image-generation - SKILL.md Agent Skill

name: image-generation description: Guide to image generation and editing in MassGen. Use when creating images, editing existing images, iterating on image designs, or choosing between image backends (OpenAI, Google Gemini/Imagen, Grok, OpenRouter).

Image Generation

Generate images using generate_media with mode="image". The system auto-selects the best backend based on available API keys.

Quick Start

# Simple text-to-image (auto-selects backend)
generate_media(prompt="A cat in space", mode="image")

# Specify backend and quality
generate_media(prompt="A logo for a coffee shop", mode="image",
               backend_type="openai", quality="high")

# Batch generation (parallel)
generate_media(prompts=["sunset over ocean", "mountain landscape", "city at night"],
               mode="image", max_concurrent=3)

Backend Comparison

Backend	Default Model	Strengths	API Key
Google (priority 1)	`gemini-3.1-flash-image-preview` (Nano Banana 2)	Fast, flexible sizes, image editing, multi-turn	`GOOGLE_API_KEY` or `GEMINI_API_KEY`
OpenAI (priority 2)	`gpt-5.4`	High quality, transparent backgrounds, continuation via response ID	`OPENAI_API_KEY`
Grok (priority 3)	`grok-imagine-image`	1k resolution, continuation via stored data URI	`XAI_API_KEY`
OpenRouter (priority 4)	`google/gemini-3.1-flash-image-preview`	Access to multiple models via single API	`OPENROUTER_API_KEY`

Key Parameters

Parameter	Description	Example
`prompt`	Text description of the image	`"A watercolor painting of mountains"`
`backend_type`	Force a specific backend	`"google"`, `"openai"`, `"grok"`, `"openrouter"`
`model`	Override default model	`"gemini-3-pro-image-preview"` for studio quality
`quality`	Image quality (OpenAI)	`"low"`, `"medium"`, `"high"`, `"auto"`
`size`	Image dimensions	See backends reference
`aspect_ratio`	Aspect ratio	`"16:9"`, `"1:1"`, `"4:5"`
`input_images`	Source images for image-to-image editing	`["photo.jpg"]`
`continue_from`	Continuation ID for multi-turn editing	`result["continuation_id"]`

Image-to-Image Editing

Transform existing images by providing input_images:

generate_media(
    prompt="Make it look like a watercolor painting",
    mode="image",
    input_images=["photo.jpg"]
)

Supported backends for image-to-image: Google (Gemini), OpenAI, Grok. The system auto-selects if your current backend doesn't support it.

Multi-Turn Editing (Continuation)

Iteratively refine images using continue_from:

# First generation
result = generate_media(prompt="A logo for a coffee shop", mode="image")

# Refine using the continuation ID
result2 = generate_media(
    prompt="Make the text larger and add a cup icon",
    mode="image",
    continue_from=result["continuation_id"]
)

Each backend uses a different continuation mechanism:

OpenAI: Passes previous_response_id (stateless)
Google Gemini: In-memory chat store (LRU, 50 items)
Grok: In-memory data URI store (LRU, 50 items)

Continuation only works for single image generation (not batch).

Google: Gemini vs Imagen

Google supports two API paths. Gemini (Nano Banana 2) is the default and recommended for most use cases. Imagen is only needed for advanced reference-image editing features.

Gemini models (gemini-*): generate_content() — text-to-image, image editing via input_images, multi-turn continuation
Imagen models (imagen-*): generate_images() / edit_image() — text-to-image with negative_prompt/seed/guidance_scale, plus style transfer, control editing, and subject consistency via reference images

For studio-quality precision and text rendering, use: model="gemini-3-pro-image-preview" (Pro-tier).

Need More Control?

Per-backend sizes, quality options, and quirks: See references/backends.md
Complete extra_params reference: See references/extra_params.md
Advanced editing (inpainting, style transfer, control, subject): See references/editing.md