Seedance 2 · Prompt Guide

How to Write Seedance 2 Prompts That Pass Filters

In production usage, many Seedance 2 rejections are false positives caused by ambiguous wording, not actual policy violations. This guide breaks the pattern into 5 practical rules you can apply immediately.

Quick note: this article is about reducing false flags and improving legitimate pass rates, not bypassing policy. Always follow Seedance rules and Creaa.ai platform guidelines.

After Seedance 2 launched on Creaa.ai, users generated a large volume of videos. One pattern was clear: many prompts were blocked not for obvious violations, but because the scene intent was under-specified and interpreted conservatively.

How the Filter Works: It Reads Intent, Not Keywords

Seedance 2 moderation behaves more like an LLM judging context than a strict keyword blacklist. The same word can pass or fail depending on framing. Your goal is not removing words, but making intent explicit.

Principle: do not fight the filter; give it a scene that is clearly understandable and non-harmful.

Tip 1: Build Safe Context Around Sensitive Elements

Isolated actions are easier to misread as threats. Place the same action inside a complete scene with environment and purpose.

Avoid

a person fires a rifle into the sky

Use

cinematic wide shot, a rider crosses a snowy valley and fires one shot into the sky as a signal, echo rolling across an empty mountain, 35mm grain

Principle: do not remove dramatic beats; add place, purpose, and cinematic framing.

Tip 2: In Image Input Workflows, Describe Roles, Not Ages

If a reference image already shows the character, you usually do not need age labels in text. Minor-related terms can trigger stricter review thresholds even for otherwise safe scenes.

Avoid

a young boy riding a horse through snowy mountains

Use

a rider on a gray horse moving through snowy mountains, striped poncho, worn leather boots, overcast sky

Principle: let the image carry identity; let the prompt carry action and scene.

Tip 3: Every Sentence Must Serve the Visual

The model evaluates one complete scene. Unnecessary backstory or unrelated references add ambiguity. Length is fine; noise is not. Structured prompts are safer and more controllable.

{
  "visual_world": {
    "light": "overcast snow light, soft diffused shadows",
    "color": "muted cold whites and grays",
    "film": "35mm grain, soft halation, 2.39:1 anamorphic",
    "atmosphere": "quiet, vast, isolated"
  },
  "sequence": {
    "duration": "10 seconds",
    "shots": {
      "shot_1": {
        "camera": "static wide",
        "action": "rider and horse beside an icy stream, still frame"
      },
      "shot_2": {
        "camera": "low angle from behind",
        "action": "horse gallops through deep snow, pines on both sides"
      },
      "shot_3": {
        "camera": "locked wide composition",
        "action": "rider and wolf face each other from distance, both motionless"
      }
    }
  }
                }
Principle: dense visual information beats narrative fluff. Every field should describe something the camera can see.

Tip 4: With Image Inputs, Faces Are the #1 Rejection Trigger

Many of these rejections happen before your text is interpreted. If uploads fail repeatedly, the issue is often the image itself, not wording.

  • Prefer back views, wide shots, or compositions where facial features are not identifiable.
  • If the subject must stay visible, reduce face recognizability via angle, occlusion, or distance.
  • If photos fail repeatedly, illustrated references may pass more reliably.

Tip 5: Use Cinematic Language as a Safety Anchor

Camera direction, lighting, lens/film texture, and aspect ratio help the model classify the request as creative production context. That reduces ambiguity versus real-world threat interpretation.

Avoid

a person on a horse fires a gun in the mountains

Use

cinematic wide shot, 2.39:1 anamorphic, overcast diffused light, a rider raises a rifle and fires once into the sky as a signal, smoke drifting in cold air

Principle: cinematic language is not decoration; it is a context anchor.

30-Second Pre-Submit Checklist

  • Did you specify purpose and scene context, not just a sensitive action?
  • Did you remove unnecessary age labels and identity tags?
  • Can each sentence be seen directly by a camera?
  • Does the reference image contain a detectable face likely to fail pre-check?
  • Did you add camera, lighting, or aspect-ratio anchors for creative context?

The Core Idea

Seedance 2 moderation is fundamentally scene interpretation. If your prompt reads like shot direction, it is more likely to pass; if it reads like ambiguous threat text, it is more likely to be blocked.

Compliance note: This guide is for reducing false positives and improving legitimate creative workflows. Do not use it to produce harmful, deceptive, or policy-violating content.