Seedance face detected: the route from our tests

Repeated face-detection refusals call for a production decision before another prompt revision. Preserve the exact refusal and input context, check the selected service’s rules, and decide whether a permitted portrait-and-voice route or separate coverage can serve the shot. This guide keeps our dated provider observation separate from official Seedance release documentation. The model selection guide places the alternative production routes alongside their separate review responsibilities.

What the refusal actually tells you

In our production in October 2026, Seedance 2.5 through one API provider refused photoreal faces in the start image, reference images, and reference audio. The recorded message described sensitive input. That establishes what happened in that integration at that date. It does not establish a universal limit across Seedance versions or services.

ByteDance’s ByteDance Seed: Seedance 2.0 lists multimodal inputs for Seedance 2.0. The ByteDance Seed: Seedance 2.0 launch discusses authorization or identity verification for real human portrait references. Those are maker statements about that release, separate from our provider-specific observation. Start with the policies and supported inputs of the service you are actually using.

Keep a useful refusal record

Save the request identifier, model identifier, provider, date, input type, and exact error. Keep credentials outside the record. A thumbnail can help identify the approved input, but the diagnostic log should not become a new collection of personal images.

Distinguish rejection from a job still running, a network interruption, or a missing output download. Our production rules check job status before resubmission because an interrupted client call does not prove the remote job ended. After a clear content refusal, repeating essentially the same input is not a new production hypothesis.

Source: suggested production checklist, October 2026.

Split the job by what the shot needs

A talking head needs identity, understandable speech, and mouth movement. B-roll needs an action that supports the spoken sentence. Combining those needs in one generation is a choice, not a requirement of the finished ad.

Our alternative keeps the talking head on a route built around an approved portrait and a fixed voice file, while Seedance supplies shots without recognizable faces. That preserves the message without treating the rejected portrait as something to disguise. If the person must be real, filming them with permission is another production option.

A refusal should leave enough context for a production decision: which model and service were selected, what input was submitted, and what the response actually said. Keep that record without reproducing a provider interface or claiming a universal portrait rule. Check the selected route’s restrictions before choosing another permitted approach, and keep official Seedance release facts separate from the dated integration observation discussed in this article.

1. Portrait
2. B-roll
3. B-roll

Approve the portrait and voice independently

Check the original portrait before making a talking head: identity, framing, clothing, and expression should fit the role. Keep its full file identity in the job so a similarly numbered image cannot silently substitute for it. Keep the same person across AI ad shots describes reference discipline across shots.

Approve the spoken script and a short pronunciation sample separately. A plausible face cannot repair a wrong product name or an unsupported statement. Fix AI voiceover pronunciation before rendering covers that audio review. The person on screen should never be presented as a real buyer describing an invented experience; a fictional spokesperson needs an honest role in the story.

Make b-roll that fits the separation

Build a shot list from the voice file rather than asking for a new scene wherever a talking head is unavailable. An exterior, a product detail, or a diagram may carry the sentence more clearly than another person. Choose imagery that genuinely relates to the explanation.

For an illustrative household-product ad, the narrator could explain a folding action while a real product demonstration supplies the evidence. A clay room could establish a setting, but it should not stand in for proof of how the actual product folds. The separation is about production roles, not permission to invent product behavior.

1. 01Record the refusalProvider, model, input role
2. 02Check the service rulesConfirm permitted inputs
3. 03Approve a separate routePortrait plus fixed voice
4. 04Review the full cutIdentity, joins, captions

Review the joins, not just the generated shots

Watch the cut where a talking head gives way to b-roll. Does the sentence continue smoothly? Does a change in clothing suggest that the narrator has become a different person? Does an unrelated hand action appear to demonstrate a claim the footage cannot support?

Listen to the audio transition as well. Generated clip sound should not introduce an extra speaker beneath the approved voice. ElevenLabs music for ads: plan the full music bed explains planning a continuous music bed. Keep the voice as the timing reference and rebuild captions after any audio change. A finished edit can look coherent while its words and mouth movement disagree.

Unhelpful conclusionUseful production decision

• Unhelpful conclusionThe whole model rejects peopleUseful production decisionThis request was refused here
• Unhelpful conclusionRetry the same inputUseful production decisionChoose a permitted route
• Unhelpful conclusionMerge portrait and b-roll needsUseful production decisionGive each shot a clear role

Make the next test a bounded decision

If you later revisit the model, make that a new test with a written reason: a newly documented supported route, a different authorized input category, or a changed provider version. Specify what result would justify using it. Do not silently generalize the old refusal or announce that the restriction has disappeared.

Keep a permitted alternative in the brief so the project does not depend on a disputed input. Turntake’s models plan marks our planned connection as Coming. The method described here runs in production workflows; it is not a claim that Turntake currently offers self-service face generation or a refusal-removal feature.

What to hand the editor

Deliver the approved voice, portrait reference, permitted talking-head output, b-roll sources, and refusal note as separate items. The refusal note explains why the shot plan changed; it is not a prompt for someone else to try the same rejected route again.

Add the source of each visual claim and identify any fictional imagery. Preserve the original approved files alongside the new version. The next editor should be able to replace one b-roll shot without reopening the portrait decision, regenerating the voice, or losing the reason for the production split.

Who wrote this We made the ads first, for SmartSolar in Denmark, Sweden and Ireland. Then we built Turntake. Where Turntake comes from