Writing prompts the model reads as cinematic.

The filter is a language model reading your whole prompt as a scene — not a keyword scanner.

Most flagged prompts aren't flagged because of a specific word. They're flagged because the prompt didn't give the filter enough context to read the scene as cinematic. Once you understand how it reads, the fixes become obvious.

Two layers run independently. The image-evaluation layer runs first and rejects on the upload itself: real identifiable faces of public figures + named copyrighted characters are image-stage hard blocks — no prompt rewrite gets past them. The text filter runs second and reads the entire prompt as one scene, judging intent and context. Words that look sensitive in isolation pass when wrapped in clear cinematic framing; harmless words get flagged when the prompt is too sparse to interpret.

Frame the whole scene, not just the action

The most common shape of a flagged prompt is one action with no surrounding context. Build outward from the action and answer four questions in the prompt:

1. Where is this happening?
2. What does it look like visually?
3. What is the camera doing?
4. What is the overall atmosphere?

Flagged · sparse, contextless

a soldier shoots someone in the street

Passes · same action, full scene context

Wide shot, war-torn Eastern European street in the 1940s. A
soldier in a grey uniform fires toward an off-screen position
during an active firefight. Smoke rises from collapsed buildings
in the background, overcast flat light, 35mm grain, documentary-
style handheld framing, debris scattered across the foreground.

Visual facts, not narrative

Strip emotional motivation, backstory, and relationship context from the prompt. The filter cares only about what the camera would see if this scene existed. A screenplay has scene description and subtext; Seedance only needs the description. Before keeping any sentence, ask: "if this were a real shoot, would this line appear on the shot list?" If not, cut it.

Production language is a context signal

Two or three production-language terms in a prompt establish the register and meaningfully shift how the filter weighs the rest. Pull from each category:

Multi-shot: native JSON

For multi-shot sequences, Seedance accepts a structured JSON prompt. A visual_world block sets the cinematic register once; each shot then only describes what the camera sees at that moment. This forces visual-fact discipline automatically.

{
  "visual_world": {
    "light": "soft overcast, diffused shadows, no hard edges",
    "color": "muted naturals, cold whites, desaturated tones",
    "film": "35mm grain, anamorphic lens, soft halation on highlights",
    "atmosphere": "quiet, isolated, expansive"
  },
  "sequence": {
    "duration": "10 seconds",
    "pacing": "slow build to rapid cuts, ends in stillness",
    "shots": {
      "shot_1": {
        "duration": "3 seconds",
        "camera": "locked off wide shot, low angle",
        "action": "Lone rider on horseback crests a snowfield ridge",
        "transition": "SMASH CUT"
      },
      "shot_2": {
        "duration": "4 seconds",
        "camera": "tracking shot from behind, handheld feel",
        "action": "Horse and rider gallop through deep snow, cloak whipping in wind",
        "transition": "SMASH CUT"
      },
      "shot_3": {
        "duration": "3 seconds",
        "camera": "static wide, fully locked off",
        "action": "Empty snowfield, a wolf stands motionless on a distant ridge"
      }
    }
  }
}

Reference uploads: cite with purpose, every time

An uploaded image is not a self-explanatory artifact. The model assumes nothing about its role unless the prompt assigns one. Two equivalent syntaxes work depending on host:

• Replicate / fal canonical: [Image1], [Video1], [Audio1]
• Morphic / Dreamina UI: @Image 1, @Video 1, @Audio 1 (the picker injects the tag for you)

Either way, list every role assignment at the top of the prompt, before any scene description:

[Image1] as the first frame.
Reference all camera movements from [Video1].
Character appearance based on [Image2].
Use [Audio1] as the background score.

[scene description follows…]

Character-image best practice: let the image do the work

Don't re-describe a referenced character in text — that creates a second, competing identity layer the model has to reconcile. The image handles appearance; the prompt handles what happens and what the camera sees. Two important consequences:

• Refer to characters by role, not by age. Words like child, kid, young, boy, girl raise the sensitivity threshold across the entire prompt regardless of what the uploaded image actually shows. Use "a small figure in a dark coat" instead.
• If image-stage block is firing, work the image: face away from camera; go wide enough that the subject reads as silhouette; swap photo for illustration (illustrated images pass more reliably); use the image for wardrobe/setting/palette rather than face/identity.

Negative prompts that actually work

Negative prompts don't unstick filter flags — but they reliably reduce visual artifacts. Keep them short and tied to actual failures you're seeing:

negative: no jitter, no warping, no flickering, no identity drift
negative: no text morphing, no garbled logos, no color shift
negative: no motion blur on face, no floating limbs, no background collapse

Long lists backfire. Two or three targeted terms outperform exhaustive enumerations.

The Chinese-language pass-rate experiment

Community pattern reported by some Morphic users: write the scene description in Chinese, keep dialogue and on-screen text in English. The reasoning is that Seedance was developed with strong Chinese-language training, so Chinese prompts may hit slightly different filter thresholds. Not guaranteed; low effort to try if a well-constructed prompt keeps getting flagged.

Most of this section synthesizes Morphic's "Why your Seedance 2.0 prompts keep getting flagged" guide, validated against our empirical findings on Replicate and fal. Their full piece is the cleanest single reference on the filter's behavior.