Does Martini train on user uploads?

Uploaded files — reference images, source video, audio, masks, control inputs — are stored against your project so the workflow can use them. They are sent to the appropriate model provider as part of a generation request, and the result is returned to your canvas. Files are retained while the project exists so you can re-run nodes, branch workflows, and edit. Deleting a project removes the associated uploads on the standard deletion schedule.

Martini does not feed your uploads into any in-house training pipeline. Martini does not have its own foundation models to train. The platform's role is orchestration — prompt routing, credit accounting, queue management, output storage — not model training. There is no internal dataset built from user uploads or generations.

Each AI model on Martini comes from a third-party provider — OpenAI, Anthropic, Google, FAL, Runway, Hailuo, and others. Each provider has its own data and training policy. Most enterprise and API-tier endpoints these providers offer are explicitly configured to not use API request data for training. Martini routes generations through API endpoints, not through consumer chat products, so the provider's API training stance applies, not their consumer-product stance.

When a provider's terms allow you to opt out of any data use beyond fulfilling the request, Martini's integration uses the opt-out path by default. If a specific provider has a different default and an opt-out is not available at the API level, that is noted on the model card for that node. Refer to the upstream provider's privacy policy for the exact terms that apply to a specific model.

We retain the operational data needed to run the platform. That includes: your project files and generated outputs (so you can keep working on them), recent prompt and job history (so you can re-run, branch, and audit), credit-spend records (so we can bill correctly), and short-lived job-queue entries during async generation (so video and other long-running jobs can complete and webhook back). None of this data is fed into model training.

Operational logs are retained for the time needed for debugging, abuse investigation, and billing reconciliation, then purged. Anonymized and aggregated metrics — total generations per model, error rates, queue latency — may be retained longer to monitor platform health, but contain no personal content.

For high-sensitivity work, use the model that best matches your privacy posture by checking each model card before generating. Delete projects you no longer need from the project menu — uploads and outputs are queued for deletion when you do. For workspace use, ask your admin about workspace-level retention controls and access policies.

Enterprise and API customers can request additional contractual data-handling terms — bespoke retention windows, regional data routing, signed processing agreements. Reach out via the help icon and ask for an enterprise contact if you need terms beyond the standard policy.