Create more. Sell more this Black Friday.Explore
Skip to article
Conceptual portrait reference card and video frames passing through a glass API gateway
Use Cases

Seedance Human Face Detected: Fix Rejected Reference Images

Work out why Seedance rejects a face reference, separate AI characters from real-person photos, and test the appropriate Seedance 2.5 workflow on EvoLink.

JessieCOO
9 min read
If Seedance rejects a reference image because it detects a human face, first check the asset type and the exact provider route. A photo of a real person, a photorealistic AI character, and a video that generated the wrong face require different fixes. Rewriting the action prompt will not resolve an input-policy rejection.
For teams making recurring character ads or spokesperson clips, the useful next step is a supported reference workflow. EvoLink's current Seedance 2.5 image-to-video documentation explicitly supports AI-generated realistic human materials. Start with a small AI-character test in the Seedance 2.5 playground; confirm the current requirements separately if the subject is a specific real person.
This guide was checked on November 12, 2026. It explains diagnosis and evaluation; it does not report a matched-provider generation test or an approval rate.

Why Seedance detects a human face in an AI reference

“Human face detected” describes a family of symptoms, rather than one universal error string. In a community discussion about Seedance 2.5, users describe photorealistic AI characters being rejected on ModelArk and fal. Those reports identify a practical problem; they do not establish a permanent restriction across every Seedance provider.
For ModelArk, the official error-code reference identifies InputImageSensitiveContentDetected.PrivacyInformation for an image that may contain a real person; the video counterpart is InputVideoSensitiveContentDetected.PrivacyInformation. These are ModelArk identifiers, not EvoLink error codes. Use the returned message to connect the privacy error to your particular input.
Check the returned category before changing the image. The fal model-error reference marks content_policy_violation as non-retryable and distinguishes it from format and infrastructure failures. Save the error type, message, affected input field, request ID, endpoint and time. A policy category alone does not prove that a face caused the rejection; the message and context must support that diagnosis.

An AI image's origin and its eligibility on a particular service are separate questions. Writing “this is AI-generated” in the prompt does not register the asset in a provider's trusted-media system.

Choose the fix that matches your material

What you haveWhat to establishPractical next step
A fictional, photorealistic AI characterDoes this route document support for AI-generated human references?Test the documented EvoLink Seedance 2.5 image workflow with one original character image.
A photo of a specific real personAre permission and the provider's portrait onboarding requirements satisfied?Confirm the supported real-person path before submitting; an AI-character capability does not settle this.
A product or location image without a personIs the error actually about URL access, file format or parameters?Fix the indicated input issue rather than assuming a face-policy block.
A completed video with a changing faceWas the input accepted, but identity continuity lost?Simplify the shot and evaluate reference consistency; this is an output-quality problem.
BytePlus documents a distinct portrait-asset workflow. On ModelArk, direct real-face uploads are restricted; documented paths include preset digital characters, eligible trusted outputs and authorized real-person assets. Trust has account, origin and validity conditions. Do not assume an image generated elsewhere inherits it, or that a ModelArk asset ID is accepted by EvoLink.

For EvoLink, use the current API documentation as the integration contract. Its explicit AI-generated-material support gives you a concrete workflow to evaluate; it does not guarantee that every reference or finished video passes review.

Consider a fictional presenter demonstrating a plain ceramic mug. The example below is an author-created test brief, not a measured Seedance output.

1. Start with a minimal reference package

Keep the original AI-generated character image and its provenance. Pick one clear subject, a simple background and a visible face. Use that image alone for the first test, so a failure is easier to attribute. If you need to reproduce a real actor, use the real-person path instead of describing the photo as synthetic.

2. Choose the route by the task

Open Seedance 2.5 and select Image-to-Video when one image should define the starting frame. The documented route also supports a first/last-frame pair. Use Reference-to-Video when separate character, product, motion or audio references should guide new footage.
If the task is to change or continue an existing video, choose Video Edit or Video Extend. The reference-to-video documentation specifically directs editing and extension tasks to those routes.

