Check the reference budget first
Omni generates 3–15 seconds in whole-second steps and accepts up to six reference images. For 3–10 second outputs, use up to three reference videos, each no longer than three seconds. For 11–15 second outputs, use only one reference video, still no longer than three seconds.
The current Gemini API does not accept audio uploads or audio references for Omni. Sound inside a reference video is ignored, although the generated result can contain native audio. Two-frame start/end interpolation has a ten-second ceiling; six subject references are not the same operation as six keyframes.
Give each reference one clear job
Pick the few references that explain the shot. One image can define product appearance while a short video illustrates motion. Say which detail should come from each input, and avoid several competing versions of the same subject. Add more inputs only when they resolve a specific missing detail.
Trim a longer motion clip before uploading it for Omni. A clip that works with another model may exceed this model’s limit. A depth conversion changes the visual representation, not the duration; a long depth clip still needs trimming.
Separate a content choice from an input error
If the request is rejected, first check count, duration, media type and selected resolution. Reducing a fifteen-second clip to ten seconds may make a multi-video reference plan valid; changing the prompt cannot make an unsupported audio file valid.
Content rules remain separate. Some recognizable people or region-specific cases may be restricted by Google. A permissive configurable threshold does not remove mandatory provider protections. Review the returned error before changing inputs; do not keep resubmitting the same rejected request.
Track the job instead of starting it again
Generation is asynchronous. Waiting, processing and saving are distinct from failure. Keep the existing result card and let it update. If a task fails or its submission is uncertain, copy the diagnostic details from the card when asking support to investigate.
For Agent workflows, discover the current catalog, quote the chosen model and resolution, then create once with a stable idempotency key. For a final, use the successful Kivo Draft job as sourceJobId. The platform handles the Google interaction behind it; an Agent should not supply a Google API key or reuse a Seedance task as an Omni source.