How to Write Prompts to Generate Uncensored AI Videos
How to write prompts to generate uncensored AI videos on Seedance 2.5: what trips error 80006, which details belong in reference images, and how to draft.
Knowing how to write prompts to generate uncensored AI videos on Seedance 2.5 starts with where moderation runs. It runs on the prompt text, on every reference file, and on the finished video, and the content_filter field routes a request between two channels without switching any of that off. No phrasing changes those checks. What prompt craft does change is how often you trigger a refusal you never intended, which accounts for a large share of failed generations.
How to write prompts to generate uncensored AI videos without tripping a refusal
Every safety refusal returns error 80006, and the code does not say which stage fired. Three stages can raise it: the text, an attached file, or the video that came out the other end. A prompt can be entirely clean and still produce 80006 because the generated scene was the thing refused.
Three categories are blocked inside the model and no setting reaches them:
- named real celebrities
- third-party intellectual property
- illegal content
Everything else is a moderation decision made per request. Realistic people are supported and need no flag. Failed jobs are refunded in full, so iterating on the text costs time rather than credits.
Names are the most reliable way to get refused
The celebrity block is written against names, which makes proper nouns the first thing to audit in any prompt.
A living actor, a musician, a politician, a public figure of any kind: naming one refuses the request before anything is generated. The intellectual property block is wider and does not need a name at all. A character, a film title, a game, a recognizable costume, a vehicle silhouette, a logo described closely enough: all of those can trip it while the prompt looks like plain description.
Both blocks are removable from the prompt in the same way. Replace the name with what it was standing in for. Age, build, hair, wardrobe, posture, expression and accent are description, and description is not what either block targets.
Describe what the camera sees
Concrete visual detail gives the model less room to invent something the output-stage check refuses. Abstract mood words leave it a great deal of room.
Work through subject, action, setting, light and camera behavior. Each of those can be checked against the result, which makes a second run purposeful instead of a reroll. A prompt that says a figure turns toward a window in low afternoon light while the camera holds still produces a predictable frame. A prompt that asks for something provocative delegates the whole scene to the model, and whatever it decides on is what moderation then judges.
Keep the action count honest against the duration you set. A single generation runs 4 to 30 seconds and can hold several shots, so write longer prompts as beats and say where the cut falls. Five seconds carries one action. Four actions compressed into five seconds arrive as noise, which then costs another run.
Push the specifics into reference images
Anything visual and exact belongs in a file rather than a sentence. One request accepts up to 30 reference images in jpeg, png, webp, bmp, tiff, gif, heic or heif, each under 30 MB, and they add no billable seconds.
Reference video is available too, up to 10 clips in mp4 or mov at 200 MB or less, 2 to 30 seconds each and 30 seconds combined, but it does add billable seconds, so reach for images first. All media is supplied as public HTTP(S) URLs; data URIs are rejected platform-wide.
Reference files are moderated on the same terms as the prompt. A photograph of a recognizable public figure refuses at that stage and returns the same 80006. With 30 images attached and no indication of which one caused it, split the set in half and resubmit until the file surfaces.
Once images are attached the prompt becomes optional and its role narrows to action, camera and timing. Delete the sentences that duplicate what a photograph already shows, since two sources for one detail is how the written version quietly overrides the visual one.
Quoted lines become speech
Audio generates in the same pass as the picture, covering speech, effects and music as a mono track. Anything inside quotation marks comes back spoken.
Write the line exactly as it should be heard, then time it out loud against the duration field, because a forty-word speech does not fit into five seconds. When two characters speak, attribute the lines in the surrounding sentence so the pairing is unambiguous. Sound that is not dialogue gets described rather than quoted.
Dialogue is also checked. A refusal on a request whose visuals are unremarkable is worth re-reading with the quoted lines isolated. For a silent clip, keep quotation marks out of the prompt entirely, or set generate_audio to false.
Draft at 480p, ship once
Prompt work is structural, and structure is fully legible at the lowest resolution. Generation without a reference clip costs $0.118589 per second at 480p, $0.266824 at 720p and $0.461856 at 1080p.
That puts a five-second draft at about $0.59 against roughly $2.31 for the same draft at 1080p. Iterate at 480p using the duration and aspect ratio you intend to ship, then run the settled text once at the resolution you want. Generation averages two to five minutes either way and polling for the result is free.
The prompt field holds up to 20,000 characters, so there is no reason to compress. Duration, resolution, aspect ratio and audio are separate fields and cannot be set by writing about them; asking for a vertical video in prose does not select 9:16.
Frequently asked questions
Can a prompt bypass the content filter?
No. Moderation runs on the text, the reference files and the output, and content_filter selects a channel rather than disabling any of it. Prompt craft avoids unintended refusals and does not remove the model-level blocks.
Which words trigger error 80006 most often? Proper nouns. Names of real public figures and anything identifying third-party intellectual property are refused before generation starts, and they usually appear in prompts that the writer considered harmless.
Should details go in the prompt or in a reference image? Exact visual details belong in reference images, which accept up to 30 files per request and add no billable seconds. The prompt is better spent on action, camera and timing.
Does a refused prompt cost anything?
No. Failed jobs are refunded in full. Check usage.credits on the failed task to confirm the reserve was returned.
The short version
How to write prompts to generate uncensored AI videos reduces to a few habits rather than a phrasing trick. Strip proper nouns, because names carry the two blocks that no setting removes. Describe what the camera sees so the output-stage check has something predictable to judge. Move exact visual detail into reference images, which are free, and keep the prompt on action, camera and timing. Quote only the lines meant to be spoken, match the action count to the duration, and iterate at 480p where five seconds costs about $0.59.
