Images in a turn

Usually you want an image, so you call generate and get one. Sometimes the picture is part of what the model is saying: it replies with a sentence and a diagram, and “make it dawn instead” changes that same picture.

Only Google’s image models do this today. If you are not using them, you want image generation instead.

Reading them

No new service. It is the turn call you already make:

const turn =
  yield *
  gemini.turn({
    model: "gemini-3.1-flash-image",
    history: [ask("Draw a lighthouse at dusk")],
  })

assistantImages(turn) // ImageSource[]
assistantText(turn) // often ""

The image is a content block on the assistant’s message, { type: "output_image", source }, holding the same ImageSource as an input_image. assistantImages pulls them out in order.

Do not count on the text. These models tend to put everything in the picture, words included, and return no output_text at all.

Changing them

Replay the turn and the model edits what it drew:

const dawn =
  yield *
  gemini.turn({
    model: "gemini-3.1-flash-image",
    history: [...first.items, ask("The same lighthouse, now at dawn")],
  })

The adapter puts the image back on the wire as a model-role part, which is what gives “the same lighthouse” something to refer to.

Streaming

ImageOutput joins the TurnEvent union:

ImageOutput: { image: ImageSource; partialIndex?: number }

partialIndex marks a preview frame where a provider streams them. Gemini sends the picture whole, so you get one event with the finished image. It also lands on TurnComplete.turn, so reading only the assembled turn misses nothing.

Showing one to a different model

Only Gemini’s wire has a slot for an assistant-drawn image. Everywhere else it is dropped on replay, with a warning saying so.

To hand the picture to another model, say so explicitly:

const described =
  yield *
  claude.turn({
    model: "claude-sonnet-5",
    history: [...imagesAsInput(previous.items), ask("What is in this image?")],
  })

imagesAsInput adds a user message carrying the same pictures as input_image. It is a call you make rather than something the adapter does quietly, because “the assistant drew this” and “here is an image, look at it” are not the same claim.

Asking for pictures

There is no shared switch for this, so it stays provider-specific. On Gemini it is the model: an image model asks for the IMAGE modality on your behalf, since answering in prose would be a disappointment rather than an error. Both knobs live on GeminiRequest:

gemini.turn({
  model: "gemini-3.1-flash-image",
  imageConfig: { aspectRatio: "16:9", imageSize: "2K" },
  responseModalities: ["TEXT"], // hold it to prose
  history,
})