Seedance 2.0 Prompts: The Complete Guide to AI Video Generation

Seedance 2.0 by ByteDance is one of the strongest video generation models of late 2026. It holds composition, understands cinematographic terminology, and rarely breaks physics. But like any video model, it rewards precise prompts and punishes vague ones. In this guide: a working six-block formula, eight types of camera movement, ready-to-use templates for 4–15 second clips, and the common mistakes that ruin results.

All examples were tested on Nexvy, where both versions are available — Seedance 2.0 (the flagship) and Seedance 2.0 Fast (faster and cheaper), with arbitrary duration from 4 to 15 seconds.

What sets Seedance 2.0 apart from earlier models

Unlike Seedance 1.5, the 2.0 release is noticeably better at holding the principle of "one frame, one idea." Earlier the model would get confused when a prompt contained two competing camera movements or a sharp scene change. Now it reads the prompt like a script: the first sentence sets the subject and atmosphere, the second describes the action, the third defines the camera. That changes the writing approach: instead of one long sentence with dozens of adjectives, write in short semantic blocks.

The second important feature is reference support. You can submit a start frame, an end frame (as hard anchors), or a subject reference (as a soft cue for a face or object). These modes are mutually exclusive: either you anchor the frames, or you provide a sample. Third is linear duration. Previously duration came in fixed steps (5, 10 seconds). Seedance 2.0 accepts any integer value from 4 to 15 seconds, which matters for social platforms: 6 seconds for a Reels preview, 9 for a TikTok hook, 15 for a full scene.

The 6-block formula

To get consistently strong results, keep six slots in mind. You don't have to fill every one — but if the result feels "off," check which slot you skipped.

1. Subject

Describe who or what is at the center. Not "a woman" but "a young Asian woman in an ivory silk dress, long dark hair, no makeup." The more specific your visual details, the less the model invents in styles you didn't want. If the subject is a product, add material and state: "a matte ceramic cup of coffee with rising steam."

2. Action

Use concrete verbs instead of generalities. Bad: "moves beautifully." Good: "slowly turns her head to the left, raises her hand and touches a strand of hair." Indicate pace: "unhurriedly," "smoothly," "abruptly." Avoid the word "fast" in combination with a fast camera — the model starts producing artifacts and motion blur.

3. Environment

Location, time of day, weather, light source. "A narrow street in old Lisbon at sunset, orange light from the setting sun reflecting in the wet tiles after rain." Lighting is the most underrated quality lever. One precise term ("golden hour," "top backlight," "soft window from the left") makes the image more expensive-looking than ten adjectives about "beautiful atmosphere."

4. Camera

One main movement per scene. If you want a pan AND a push-in AND an orbit — the model will make mush. Pick one: dolly-in, pull-out, pan, tracking, orbit, aerial, handheld, or locked-off. More on each below. Add angle and focal length if you like: "low angle, 24mm wide lens."

5. Style

References are clearer than abstractions. "In the symmetrical style of Wes Anderson," "Apple keynote aesthetic," "faceted neon in the spirit of Blade Runner 2049." You can also specify film format ("shot on 35mm"), color palette ("warm ochre tones, deep indigo shadows"), or era ("70s, grain, faded contrast").

6. Constraints

What to exclude. "No on-screen text, no logos, no other people in the frame, no cartoon style." Especially important for commercial videos, where a random logo on a background can be expensive.

Eight types of camera movement

Dolly-in (push-in): the camera moves toward the subject. Works well for dramatic close-ups: "slow dolly from a medium shot to the heroine's face."

Pull-out: the camera retreats, opening up the context. Ideal for closing shots in product ads: first you see a detail, then the whole interior.

Pan: rotation around the camera's own axis, left or right. Use it for wide landscapes and to transition the viewer's gaze from one subject to another.

Tracking shot: the camera moves parallel to the subject. Classic for running, riding, walking heroes. Specify the side: "left to right along the direction of motion."

Orbit: the camera moves along an arc around the subject. An expensive effect that instantly makes the shot cinematic. Works best with a stationary subject.

Aerial: top-down view, like a drone. Specify altitude: "low flyover at 5 meters above the rooftops," otherwise the model will default to a generic "satellite" view.

Handheld: a slight shake, documentary feel. Good for intense, emotional scenes — but don't overdo it, or the viewer will tire within 8 seconds.

Locked-off (static camera): the camera doesn't move. The most underrated option. When the subject does something striking — dances, cooks, speaks — a static camera amplifies attention.

