Seedance 2.5 Prompt Guide: Better AI Video Prompts

Seedance 2.5 prompts work best when they read like a compact directing brief. A weak prompt describes a beautiful frame. A strong prompt describes what changes, what the camera notices, what must remain consistent, and what the audience hears.

That difference matters because Seedance 2.5 can generate a continuous video up to 30 seconds, use multiple references, and create synchronized sound. More duration and more inputs give you more control, but they also create more ways for a prompt to contradict itself.

Quick answer: build your Seedance 2.5 prompt in this order: subject and event, setting, visual treatment, camera, sound, timeline, continuity rules. Label every reference by its upload order. For a long clip, give each stage one main event and a visible end state that the next stage can inherit.

Seedance 2.5 at a Glance

Seedance 2.5 is ByteDance's joint video-and-audio generation model. According to the current BytePlus Seedance 2.5 tutorial, its main production features are longer clips, multimodal references, native audio, first-and-last-frame generation, video editing, and video extension.

This distinction is important. The official model can do more than every provider integration exposes. In Melies today, you can generate from text, animate one image, guide a shot with several image references, or control its first and last frame. Do not upload a video and write an editing prompt in Melies until that workflow appears in the tool.

Watch the official Seedance 2.5 examples

This official BytePlus reel shows long-form scenes, reference control, performance, camera direction, and synchronized sound produced with Seedance 2.5. Treat it as a capability demonstration, not a promise that every prompt or seed will produce the same result.

The examples are published by BytePlus, ByteDance's official enterprise platform. Pause on a shot and identify the visible subject action, camera rule, sound cue, and end state before reading the prompt patterns below.

The Seedance 2.5 Prompt Formula

Start with six questions:

1. Who or what is the subject?
2. What changes during the clip?
3. Where does it happen?
4. How should it look and feel?
5. How does the camera observe it?
6. What should we hear?

A useful short formula is:

[Subject] performs [specific action] in [setting].
The image uses [light, texture, palette, and medium].
The camera uses [shot size, angle, movement, and cut behavior].
Audio contains [dialogue, ambience, effects, or music].

For example, compare these two prompts.

Weak:

A woman waits at a train station, cinematic.

Stronger:

A tired commuter in a wet navy coat waits alone on an outdoor train platform before dawn. She checks the empty tracks, hears a distant metallic vibration, and slowly looks toward the tunnel. Begin in a locked wide shot, then make one restrained push toward her as the platform lights flicker in sequence. Cold rain catches the fluorescent light. Keep the station almost silent except for rain, electrical hum, and the approaching vibration. No music and no subtitles.

The stronger prompt does not win because it contains more adjectives. It wins because it defines an event, a camera decision, a change in attention, and a sound progression.

Write Motion, Not a Photograph

Image prompts describe what exists. Video prompts describe what happens.

Replace static attributes with observable change:

Concrete physics are especially useful. Name momentum, weight, drag, contact, balance, or secondary motion when they matter. Do not add every physical detail to every prompt. Pick the two or three interactions that sell the shot.

Label Every Reference and Give It One Job

References are Seedance 2.5's most useful control surface. They are also a common source of identity swaps, duplicated objects, and unwanted backgrounds.

Use the upload order in your prompt:

@Image 1 defines Mara's face, short black hair, and yellow raincoat only. Do not use its background.
@Image 2 defines the underground platform, tiled columns, and green fluorescent light. Do not use the people in the image.
@Image 3 defines the red hard-shell suitcase and its worn silver handle only.

The official BytePlus prompt guide recommends explicit mapping because labels placed inside an image are not a reliable substitute for prompt instructions.

Use narrow roles

@Image 1 defines the woman is too broad. The model can inherit her wall, light, pose, or another person standing beside her.

Prefer:

@Image 1 defines the woman's facial identity, hairstyle, and green jacket only. Ignore the background, pose, and lighting.

Say when several images show one object

Multiple views can become multiple objects unless you explain the relationship:

@Image 1 shows the front of one folding desk lamp.
@Image 2 shows the left side of the same lamp.
@Image 3 shows the back of the same lamp.
All three images define one lamp. Only one lamp appears in the video.

Build a reference hierarchy

When references disagree, the model needs priorities. A practical order is:

1. Character identity
2. Essential prop or product geometry
3. Wardrobe
4. Location and spatial layout
5. Lighting and texture
6. Motion and camera style

State the priorities at the end of a demanding prompt. Sacrifice background detail before identity or product geometry if those are the reason for the shot.

Direct 30-Second Video as a Sequence of States

A 30-second prompt is not a longer five-second prompt. It is a small continuity system.

For every stage, define:

• the initial state;
• one main event;
• the camera behavior;
• the visible end state.

The end state is the handoff. It tells the next stage where each person and object begins.

[0-7s - Arrival]
Initial state: Mara enters from frame left carrying the red suitcase in her right hand.
Event: She crosses the empty platform and stops beside the second tiled column.
Camera: A medium tracking shot moves parallel to her, then stops with her.
End state: Mara faces the tunnel; the suitcase stands upright beside her right leg.

