Seedance 2.0 content filter: why video edits get flagged (and how to prompt around it)
If a Seedance 2.0 Video Edit job climbs to about 90% and then sits there before failing, it was probably flagged by a content screen that has nothing to do with Vixxxen. This guide explains what triggers it, what happens to your credits, and how to prompt so it passes.
TL;DR: The Seedance 2.0 model isn't ours and we don't add any censorship on top of it — but ByteDance (the model's maker) and the hosting provider run their own content screening on Video Edit jobs, and it's stricter than the other video modes. If your edit gets flagged, it looks like a generation stuck at ~90% that eventually fails — and your credits are automatically refunded in full. This guide explains what triggers it and how to prompt around it.
What's actually filtered (and by whom)
Two layers sit between your prompt and the video, and neither one is Vixxxen:
ByteDance's model-level filter. After legal pressure from major film studios, ByteDance added aggressive filters to Seedance 2.0 — specifically around realistic human faces used as reference images and anything that reads as full identity replacement (deepfake-style edits). This applies even to AI-generated faces that never belonged to a real person.
The hosting provider's input screen. The API host runs its own automated screening on prompts and input images before the job renders. There is no toggle to disable it and no appeal — the only fix is changing your input.
We can't turn either of these off. NSFW content itself is not the primary trigger — identity replacement is.
How a flagged job behaves (important)
Flagged Video Edit jobs often don't fail immediately. The provider holds them in a review queue, so on your screen it looks like:
Progress climbs normally, then sits around 90% for a long time
After roughly 20 minutes the job fails
Your credits come back automatically within seconds — no need to open a ticket
If this happens, don't re-run the identical prompt — it will almost certainly flag again. Change the prompt using the tips below.
What we've seen pass vs. get flagged
From real jobs (same user, same reference images, same session):
Passed
"replace the woman in the video with the woman from the first image, match the motion, keep the music the same"
Flagged
"replace the woman from the input video with the woman from image 1, replace the background from the input video with the background from the reference image, match the motion and lip sync, keep the music the same"
Same faces, same clips — the difference was scope. Asking for face swap plus background replacement plus explicit lip sync pushed it from "an edit" into "full identity replacement," which is exactly what the deepfake filter targets.
Prompting do's and don'ts
Do
Keep the edit scoped: swap one thing per request (the person or the background, not both at once)
Say "match the motion" — it works and passes
If you need both a person swap and a background change, do them as two sequential edits
Use reference images that look styled or produced rather than candid ID-photo-style headshots
Don't
Ask for "lip sync" explicitly — "match the motion" covers it and doesn't trip the filter
Stack replacements ("replace the woman AND the background AND...") in one prompt
Name real people, celebrities, or trademarked characters/styles — instant flag
Re-run a flagged prompt unchanged hoping it passes — the screening is consistent
Frequently asked questions
Did I lose credits when my Seedance Video Edit was flagged?
No. Flagged jobs are refunded in full, automatically, the moment they fail. Check your credit history if you want to confirm.
Is Vixxxen censoring my content?
No. This screening happens on ByteDance's model layer and the hosting provider's input screen, and it applies to everyone using Seedance 2.0 Video Edit anywhere. Other Seedance modes — text-to-video and image-to-video — are screened less aggressively than Video Edit.
Why does it waste 20 minutes instead of failing fast?
The provider's review queue doesn't tell us the job is flagged until it's done reviewing. We're shipping an update so flagged jobs show you the real reason ("content flagged — adjust your prompt") instead of a generic timeout.
Will this get better?
The filter belongs to ByteDance and shifts over time. We'll keep this guide updated as we learn what passes. Share your own pass/flag examples in the Discord — real data is how this guide got written.