Ready-to-use templates by scenario

Cinematic scene (10 seconds)

A young man in a black trench coat stands on the roof of a high-rise at sunset. He slowly turns his head toward the camera, the wind lifting the hem of his coat. A thin band of orange light on the horizon, the rest of the frame in warm gray-blue tones. Camera orbit along an arc from left to right, low angle. Style of Denis Villeneuve, cinematic color grading, 35mm grain. No on-screen text, no other people.

Product ad (6 seconds)

A matte ceramic espresso cup sits on a dark oak table. A thin trail of steam rises vertically and curves slightly to the left. Warm window light from the right creates soft highlights on the rim of the cup. Slow dolly-in from a medium shot to a close-up. Premium coffee shop aesthetic, dark color palette, contrasting shadows. No text, no logos on the cup.

Nature and atmosphere (12 seconds)

Fog slowly drifts across the surface of a mountain lake at dawn. The black silhouette of a lone tree on the far shore gradually emerges as the fog thins. Cold blue light shifts to warm peach. Locked-off camera, long take. National Geographic style, minimalism, serenity. No people, no buildings.

Character and emotion (8 seconds)

An elderly craftsman in an apron works at a wooden bench in his workshop. He carefully polishes a wooden figurine, focus and a slight smile on his face. Warm light from the window on the left, dust particles in the beams. Slow dolly-in from a medium shot to his face. Documentary aesthetic, warm palette, soft contrast. No dialogue, no captions.

Action and sport (5 seconds)

A surfer glides along the crest of a wave at sunset, the orange light catching the spray. Tracking camera moves parallel to the surfer from left to right. Low angle, almost at the surface of the water. Cinematic style, saturated colors, high water detail. No logos on the board, no other people in the frame.

Fantasy and sci-fi (15 seconds)

A giant whale made of clear glass slowly drifts above an empty future city at night. Inside the whale, living starlight is visible, neon beams shining through the transparent skin. The camera slowly pulls out from below upward, revealing the scale. Blade Runner 2049 aesthetic, saturated neon, deep shadows, volumetric fog. No people, no text.

ASMR and calming content (9 seconds)

Raindrops slowly fall onto the wide green leaves of a tropical plant, close-up. Each drop trails down a vein of the leaf, catches the light, and breaks off downward. Locked-off camera, macro shot, shallow depth of field. Natural soft diffused light. Minimalism, calm, water texture. No extraneous elements in the composition.

Common mistakes

Vague adjectives. "Epic," "cool," "incredibly beautiful" — these words mean nothing to the model. Replace with specifics: instead of "epic sunset" — "orange sky with purple clouds, low sun on the horizon."

Two camera movements at once. "Push-in and at the same time an orbit" — the model can't handle it. Pick one main movement. If you really need a combo — generate two scenes and stitch them in editing.

Prompt overload. Optimal length is 60–120 words. Past 150 words the model starts losing details from the beginning. If you want to add more, split into scenes.

Conflict between subject and camera motion. If the subject runs and the camera moves quickly in the other direction, you get a disorienting artifact. Either synchronize (tracking) or leave the camera static.

The word "fast." Seedance 2.0 interprets "fast" aggressively — it cranks blur, loses detail. If you want dynamics, use concrete actions ("jump," "strike," "flash"), not a generic speed adjective.

Frequently asked questions

What's the difference between Seedance 2.0 and 2.0 Fast?

Fast is the same model with a simplified pipeline: roughly twice as fast and cheaper, but with slightly less detail in complex scenes. For drafts, concepts, and quick prompt testing — go with Fast. For the final version — the flagship.

What duration should I choose?

4–6 seconds — for Reels and TikTok hooks. 8–10 seconds — for ads and storytelling. 12–15 seconds — for full cinematic scenes. Price on Nexvy scales linearly with duration, so you pay for exactly the length you need.

Can I use reference images?

Yes. Seedance 2.0 supports a start frame, an end frame (as hard anchors), and a subject reference (as a soft cue). These modes are mutually exclusive — pick one. Details are available in the tooltips inside the Nexvy interface.

Why doesn't the result match my prompt?

Most often — because the prompt describes "mood" rather than "frame." Rewrite it as a storyboard: what's in the center, what it does, where it is, how it's shot, in what style. If after that the result still isn't right — try Seedance 2.0 (flagship) instead of Fast.

Conclusion

Seedance 2.0 is a tool that loves structure. Six blocks, one camera movement, concrete verbs, and a precise description of light — that's enough to consistently produce cinematic results. Don't try to fit an entire script into one scene: 4–15 seconds is one shot, not a short film. Take one of the templates above, swap in your subject and environment, and generate the first version on Nexvy. After ten to fifteen iterations you'll develop your own prompting style — and that will save you hundreds of hours and hundreds of dollars in production.