Why Does My AI Video Face Look Different? What to Check

First identify whether the request completed

A queued or in-progress request has not produced its final checked video yet. A completed request means the workflow saved a playable result after media checks. It does not certify that every face looks exactly like the portrait or that every crowd member changed.

If the request is failed, moderated or could not pass output checks, read its status and credit accounting message before retrying. A provider moderation response may not identify whether the source, prompt or a reference triggered it, so do not assume a different crop alone will fix the request.

Check which person the photo was meant to replace

Compare the selected portrait thumbnails with the template's role labels. Two-role templates can reverse the intended cast if you swap the photos accidentally. APT performers also move between screen positions, so evaluate identity across the performer rather than one side of the frame.

Storm Dance repeats the second portrait across background people, while one-photo templates target a particular lead. If a crowd member remains unchanged, the problem is a partial replacement, not necessarily an incorrect lead photo. Check the front and back rows separately.

For the Ghostface phone-scene generator, the reference changes the phone-holder; the masked character and dog are intended to stay. The Baby Knight AI studio assigns Image 1 to the baby and Image 2 to the mother, even when they move or overlap.

The Closed on Sunday AI Video Generator repeats an optional second identity across the choir. Leaving that slot empty keeps the choir unchanged. In the Rumpelstiltskin duo dance effect, each reference follows its assigned character through shot changes; a rear view is not a new upload role.

Find your template's role mapping

Compare the actual reference with the difficult frame

A distant face, heavy filter, deep shadow or sunglasses can hide the features you want the model to preserve. Choose a sharper, single-subject reference with visible eyes and face shape if those details are unclear in the uploaded file.

Then inspect the source moment: a profile, rear view, moving hand or another person may partly hide the face. If the identity is recognizable in front view but changes on a turn, note that specific transition before deciding on another attempt. A clearer photo may help; it does not guarantee the next render.

Separate cropping from identity quality

A 9:16 crop can remove part of a head, the hands or a second performer. That is a framing decision rather than evidence that the model used the wrong photo. Compare the crop outline with the generated action and consider the original frame for a new request.

Raising resolution changes output detail and credit cost, but it does not by itself prove better likeness. Review the reference and role assignment before spending more on a higher-resolution repeat.

Review square and vertical crops

Choose the next action without duplicating a saved request

If submission timed out, check the last request or My videos first: the original request can still exist. Refreshing that result page checks it again. Avoid sending a second request merely because the first page did not immediately finish loading.

For a confirmed failed request, use the offered retry flow after checking its credit status. For a completed output, changing the portrait or settings creates a new generation with its own quote. If there is a file, sound or accounting issue, give support the existing request reference so it can be checked directly.

Check saved videos · Get help with a request