A completed AI video can be technically usable and still fail your casting brief. Check whether the actual people, visible count, and actions meet the agreed requirements before accepting it. If ten distinct surrounding characters are essential, a handsome scene with duplicate faces or extra people is a failed result for that task, even when the file downloads and plays normally.
What does a completed job actually prove?
Completion establishes a delivery state. Confirm that the file exists, decodes, and contains the intended duration and dimensions. Those checks do not prove that the model preserved a particular person, a fixed crowd size, or each character's assigned action.
Our documented single-run H3 test obtained a real output after one submission. The requested inputs included one fictional lead and a separate reference of ten fictional surrounding adults. The output broadly retained the lead's clothing and a group action, but repeated appearances and extra crowd members remained visible. At eight seconds, the reviewer could distinguish at least sixteen surrounding adults; that was a visible lower bound, not an exact scene-wide count.
The test therefore passed the API pipeline and failed the full cast contract. It does not estimate how often another run would succeed. It also did not demonstrate a public website workflow with two image uploads.
How do I write a cast contract I can inspect?
Use observable anchors instead of asking for an unspecified crowd to be perfect. A worksheet can begin like this:
| Requirement | Evidence to inspect | Result to record |
|---|---|---|
| One lead remains central | Wide and medium views throughout the clip | Present, missing, or uncertain |
| Ten distinct surrounding adults | Clearest unobstructed views | Count or visible lower bound |
| Lead stays upright while group bends | Shared timestamps during the group action | Independent, synchronized incorrectly, or hidden |
| No recurring cloned appearance | Hair, clothing and facial features across positions | Observed duplicate or not established |
Make the contract proportionate to the task. A generic party atmosphere may allow an approximate crowd. A personalized film representing named people needs stricter review and appropriate consent.
Why is counting one frame insufficient?
People can enter, disappear behind others, or reappear with changed features. Inspect the beginning, the widest view, and moments when bodies cross. Record what is visible rather than guessing the number behind an obstruction.
A matching jacket does not prove a stable identity. Compare several anchors and the person's position across time. Keep uncertain observations separate from confirmed defects. Face recognition scores would require a justified method; this review used direct visual observation, not a calibrated identity metric.
Also check role behavior. Two people can look different while performing the wrong assigned actions. In the test, the central lead largely stayed upright when the group bent, but that limited success did not correct the crowd size or the choreography timing.
Should I crop, edit, or generate again?
Crop only if the resulting composition still serves the story and hides the defect throughout the relevant interval. A duplicate that moves into the retained area is not repaired by a successful first-frame crop. Check the final export after any reframing.
If a named person or an exact group is essential, simplify the generation contract. Consider separate character shots or controlled editing instead of demanding simultaneous identity, count, dance, and complex camera changes. Confirm the selected tool's supported inputs before planning that approach; a prompt does not create a missing multi-character control feature.
What should I tell the person receiving the video?
Describe the result according to the checks it passed. Call it a generated interpretation when casting or movement departs from the source. Do not present a fictional crowd as a record of real attendance.
Keep the cast worksheet with the accepted version and note unresolved observations. The original test media are not republished with this guide because a public reference used for internal evaluation does not by itself establish commercial publication rights. The evidence record preserves the measured delivery facts and the review limits without turning one output into a product guarantee.