[7-18s - Signal]
Continue with the same identity, wardrobe, suitcase position, platform layout, and light.
Event: The fluorescent lights switch off one by one from the far end toward Mara. She does not move until the final light above her goes dark.
Camera: Hold the same axis and push in slowly.
End state: Mara remains beside the second column in silhouette, one hand on the suitcase handle.

This method came up repeatedly when reviewing long-form examples: continuity improves when a prompt treats cooking, dressing, travel, damage, or prop placement as irreversible state changes. Raw food cannot return after it has been cooked. A wet coat does not become dry between cuts. A suitcase placed on the right should not jump to the left without an action that moves it.

Use timestamps as budgets, not frame-accurate edit points

Seedance 2.5 supports whole-second time ranges. Keep them consecutive and do not leave gaps:

0-5s: ...
5-12s: ...
12-20s: ...
20-30s: ...

Too little action gives the model room to improvise. Too much action produces rushed cuts or omitted beats. One main change per range is a good starting point.

Do not use timestamps as a mechanical counter. A direction such as "blink three times in one second" is less reliable than a visible performance beat such as "she holds eye contact, blinks once, then looks away."

Direct Camera Movement With a Start and an End

Seedance 2.5 understands standard shot and camera terms, including wide shot, medium close-up, low angle, overhead, push in, pull out, pan, track, orbit, tilt, handheld, one-shot, dolly zoom, and rack focus.

Naming a term is only the start. A useful camera instruction also states:

• what the camera follows;
• where it begins;
• where it ends;
• what motivates the movement;
• what must remain readable.

Vague:

Dynamic orbit around the dancer.

Directed:

Begin at ankle height behind the dancer's right heel. As she turns, rise into one clockwise orbit and finish at eye level in a medium close-up. Her face remains readable through brief foreground occlusions. The camera slows when her body becomes still. No collision, weightless floating, or speed ramp.

If a technical term can be interpreted several ways, explain the visible result:

Rack focus: the foreground leaves gradually soften while the face behind them becomes sharp.

One motivated move is usually stronger than a pile of moves. If the prompt says orbit, crane, zoom, whip pan, and handheld in the same five seconds, the model has no stable camera rule to follow.

Write Performance as Observable Behavior

"She becomes sad" is an emotional label. It does not direct a performance.

Use a four-part beat:

1. Trigger
2. Immediate reaction
3. Gradual change
4. Held expression

When the station announcement says the train has been cancelled, Mara's eyes remain on the tunnel for one beat. Her grip loosens on the suitcase. She exhales through her nose, lowers her gaze, and lets one restrained smile appear as if the bad news confirms something she already knew. She holds that expression without looking into the lens.

Two to four physical signals are enough for one change. More detail can make the face unstable or turn a quiet performance into exaggerated mime.

For a talking head, connect each performance beat to a specific part of the dialogue. Do not write four emotional stages and leave the model to guess which sentence belongs to each one.

Direct Dialogue, Sound Effects, Music, and Subtitles

Seedance 2.5 can generate sound with the image. Keep the sound plan as disciplined as the camera plan.

The official guide uses four markers:

Plain language also works. The markers are most useful when a sentence could be mistaken for a subtitle, lyric, or spoken effect.

For dialogue:

• name the speaker;
• name the language before non-Chinese dialogue;
• keep the line short enough for the available time;
• describe delivery before the line;
• state No subtitles if you do not want the line rendered on screen.

Dialogue language: natural American English. Mara speaks quietly, without turning toward the camera: {I knew I had missed it.}
No subtitles. No background music. Keep only her voice, rain, and the electrical hum of the platform.

Record yourself reading the dialogue at the intended pace before generation. If the spoken line already fills the entire time range, the model has no room for silence, reaction, or physical action.

Use Positive Priorities Before a Long Negative List

Positive instructions tell the model what to build. Negative instructions are best kept for a small number of costly failures.

Start with:

Keep the camera locked at waist height. Preserve one subject and one red suitcase throughout. Use only rain, footsteps, and station ambience.

Then add targeted exclusions:

No subtitles, no background music, no duplicated suitcase, no new passengers, and no cut away from the platform.

A giant generic negative prompt often contains conflicts. For example, excluding blur while asking for shallow depth of field, or excluding camera shake while asking for handheld footage. Protect the aesthetic as well as the anatomy.

A Production-Ready Seedance 2.5 Prompt Template

Use only the sections your shot needs. A simple clip should remain simple.

[Generation Goal]
Generate a [duration]-second [format or genre] video in which [main dramatic event].

[Reference Roles]
@Image 1 defines [identity or object feature] only. Ignore [unwanted content].
@Image 2 defines [location, wardrobe, prop, or light] only. Ignore [unwanted content].

[World and Visual Treatment]
Describe the location, time, light, palette, texture, atmosphere, and relevant physical behavior.

[Camera Rule]
State the opening shot, subject, movement, axis, end framing, and cut behavior.

[Timeline]
[0-Xs] Initial state. One main event. Camera behavior. End state.
[X-Ys] Continue all invariants. One main event. Camera behavior. End state.

