Image Generation

Create images from text descriptions directly within Starfish. The fastest path is the in-chat image tool — ask for an image in a conversation and it renders inline — and each generation is also kept in the Media library alongside your other work.

Available Models

Model	Backend	Notes
Nano Banana 2	Google gemini-3.1-flash-image	The default — best balance of quality, speed, and cost
Nano Banana 2 Lite	Google gemini-3.1-flash-lite-image	Fastest for drafts, variations, and quick edits
Nano Banana Pro	Google gemini-3-pro-image	Highest-detail Gemini option for polished assets
GPT Image 2.5 Flare	OpenAI gpt-image-2.5-flare	Fast everyday images with sharp detail and typography
GPT Image 2.5 Sunburst	OpenAI gpt-image-2.5-sunburst	Highest fidelity for polished work and precise edits
GPT Image 2	OpenAI gpt-image-2	Strong for typography, layouts, and illustrations
Seedream 4.5	ByteDance seedream-4.5	Realistic marketing visuals
Nano Banana	Google gemini-2.5-flash-image	Legacy fast option for familiar results

All models route through the Vercel AI Gateway with a single key — no per-provider account needed.

Capabilities

• Reference images — up to 14 on Nano Banana 2 / 2 Lite / Pro, and up to 4 on the legacy Nano Banana, the GPT Image models, and Seedream.
• Multiple images per request — the GPT Image 2 family can return up to 10 images in a single call. Every other model returns one.
• Extra controls — size, quality, background, and output format are exposed by the GPT Image 2 family only. The in-chat tool can pass them; the studio uses the model's defaults.
• Aspect ratios — the Gemini models and Nano Banana cover 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, and 21:9. GPT Image and Seedream 4.5 support 1:1, 9:16, and 16:9.

Generating in Chat

The agent can generate images directly in a conversation. This is useful when image creation is part of a bigger task — "summarize this email thread, then design a banner for the campaign" works in one chat.

Enabling it

Open the + menu in the chat composer (the same menu you use to add files or apps) and choose Create an image. The composer switches into image-creation mode for that message; leave the model on Automatic or open the submenu to pick a specific one. Nothing is generated without this — images have a real per-image cost.

Once armed, the agent has access to a generate_image tool it can call when your message clearly asks for an image.

How it shows up

While the agent works, the reply shows a grid-reveal placeholder — a mosaic that fills in as the picture arrives, resolving into the real image in the same slot. Ask for several images (on a model that supports multiple per request) and they render as one grid, with a placeholder cell for each image still in flight. If a job runs long, the frame shows a note that it can take a few minutes and that you can switch away and come back.

The image is also saved as an Artifact, so it shows up in the Media library (/designs) under the Images tab. HTML/PDF/CSV artifacts the agent produces live alongside it under the Designs tab.

Image Sessions

Each image generation is a conversational turn. Open a session from the Images tab of the Media library to view it full-size and continue from there:

Multi-Turn Refinement

Image sessions are conversational. After generating, you can:

• Describe specific changes ("make the background darker")
• Request variations ("same composition but in watercolor style")
• Adjust details ("remove the text in the top-left corner")

The model sees the prior turns' images and prompts as context, so you refine without starting over.

Save to a Folder

Click Save folder in the composer toolbar to pick a destination on your disk (uses the macOS File System Access API). Every subsequent generation in that session is auto-saved to a generated/ subfolder, and uploaded references to a references/ subfolder.

Continue in Chat

Hover any generated image and click Continue in chat. Starfish opens a fresh chat session with the image pre-attached as a multimodal input — the chat agent can see the image and reason about it (suggest edits, write copy that pairs with it, etc.). A Continue in chat image can also be reused as a reference for the next prompt via Edit this.

Automations

If you've added image creation to an Automation, the scheduled agent run can generate images on its own (e.g. "every Monday, draft a social card for this week's announcement").

Storage

• Generated bytes are saved to a local IndexedDB store keyed by session — survives reloads, no network round-trip on re-view.
• When the agent generates an image in chat, the bytes are written to the local artifacts table (SQLite) and the image is reachable from the Media library at /designs.
• Nothing leaves your machine after generation except the round-trip to the Gateway during creation.