Media¶

Kiwi agents create and understand media through Media Studio. Media can be part of a larger conversation: develop a concept, generate an asset, review it, and ask for a variation without starting the brief again.

Available media tools¶

Task	Tool
Create or edit an image	`generate_image`
Create or edit a short video	`generate_video`
Generate music	`generate_music`
Generate a sound effect, ambience, or sonic logo	`generate_sound_effect`
Generate speech from text	`generate_speech`
Describe an image or video	`caption_media`

The models and options available for each task can change. Check Supported models for the current documented catalogue.

Generate and iterate¶

State the purpose, audience, format, and important constraints in your request.
Ask the agent to generate the asset. It chooses an enabled media tool and sends the necessary prompt and source material to Media Studio.
Review the result rather than treating it as final. Check accuracy, rights, representation, brand requirements, and suitability for the intended channel.
Refine by describing the change. Images can use source images for edits or visual guidance; supported video flows can use image or video guidance.

The visual editor can also apply annotated, region-specific instructions to an existing image. This is useful when you want to preserve most of an approved image and change only selected areas.

When you omit generation settings¶

Kiwi applies documented model-specific defaults when a request does not specify settings such as image shape, image size, video duration, video resolution, sound-effect duration, or speech voice. This makes repeated requests more predictable than relying on a provider's changing defaults.

Explicit instructions take priority. Products can also supply request-level media defaults between your instructions and Kiwi's defaults. The final value must still be supported by the selected model.

Some source-guided modes preserve the source instead of forcing a scalar default:

Google and OpenAI image generation can derive shape from conditioning images when no aspect ratio was requested.
Firefly image edits derive dimensions from their reference images.
Omni video follow-ups derive shape from the previous interaction.
Visual Edit preserves the original source framing.

See Supported models for the current default settings.

Inputs, outputs, and links¶

Prompts and any selected source media needed for the operation are sent to Media Studio. Turning off web search does not prevent this.
Generated assets are stored in Kiwi as files so they can be reused and associated with the conversation.
Media links are time-limited signed links. Treat the underlying asset according to its classification even though the link expires.
A generated output can be plausible but wrong or inappropriate. Human review is required before publication or consequential use.

See Security and data handling before using confidential source media or content subject to client, licensing, or personal-data restrictions.