Answer: Seedance reviews a job twice. Layer 1 screens the prompt before any task exists — a clear violation is refused and nothing is charged, while uncertain prompts are allowed and recorded. Layer 2 is the selected model's own safety review, which is stricter: it currently refuses real-person depictions, branded or trademarked content, and political material on the standard route, and it can also fail a job after generation if the frames themselves violate policy.
Knowing which layer refused you tells you what to change. A layer-1 refusal is a wording problem. A layer-2 refusal is a category problem — rewriting adjectives will not help; you need a different route or a different subject.
The two layers, side by side
| Layer 1 · prompt screening | Layer 2 · model safety review | |
|---|---|---|
| When | before the task is created | before and during generation |
| What it looks at | the prompt text | prompt, references and generated frames |
| Clear violation | refused, nothing charged | job fails, credits refunded after reconciliation |
| Uncertain case | allowed, recorded for audit | may pass or fail downstream |
| What to change | wording | route, subject or references |
Neither layer is a local keyword list. Our screening lets uncertain prompts through and records them, which is why the same sentence can pass on one route and fail on another.
What layer 2 refuses on the standard route
- Real-person depictions. The published page lists them as prohibited, and in practice the standard route still passes most ordinary real-face jobs — it is the model's review being stricter than the brief. When it refuses, the ultra-wide route is the fallback.
- Branded or trademarked content. Brands, franchises, studios, games and licensed characters are likely to fail. Generic visual descriptions survive.
- Political material.
- Generated frames that violate policy. The prompt can pass and the render still fail, because the check also sees the pictures.
The two errors you will see, and what they mean
Your content violated community guidelines.
That is a guideline refusal. It is not a billing event: a prompt-level refusal creates no task and charges nothing; a job that fails later is refunded automatically after reconciliation, and the activity page shows a refund note on the failed task.
Video generation failed due to a server error. Please try again.
That one is not a policy decision at all — it is a temporary model-service failure. Retrying is the correct response, and it is also the message people most often misread as a ban.
Rewrites that change the outcome
| Refused | Why | Rewrite |
|---|---|---|
| A photo of a real client | review refused the person | retry on the ultra-wide route, or use a generated face |
| "in the style of [franchise]" | trademarked property | describe the visual: palette, lens, texture, era |
| A named product on the bottle | brand reference | "unlabelled amber glass bottle" |
| A crowd scene with a real politician | political content | remove the figure, keep the composition |
| Explicit prompt | sexual content | cut the act, describe framing and wardrobe |
The pattern in every row: replace the identity the model is not allowed to reproduce with a description it is allowed to render.
What good behaviour looks like in practice
- Start generic, then add specificity. Pass the scene first; add the reference second. When something fails you will know which addition did it.
- Keep a note per rejection. "Brand name", "real face on standard", "visual layer" — three different fixes that people otherwise treat as one.
- Simplify on the second failure. Repeated refusals on a long prompt are usually one clause; shortening the prompt finds it faster than editing adjectives.
- Do not retry a policy refusal. A refusal is deterministic. A server error is not — those, retry.
Where this is genuinely annoying
- Passes screening, fails at render. You paid the wall-clock time and got a refund, not a clip.
- Nothing local to test against. The rules are the model's, so a prompt that worked last week can fail today if the review changes.
- Ultra-wide is not immunity. It is a broader path with the same platform limits, and it costs more.
FAQ
Do refusals cost credits? Not on layer 1. On a later failure, credits are refunded automatically after the task is reconciled as failed.
Is the ultra-wide route unfiltered? No. It is a wider review path; the acceptable-use policy still applies and requests can still be refused.
Can I turn the safety check off? The generator exposes an optional pre-generation content check, but turning it off does not remove the platform's own rules — it only stops sending that optional flag.
Why did a prompt that worked before stop working? Because these checks belong to the model and can change. Treat prompt behaviour as dated, like a price.
The published rules are at the Seedance community guidelines, the wider route is described at the wide-limit page, and the prompt library — 122 prompts written to survive the review — is at the Seedance prompt atlas. If a refusal here contradicts the published rules, that is the bug report worth sending.
Top comments (5)
The two-layer split maps almost exactly to how we think about safety in a desktop AI client: a cheap local policy check on the outbound prompt (your Layer 1) catches the obvious stuff before we spend a token, and a separate server-side model review (Layer 2) catches what the local list cannot. The trap we hit in production is exactly the one you called out — a guideline refusal and a transient 5xx look identical in the UI, and retries on a hard refusal just burn quota while blindly aborting on a server error loses good jobs. Your per-layer 'what to change' table is the kind of runbook we should have shipped two versions ago.
A guideline refusal not being a billing event is worth stating plainly; the refund-after-reconciliation note answers the 'do I lose credits' panic.
The layer-1 vs layer-2 split is the most practical framing I've read: wording problem vs category problem tells you exactly what to change.
Real-person, branded and political as layer-2 categories matches what I've seen; ultra-wide only helps the first one.
The same sentence passing one route and failing another is the part people don't expect from a 'review' — it really isn't a keyword list.