Seedance likeness filter: what works (and what does not)
September 26, 2026
Why Seedance rejects photoreal faces, why prompt tricks fail, and the legitimate paths — fictional characters, trusted URLs, T2V, and Kling.
If you generate talking-head video with Seedance, you may hit an error like:
The images or videos provided may contain likenesses of real people or other private information that cannot be processed.
This is Seedance's likeness / privacy precheck on input images. It is not fixed by rephrasing "make a video of this person" to "make a video of me."
What the filter is checking
Seedance image-to-video asks: does this input look like a real identifiable person without provenance?
That includes:
- Uploaded photos of real people
- AI portraits that read as camera photographs
- Third-party faces with no trusted chain of custody
Changing the motion prompt does not vouch for the still. The rejection happens before generation.
What does not work (and why)
| "Hack" | Reality |
|---|---|
| Crop the face smaller | Still often classified; policy risk |
| Blur or pixelate | Unreliable; degrades output |
| Contact sheet / grid layout | Unreliable; against guidance |
| Swap pronouns in the prompt | Precheck is on the image, not wording |
IQON's docs treat these as fail-closed. Do not build a production workflow on them.
What does work
1. Fictional character framing (text still)
Describe a specific fictional face in an ordinary room — pores, asymmetry, front-camera framing — without claiming "real person" or "photograph of a celebrity."
Seedream text-to-image can produce a still tuned for this. IQON omits phrases that trigger CGI or "framed painting" artifacts.
2. Trusted URL pipeline (Seedream → Seedance I2V)
When Seedream returns BytePlus trusted URLs (return_byteplus_urls: true), Seedance can accept that still for I2V because provenance is attached.
Critical detail: re-hosting the image breaks trust. IQON sends Seedance the trusted URL, not a copied file in Supabase storage.
Trusted URLs expire (~24 hours). Retry must mint a fresh still.
3. Text-to-video fallback
If I2V fails or Seedream flags the face, describe the same iPhone front-camera scene in prose with no input image. Same fictional face, same delivery — no likeness upload to reject.
4. Use a different video model
Kling 2.6 Pro image-to-video does not share Seedance's likeness filter. IQON's current influencer test: Nano Banana Pro character → first frame → Kling I2V with quoted speech.
A photoreal Nano Banana face Seedance rejects may still work as a Kling start frame.
5. Stock or cleared characters
Use stock actors or uploaded characters scoped to your brand — faces intended for generation, not random web portraits.
Practical workflow on IQON
- Pick Influencer video — orchestrator tries Seedream + Seedance, with T2V fallback.
- Or rely on the Kling UGC path when active for onboarding talking heads.
- On failure, retry — fresh still, do not reuse stale trusted URLs.
- Upload your character only when you have rights to that likeness.
Bottom line
You are not "bypassing a filter." You are choosing inputs Seedance can legally process: fictional or cleared faces, trusted pipeline outputs, or a model without that precheck.