Why did my AI video generation fail?

Video models run on third-party GPU clusters with their own queues and capacity limits. Transient provider errors — 5xx responses, queue overload, model timeout, internal provider errors — are the most common failure cause. The error message will say something like Provider error, Timed out, or Internal error. These are not your fault and credits are returned automatically when the failure is detected.

The fix is almost always a retry. Wait a few seconds and click Run again. If retries keep failing on the same model, try a different model — provider outages tend to be model-specific, and Martini supports multiple video models so you can route around an outage. Status reports for major outages are posted in the help icon notifications when they occur.

Each video model provider runs its own content-safety filter on prompts, source images, and generated outputs. If the prompt or input asset trips the filter, the request is rejected with a Content policy error or Safety filter error. Common triggers include depictions of real public figures, named brand or character likenesses without context, sexual content, graphic violence, and politically sensitive scenes.

The fix is to soften the prompt and rerun. Replace named people with generic descriptions (a person, a character), remove brand and IP references, and rephrase any scene description that could be read as violent or sexual even when innocent. If the source image is the trigger, swap to a different reference. Credits are typically refunded when the rejection is at the provider level.

Image-to-video and reference-driven workflows fail when the source asset doesn't match what the model expects. Common asset issues: unsupported format (HEIC, BMP), resolution outside the model's accepted range, extreme aspect ratios that don't match any preset, file corruption, or oversized files that exceed the upload limit. The error usually names the field and the constraint, like Image too large or Aspect ratio not supported.

Convert the source to a common format (JPG or PNG for images, MP4 for video), check the aspect ratio against the model's supported presets shown in the node, and resize down if the file is too large. The node panel lists the constraint for each input — match the constraint and rerun. These failures generally do not deduct credits because they are caught before the model is invoked.

A few failure modes come from prompt or parameter combinations the model cannot handle. Empty or very short prompts on text-to-video models, durations outside the supported range, resolutions the model does not produce, or conflicting parameters (for example, requesting a still-frame mode and a motion strength at the same time) all surface as Bad request errors.

The node panel shows the supported parameter ranges. Reset to defaults and edit one parameter at a time to isolate the issue. Add detail to the prompt — most video models perform better with descriptive 1-3 sentence prompts than with single-word prompts. If the same prompt works on a sibling model, the issue is parameter-specific to the failing model rather than a content issue.

If a generation fails repeatedly with the same error after you have addressed the cause described in the error message, capture the project ID and the failing job ID and message support from the help icon. Include screenshots of the error, the model used, and the inputs. Support can check the upstream provider response and confirm whether the failure was provider-side (credits refunded) or whether something in the request needs adjustment.

For account-wide failures across multiple models, check the help icon notifications for any active platform-level incident. If nothing is reported and the issue persists, contact support — repeated cross-model failures are a sign of an account-level configuration issue worth investigating directly.