3. Test one action before a full ad

Sign in to EvoLink, remove any prefilled demo images or saved references, and upload only your character image. Confirm that there is one input image, so an old sample does not become the first frame while your character becomes the last. Replace the demo prompt with the example below, select 5 seconds and 480p, check the estimated credits and your available balance, then generate. In an API integration, handle 401 as an authentication issue and 402 as insufficient quota; neither alone diagnoses a face rejection.

Keep the camera still and ask for one action:

The current image-to-video documentation explicitly describes Chinese and English prompts. This copy-ready example uses English; we have not verified other prompt languages for this workflow.

The fictional adult presenter in the reference image lifts a plain ceramic mug, looks toward the camera, and smiles briefly. Preserve the character's facial features, short dark hair and teal jacket. Soft daylight, one continuous shot, no on-screen text.

This wording tests continuity and motion. It is not a control for changing moderation decisions. Add dialogue, camera movement and extra references only after the basic shot is usable.

4. Check completion and save the evidence

For an API integration, follow the documented asynchronous task flow through to completed or failed. Task creation alone is not a successful video. Keep the request settings, task ID, final error or result, and actual usage. Download accepted output promptly; the documentation specifies that generated video links are valid for 24 hours.

An accepted face reference still needs an output check

Conceptual contact sheet showing the same fictional character at three moments for a Seedance face-continuity inspection
Conceptual contact sheet showing the same fictional character at three moments for a Seedance face-continuity inspection
Editorial illustration of the review method; these frames are not Seedance test results.

Compare the reference with the beginning, middle and end of the clip. Inspect face shape, hairline and clothing, then watch the full video at normal speed for identity changes during head turns. Check the hand and mug for deformation. If you add speech, review mouth timing and pronunciation separately.

Record two outcomes: did the request complete, and is the shot usable? A completed clip with a different character still fails the brief. Change one variable on the next attempt: simplify a head turn, remove a conflicting reference or shorten the action. If the failure is a policy rejection, follow the supported asset path or ask for review with the request ID instead of running an unchanged retry loop.

When switching providers helps—and when to stay

Testing EvoLink is useful when you need its documented AI-character reference workflow and want image and multimodal video routes through a unified API gateway. Check the existing Seedance 2.5 pricing section for the selected mode and references before expanding the test.

If your current provider already accepts the assets and delivers usable shots at an acceptable cost, a face-related forum post is not a reason to migrate. For a specific real person, establish the supported authorization process first. A gateway change does not itself prove that an unsupported asset becomes eligible.

Compare actual spend per accepted shot, including completed clips you discard and manual rework. This article has no matched test showing that EvoLink has fewer rejections or lower total cost than fal. For broader generation trade-offs, use the Seedance 2.5 vs 2.0 guide; the Seedance 2.0 real-human guide covers that version's workflow.
Test a Seedance 2.5 Character Reference

FAQ

Does a human-face error mean Seedance cannot generate people?

No. A reference-input restriction and the ability to generate a person are different. Check the provider, asset type and error details before deciding what to change.

Why was my AI-generated face rejected on fal?

Users report this symptom, but the exact request needs its own diagnosis. Inspect the error type, message and input location. Photorealistic appearance alone does not establish eligibility on that route.

Yes. The current Seedance 2.5 image-to-video and reference-to-video documents explicitly list that support. Evaluate your own assets; the statement is not an approval-rate guarantee.

Can I upload a real actor's photo instead?

Treat that as a separate asset category. Confirm permission and the current provider-specific onboarding requirements. Support for AI-generated people does not establish acceptance of every real-person photo.

Will changing the prompt or retrying fix the rejection?

For a confirmed input-policy rejection, use the documented asset workflow or request a review. Prompt edits can address the shot's creative direction; unchanged retries do not resolve an eligibility requirement.

Which Seedance 2.5 mode should I use?

Use Image-to-Video for a starting frame or first/last-frame pair; Reference-to-Video for references guiding new footage; Video Edit or Video Extend for an existing clip.

Is a completed task enough to approve the result?

No. Check identity through the full clip, hand-object interaction and any speech. Keep the task status and your creative acceptance decision separately.

Do I need to move my entire workflow from fal?

No. Test a representative, supported reference job first. Keep a working route unless the new route improves your measured workflow, integration burden or accepted-shot cost.

Sources

Connect AI models through one API

Compare models and current pricing, then follow the API documentation to integrate your chosen model.