Do failed AI generations restore credits?
Short answer
Yes. When a generation fails on the provider side, Martini automatically returns the credits to your account. The refund posts back to the same balance the credits were spent from — subscription quota first, wallet second — and the failed run is excluded from your usage history's billable total.
When credits are refunded
Refunds happen automatically for provider-side failures. That includes model server errors, queue timeouts, content-safety rejections issued by the model after credits were charged, malformed responses from the upstream API, and any internal Martini error during job orchestration. You do not need to file a request — the credits return as soon as the failure is detected, usually within seconds.
If a generation finishes but the output is low quality or not what you expected, that is not a failure and credits are not refunded. The model successfully ran and the provider charged Martini for the compute. If you are unhappy with the output, adjust the prompt or settings and run again. The platform's job is to deliver the model's output, not to judge whether the output meets your taste.
When credits are not refunded
User-side errors that are caught before the model is called do not deduct credits in the first place — there is nothing to refund. That includes invalid prompt formats, missing required inputs, file uploads that fail validation, and disconnected nodes. The Run button blocks the request and no charge is made.
A generation that completes but is rejected at your end (you delete the result, the workflow downstream errors, you cancel the project) is still considered a successful generation and the credits stay deducted. The provider already ran the compute. Cancellation only stops future steps from running.
Partial completion and batch jobs
When you generate a batch — for example, four images at once — each output is billed and refunded independently. If three succeed and one fails, you are charged for three and refunded for one. The same logic applies to multi-shot video workflows: each shot is billed and refunded on its own.
For queued video jobs that take minutes to return, the credit deduction happens up front when the request enters the provider queue. If the job later fails or times out, the credits are returned at that point. Check the job status in the node — a failed status will be paired with a refund entry on the Billing page within a minute or two.
Examples
- A Veo video request that times out on the provider returns credits automatically.
- A 4-image batch where 1 image fails refunds the credits for that 1 image only.
- A content-safety rejection from a video model after submission returns credits.
- An invalid prompt that is blocked before sending does not deduct credits in the first place.
- A generation that finishes but is later deleted from the canvas keeps the credits deducted.
Edge cases
- If you do not see the refund within a few minutes, refresh the Billing page and check the recent activity row for that job.
- Workspace generations refund to the workspace balance, not to your personal balance.
- Manual retries of a failed generation are billed as new runs — the refund and the retry are separate transactions.
- If a model returns a watermarked or partial output instead of a clean failure, that's a successful generation and is not refunded.
What to do next
- Check Settings, Billing for the refund entry if you don't see your balance updated.
- See the why-video-generation-failed article for diagnostic steps before retrying.
- Contact support from the help icon if a refund hasn't posted within an hour of a confirmed failure.
Related help articles
Related docs
Still need help? Contact support.