAGE_API_DIALECT	OpenAI-compatible endpoint dialect (openai-native or ratio-metadata)
DASHSCOPE_BASE_URL	Custom DashScope endpoint
REPLICATE_BASE_URL	Custom Replicate endpoint

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

Model Resolution

Priority (highest → lowest), all providers:

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

Agent MUST display model info before each generation:

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

Provider Selection

1. --provider specified → use it
2. --ref provided + no --provider → Tuzi > Google > OpenAI > Replicate
3. Only one API key available → use that provider
4. Multiple available → Tuzi first

OpenAI-Compatible Dialect

When provider=openai, the endpoint may still be an OpenAI-compatible gateway rather than native OpenAI infrastructure.

• openai-native: standard OpenAI Images API request body
• ratio-metadata: sends aspect-ratio size plus metadata.resolution, useful for compatibility gateways in front of non-native image backends

Current limitation: ratio-metadata is text-to-image only. When using --ref, stay on openai-native.

Quality Presets

Preset	Tuzi	Google	OpenAI
normal	1k	1K	1024px
2k (default)	2k	2K	2048px

--imageSize 1K|2K|4K overrides quality for Tuzi and Google.

Generation Mode

Default: Sequential (one at a time).

Parallel: Only when user explicitly requests. Use Task tool with run_in_background=true, recommended 4 subagents (max 8).

Error Handling

• Missing API key → ⛔ MUST run API key setup from Step 0.1 (guide user to https://api.tu-zi.com/token, store in .tuzi-skills/.env). Do NOT suggest GOOGLE_API_KEY or other provider keys unless user explicitly chose a different provider.
• Generation failure → auto-retry once
• Tuzi PROHIBITED_CONTENT → content rejection error
• Tuzi NO_IMAGE → prompt too vague, suggest more explicit prompt
• Async timeout → error after 30 minutes
• Invalid aspect ratio → warning, proceed with default

Replicate Models

Format: owner/name or owner/name:version

npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "A cat" --image out.png --provider replicate
npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "A cat" --image out.png --provider replicate --model google/nano-banana-2 --ar 16:9
npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "A fashion photo" --image out.png --provider replicate --model bytedance/seedream-4.5 --size 4K
npx -y bun ${SKILL_DIR}/scripts/main.ts --prompt "A cinematic scene" --image out.png --provider replicate --model wan-video/wan-2.7-image-pro --size 4K

Supported Families

Family	Example models	Reference images	Aspect ratio	Size handling	Notes
Nano Banana	google/nano-banana, google/nano-banana-pro, google/nano-banana-2	Up to 14	Standard documented ratios	--quality maps to 1K/2K, --size can infer ratio + 1K/2K	Best default compatibility
Seedream 4.5	bytedance/seedream-4.5	Up to 14	Standard documented ratios	--size supports 2K, 4K, or custom WxH	Good for high-resolution output
Seedream 5 Lite	bytedance/seedream-5-lite	Up to 14	Standard documented ratios	--size supports 2K or 3K	Simpler size presets
Wan 2.7 Image	wan-video/wan-2.7-image	Up to 9	W:H ratios converted into size	--size supports 1K, 2K, or custom WxH	Reference images use images field
Wan 2.7 Image Pro	wan-video/wan-2.7-image-pro	Up to 9	W:H ratios converted into size	--size supports 1K, 2K, 4K, or custom WxH	4K only for text-to-image (no --ref)

Replicate Notes

• Current tool saves exactly one image per request, so Replicate supports --n 1 only.
• --imageSize is not used for Replicate. Use --quality, --ar, or --size instead.
• Unknown Replicate models still work for prompt-only generation, but --ref, --ar, and --size are only guaranteed for the families listed above.
• Documented aspect ratios shared by Nano Banana and Seedream in this tool: 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9.

Extension Support

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