Read your failed takes | Academy

When every take in a batch of four comes back wrong, the seed isn't the problem — the prompt is. A single bad clip can be a fluke; four bad clips out of four means Seedance never had enough to work with, and no amount of rerolling fixes that.

One of four failed takes — man spread, no dynamics, flat angle. Verdict: rewrite the prompt, don't reroll.

The tells are specific, not vibes. A third chair drifting into a two-person scene means the prompt never locked the furniture count, so Seedance improvised.

A third chair appears from nowhere — the prompt didn't lock what's in the room, so the model filled the gap itself.

Boring angles and dead-eyed acting are a different failure: the prompt described the setting but not the performance. Fixing that means writing the movement and the emotion as specifically as the room.

Never write Seedance a negative

Seedance doesn't parse "he's not crying" as an instruction — it reads "crying" and gets confused, sometimes doing the opposite of what you meant. State the emotion you want instead: "an anxious look," not "not calm."

Drop every negative prompt with Seedance 2.0. Say what the shot should look like, never what it shouldn't.

Rewriting with positive, specific direction changes the failure mode, not always the outcome. The next batch still drifts — the location shifts and extra props spawn in — proof that one fix rarely clears every problem at once.

Same rewritten prompt, second batch — the location drifted and stray objects spawned in. Emotion fixed, staging still loose.

The last take in this batch is the one worth keeping around: the emotion finally reads right, but the composition is still plain coverage — a static wide two-shot with no dynamics, no reason to look at it twice.

Positive phrasing fixed the emotion, but the shot is still flat, basic coverage. Good enough to build on, not to lock.

Basic isn't broken. It's the signal to stop tuning emotion and start tuning camera — the movement alone was never going to carry the scene.