# ANOMALY Knowledge · Seedance 2.0 reference

**The default animator.** You never generate video from text — you animate a still you already approved. That is how the character stays identical across a whole video.

**Model id on Higgsfield: `seedance_2_0`.**

---

## Settings

| Setting | Value | Why |
|---|---|---|
| Start image | The approved still | Never text-to-video |
| Aspect ratio | `9:16` | |
| Resolution | `1080p` | |
| Duration | `4`–`5` seconds | Videos are built from several short beats, not one long generation |
| Mode | `std` | |
| Bitrate | `high` | |
| **Enhance** | **OFF** | Enhance silently rewrites your prompt and causes drift. This is the most common mistake |
| Concurrency | Max ~8 jobs at once | Queue beyond that |
| Runs per prompt | 2–3, pick the best | |

---

## The single most important line in this file

The corpus is unambiguous here: **110 of 160 winners are handheld, across 43 of 47 accounts.** A locked-off clip reads as an animated photograph, which *is* the AI tell. So this phrasing goes in essentially every Seedance prompt:

```
handheld phone video with small organic sway and micro-corrections,
as if a real person is holding it
```

That is a proven, engine-specific line. Use it verbatim.

Note the distinction from the still-image side: the *camera* should have organic handheld sway, while the *subject action* stays small and simple. Sway ≠ swooping camera moves.

---

## The prompt shape

```
Opening frame: [exactly what the approved still shows].
Motion: [the one thing that moves, plainly described]. One action only, slow and realistic.
Camera: handheld phone video with small organic sway and micro-corrections,
as if a real person is holding it. Background is static.
Duration: [4-5]s.
Audio: [natural ambient sound of the action].
Style: iPhone UGC, natural light, realistic motion.
Hands move slowly and stay in frame.
No morphing, no warping, no text overlays, no rendered signage.
```

---

## Motion rules that keep clips usable

1. **One action per clip.** "She cuts the fabric, then holds it up, then smiles" is three clips, not one.
2. **The action happens after any speaking, never during.** Mouth movement plus hand action simultaneously is where faces break.
3. **Small, real motion beats big, impressive motion.** Every clip in an organic video is one simple beat.
4. **Hands are the weak point.** Slow hand motion, or frame so hands are partly out of shot.
5. **Say what must NOT move.** "Background is static" prevents a lot of weirdness.
6. **Never ask for on-screen text.** Every word is burned in at edit time.

---

## How long the clips should be, from the corpus

These come from real ffmpeg cut detection across 160 winners, so they are measurements, not opinions:

- **Total runtime 8–16 seconds, target 12.** 130 of 160 winners land in that band.
- **4–7 beats, roughly 2 seconds each.** Median is 6 beats at 1.9s per shot. A beat over 3 seconds needs a reason, and a 5-second static beat is the AI tell.
- **First beat is the hook, 3.0–3.5s** (4–5s when there is no build). It is the hook in 160 of 160 videos. Do not trade hook content for hook length — duration is not a lever.
- **Use exactly one build beat, or none.** Two or three ordinary build shots is the worst-performing structure in the corpus.

So a typical Seedance job list for one video is 4–7 separate 2-second-ish generations, not one long clip.

---

## Audio

Organic videos are almost always cut to a trending song, so generated audio usually gets replaced. Still describe the ambient sound of the action — it makes the model animate that action more convincingly, and the natural sound is occasionally worth keeping under the music.

The corpus is clear that music-and-text is the house style: **147 of 160 winners have no spoken voiceover** and 126 of 160 ride trending music. Do not write dialogue or a voiceover script unless the concept specifically calls for one.

---

## Failure modes

| Symptom | Cause and fix |
|---|---|
| Face morphs | Action too big, or the character was speaking. Simplify the motion |
| Hands warp | Slow the hand motion or reframe so hands are partly out of frame |
| Product changes shape | The still was ambiguous. Regenerate the still bigger and clearer, then animate |
| Scene brightens or shifts | Add "preserve the exact lighting from the first frame" |
| Looks like an animated photo | You omitted the handheld sway line, or Enhance was left on |
| Prompt ignored | Enhance was on and rewrote it. Turn it off |
| Still looks AI | Cut the clip shorter. Short clips hide flaws |