How to Prompt Realistic Motion and Physics in Seedance 2.0 Mini

Believable generated motion starts with a visible chain of cause, movement, contact, and consequence-not an action label such as "runs realistically" or "cinematic collision." For each short shot, describe the starting pose, direction and pace, support or contact point, weight shift or reaction, and camera behavior separately. Then test small prompt variations and keep the take that best communicates the physical event.

Seedance 2.0 is described as improving physical accuracy, motion stability, realism, and controllability compared with version 1.5. It can still produce generation flaws, however, so treat it as a creative video tool rather than a deterministic physics simulator. The goal is not to write a perfect simulation specification. It is to remove the ambiguity that makes movement look floaty, jittery, or disconnected from the scene.

Build the Prompt Around One Physical Event

Start by deciding what the viewer must believe happened in the shot.

Not "a runner in a city." Instead: "a runner pushes off from a planted foot, accelerates across frame, then slows before the curb."

That distinction matters because the second version supplies both spatial information-where the subject begins and travels-and temporal information-how the movement changes over time.

A useful prompt pattern is:

[Subject and starting stance] + [one action and cause] + [path, pace, and timing] + [contact or support] + [reaction or follow-through] + [camera behavior] + [environment/style]

This is a writing pattern, not required Seedance syntax. Keep only the details that help the shot read clearly. Guidance for Seedance 2.0 Mini favors concise visual instructions over long prompt scripts; a basic structure can include the subject, specific action, camera movement, mood, and, where useful, an ending detail. See the Mini prompt-structure guidance.

Replace labels with observable movement

Vague

A woman jumps over a puddle realistically, cinematic slow motion.

More directed

A woman stands with both feet planted at the puddle's edge, bends her knees, swings her arms forward, and jumps across in one motion. Her shoes leave the ground together; she lands on both feet with a brief knee bend and a small splash behind her. Camera: locked-off medium-wide side view.

The second prompt does not guarantee a perfect jump. It gives the generation a clearer sequence: support, compression, push-off, flight, landing, and response.

For most action shots, prioritize this order:

1
1. Start state: Where are the feet, hands, wheels, or object before motion begins?
2
2. Primary action: What changes?
3
3. Path and pace: In which direction, at what speed, and does it accelerate or slow down?
4
4. Contact: What supports, catches, pushes, strikes, grips, or resists the subject?
5
5. Consequence: What visibly reacts after the contact?
6
6. Camera: Is the viewpoint still, tracking, panning, or moving independently?

Style, lighting, wardrobe, and setting still matter, but add them after the physical event is unambiguous.

Prompt the Contacts That Make Motion Feel Heavy

"Realistic physics" is too broad to be a useful instruction. Instead, identify the one or two contacts that make an action believable: a shoe meeting pavement, fingers holding a bottle, tires loading into a turn, or an object striking water.

These are prompt patterns based on visible cues, not claims that the model calculates anatomy, forces, collisions, or fluid dynamics precisely.

A helpful test is to ask: What is carrying the weight? If the answer is unclear, the shot may read as floating.

For people, name the support point before naming the gesture. For objects, name the grip or surface contact before naming the movement. For impacts, describe the approach and the after-effect-not just the collision itself.

Avoid stacking unrelated actions into one brief generation:

A dancer runs into frame, grabs a bag, jumps over a puddle, spins, throws the bag, and the camera circles around her.

Even if every action is individually clear, their combined timing and interactions create competing instructions. Turn that into separate shots: the run and grab, the jump, then the throw.

Keep Camera Movement Separate From Subject Movement

A moving camera can make a stationary subject appear to move, while a moving subject can be mistaken for a camera move. State both independently.

Name the camera behavior rather than relying on a generic request for "dynamic camera movement." Separating camera direction from subject action reduces the chance that the shot's movement becomes unclear.

In one documented Mini-hosted workflow, Seedance 2.0 Mini supports text-to-video and image-to-video generation. Check the current Dreamina Seedance 2.0 workflow before planning a reference-dependent production, because available inputs and controls can vary by host.

Where the current interface supports references, use them with a narrow purpose:

1
1. Use an image to communicate composition, subject appearance, or visual style.
2
2. Use a video reference to communicate a desired action rhythm or camera path.
3
3. Repeat the crucial motion instruction in text anyway.
4
4. Do not assume a reference will be copied exactly.

A reference can reduce uncertainty, but it does not replace a clear statement of the shot's physical event.

Diagnose the Failure, Then Revise One Detail

When a result feels wrong, resist the urge to rewrite the entire prompt. Locate the missing cue and change one variable at a time. That makes comparisons easier and helps reveal whether the issue is the action, contact, timing, or camera instruction.

Long, ambiguous, or internally conflicting instructions can degrade generated edits. This prompt guide discusses avoiding ambiguity and conflicting directions. If your host exposes seed controls, reusing a seed can support cleaner prompt comparisons, but do not treat it as a guarantee of identical output.

Use a simple loop:

1
1. Generate a short test focused on one physical event.
2
2. Identify the first broken cue-for example, the planted foot, grip, impact, or camera path.
3
3. Revise only that cue and generate another variation.

If the sequence needs multiple consequential actions, split it. A shot of a person lifting a package can be one clip; the turn and walk away can be another. Assemble the strongest takes later rather than asking a single generation to preserve every contact and transition.

Verify the Host Controls Before You Design the Shot

Seedance 2.0 Mini access is host-dependent. Do not design a workflow around controls you have not confirmed in the interface you are using.

Before production, check:

1
1. Whether your current workflow supports text-to-video, image-to-video, or other reference inputs.
2
2. Which duration options are available. One documented Mini-hosted workflow lists 4-second, 8-second, and 15-second options, but those are not universal Mini specifications.
3
3. Available aspect ratio, resolution, export, and reference-file settings.
4
4. Whether seed selection is exposed.
5
5. Current platform rules for permissions, consent, and use of client, branded, or third-party assets.

The available evidence does not establish a universal Mini-specific negative-prompt syntax for preventing floating, skating, warped anatomy, or impossible impacts. If your host provides negative-prompt controls, follow that interface's current documentation rather than assuming terms or syntax transfer across Seedance tools.

Mini is positioned as a faster, lower-cost option, while quality-focused models may be a better fit when a final deliverable requires maximum detail, realism, or polish. For action work, that makes shot design especially important: short, simple, well-defined events are easier to assess and refine than a long chain of interactions.

Choose one action for your next test. Write a single-beat prompt with a clear start, force, contact, and finish; generate a few small variations; and keep only the take whose motion reads credibly. Then bring the selected clips into CapCut to trim, sequence, reframe, add sound cues, and shape them into a paced final edit. Better prompts reduce ambiguity, but careful shot design, iteration, selection, and editing are what make the finished sequence convincing.