Seedance 2.5 is rolling out to every Supervivid account today, and it is the largest quality jump we have shipped since the suite launched. Shots hold together longer, motion reads as physical instead of floaty, and text renders cleanly enough to put signage and titles directly in frame.

What actually changed

The headline numbers are easy to list — native 1080p output, shots up to 12 seconds, roughly 40% faster generation than Seedance 2.0 — but the difference you will feel is consistency. Previous video models tended to drift: a character’s jacket changed color mid-shot, a background building quietly rearranged its windows. Seedance 2.5 was trained with a much longer temporal context, so what is true in frame 1 stays true in frame 300.

Three improvements stand out in day-to-day use:

  • Camera language. The model now understands standard cinematography terms. “Slow dolly-in”, “handheld tracking shot”, and “crane up to reveal” produce exactly the moves a DP would expect, instead of a generic drift.
  • Physical motion. Cloth, hair, water, and smoke follow believable physics. Fast action — a sprinting animal, a car cornering — no longer smears into abstraction.
  • In-frame text. Storefront signs, jersey numbers, and title cards render legibly and stay stable across the shot. This was the single most requested fix from the community, and it is finally here.

Prompting Seedance 2.5

The model rewards specificity. A useful structure is subject, action, setting, camera, and mood — in that order:

A red fox sprinting through fresh snow at dawn, kicking up powder with each stride. Dense pine forest, low golden light. Handheld tracking shot, shallow depth of field. Quiet, wintry, cinematic.

Two practical tips from our internal testing. First, describe one continuous shot per generation; if you want a cut, generate two clips and edit them together. Second, put the camera move in its own sentence — burying “slow push-in” in the middle of a long clause makes the model far more likely to ignore it.

Pricing and availability

Seedance 2.5 costs the same credits per second as 2.0, and it is included in the free daily credits every account receives. Existing projects keep their original model version, so nothing you have already rendered will change — but the model picker will default to 2.5 for all new generations starting this week.

We think this is the first video model where the gap between “AI clip” and “shot from a real production” genuinely starts to close. Open the editor, spend your free credits, and see for yourself.