Images and video

Choose media capabilities, manage local model preparation and inspect hosted jobs.

Mellow separates language-model conversation from media generation. An agent can discuss an image using one model and request generation using another. Select a media model for the operation you want: generation, editing or upscaling.

Local image work uses installed assets on the Mac. Hosted media requires a configured service and can send prompts and source images off the device. Video depends on the hosted path available in your installation.

Get started

Open Settings → Images and inspect the model catalog. Download a supported local model or select an available configured media service. Check its operation capabilities before starting a job.

For the first generation, use a short prompt and the model's default dimensions: “A small lilac reading nook, morning light, soft paper textures.” Wait for a saved result before increasing resolution or enabling the capability in an agent workflow.

Choosing a model

The model catalog and runtime validate these differences. A generation-only bundle cannot be made into an edit model by attaching a source image.

Making pictures in chat

Select an available image model and describe the composition, subject and output you need. Use only the controls supported by that model.

Save the model and settings with a result you want to reproduce. A seed alone does not guarantee identical output across model versions or runtime changes.

Editing a picture

Choose an edit-capable model, attach the intended source and specify what should change. Keep the request concrete: “Keep the layout and replace the red chair with a pale green chair.” Review the complete image afterward, including areas you wanted to preserve.

Use the exact edit capabilities advertised by the model. Do not assume masks, multiple references or regional editing work for every installed bundle.

Letting an agent make pictures and videos

Configure the agent's image or video capability and its target in the agent settings. This configuration is separate from the language model answering the conversation. Give a small test request and inspect the generated artifact before adding it to a routine.

Media generation may compete with the chat model for memory. The local-model swapping policy can unload a language model while media work runs, then resume the conversation. A pause during model preparation is different from a failed generation.

Cloud images and videos

Choose a hosted target explicitly and inspect the service, operation and any quoted cost or approval before dispatch. Source images and prompts needed for the operation are sent to that service.

Workspace authentication alone does not guarantee a media catalog. If catalog discovery is unavailable, configure a supported media provider or use installed local image assets.

For video, distinguish a submitted job from completed content. Check its status and retrieve the resulting artifact while it is available. A pending cloud job can outlast the UI that submitted it; closing a window is not a reliable cancellation action.

Progress and results

Local generation reports preparation and generation stages. Inspect the terminal state: completed, failed or cancelled. An intermediate preview is not the final saved result.

If cancellation was requested, wait for confirmation before starting a replacement job with the same large model. Keep track of hosted job IDs when investigating a timeout so you do not submit duplicate paid work merely because the UI stopped waiting.

HTTP API

The local server exposes image discovery and operation routes. Use the address and authentication from your server settings. Query /v1/images/models for installed IDs and capabilities before constructing an operation request.

Hosted video has a quote-and-job lifecycle; consult HTTP API for that contract. Do not turn a local request into a paid remote target as an automatic recovery step.

Troubleshooting

For dependable workflows, validate one complete operation with your actual model and provider. Then add larger sizes, agent delegation or automation one at a time.