[Performance]
State the trigger and two to four visible changes in eyes, breath, mouth, hands, posture, or gaze.

[Audio]
State dialogue language and lines, ambience, effects, music, and what must remain absent.

[Maintain Consistency]
List identity, wardrobe, prop ownership, screen direction, spatial layout, light, and sound invariants.

[Priority]
First priority: ...
Second priority: ...
Sacrifice ... before either priority.

Five Original Seedance 2.5 Prompt Examples

These prompts are designed as starting points, not magic spells. Replace the references and simplify any detail that does not serve your shot.

1. Cinematic one-take with three references

[Goal]
Generate a 10-second realistic one-take in an empty municipal swimming pool after closing. A young maintenance worker discovers that one lane still contains moving water although the pumps are off.

[References]
@Image 1 defines the worker's face, cropped hair, orange coveralls, and black rubber boots only. Ignore the background and pose.
@Image 2 defines the drained tiled pool, lane ropes, overhead windows, and sodium work lights. Do not use the people in the image.
@Image 3 defines the old brass maintenance key in his left hand only.

[Shot]
Begin in a low wide shot inside the empty pool. Track backward as the worker walks along the pool floor. His boots scrape dry tile and his flashlight passes across the lane markings. At 4 seconds, water begins moving through the empty lane beside him without appearing from nowhere: a shallow ripple travels from the dark end toward his boots. He stops. The camera stops with him, then makes one slow push to a medium close-up as the ripple reaches frame foreground. He looks down, holds his breath, and closes his hand around the key. End with the worker still and sharply readable while the moving reflection crosses his face.

[Sound and continuity]
No dialogue and no music. <Boots scraping tile.> <Distant ventilation hum.> <A thin wave moving over ceramic.> Keep one worker, one key, the same drained pool geography, and one continuous camera axis. No cuts, no flooding pool, no extra people, and no weightless camera movement.

Why it works: the impossible event has a physical direction, the camera reacts to the character, and the final image resolves the shot.

2. Multi-reference fashion film with product details

@Image 1 defines the model's face, shaved head, and proportions only.
@Image 2 defines one silver jacket, including its asymmetric collar and red stitched cuff.
@Image 3 defines one pair of black high-top shoes with translucent soles.
@Image 4 defines the concrete underpass and green fluorescent light. Ignore all people and clothing in the location image.

Generate a 15-second raw late-1990s fashion film. 0-5s: handheld wide shot as the model walks toward camera through the underpass; the heavy jacket lags behind each step and the shoes remain fully visible. 5-10s: hard cut to a side medium close-up; a passing train produces one gust that lifts the asymmetric collar and reveals the red cuff when the model raises a hand to shield their eyes. 10-15s: low rear tracking shot as the model walks into the darker end of the tunnel; the jacket settles and the translucent soles catch the last green light.

Use harsh fluorescent exposure, mild camcorder noise, imperfect autofocus, and real fabric weight. Keep the same face, one jacket, one pair of shoes, and the same walking direction. Natural train rumble, footsteps, and fabric movement only. No polished studio lighting, no gimbal smoothness, no logos, no subtitles, and no music.

Why it works: every reference has one job, product details become visible at planned moments, and the negative controls protect the rough capture style.

3. Talking-head performance with dialogue beats

@Image 1 defines Nia's face, shoulder-length curls, blue shirt, and softly lit apartment background. Keep the same framing and room light.

Generate a 20-second handheld phone video, chest-up at arm's length. Nia speaks directly to one close friend, not to an audience. Subtle natural drift, occasional refocus, quiet room tone, no hand gestures entering frame.

0-6s: She holds warm eye contact and smiles with slight embarrassment. Dialogue language: natural British English. Nia says: {I rehearsed the whole speech in the lift.}
6-13s: The smile fades. She looks briefly toward the closed door, lowers her voice, then returns her eyes to the lens: {Then the door opened, and I said absolutely nothing.}
13-20s: Her shoulders release. She gives one small head shake, laughs through her nose, and ends on a calm, genuine smile: {So I suppose this is the speech now.}

Accurate lip sync, natural blinking, and small breaths between lines. Her face, hair, shirt, room, and light remain unchanged. Her voice only. No captions, music, beauty filter, exaggerated acting, cutaways, or extra hands.

Why it works: each line has its own visible performance and enough time for pauses.

4. Thirty-second scene with irreversible continuity

[Goal]
Generate a 30-second documentary-style scene of a ceramic artist finishing, packing, and handing over one pale-blue cup at dawn.

[References]
@Image 1 defines the artist's identity and green apron only.
@Image 2 defines the studio layout, wheel, shelf, window, and dawn light. Ignore the person in the image.
@Image 3 defines one pale-blue cup with a hairline gold repair. Only one finished cup appears.

[0-8s - Finish]
The artist lifts the wet cup from the stopped wheel with both hands and places it at the center of the workbench. Static medium shot. End state: cup centered, wheel stopped, towel to frame right.

[8-17s - Inspect]
Continue the same layout. Slow push in as the artist dries their hands, rotates the cup once, and finds the gold repair in the window light. End state: cup upright, handle facing left, artist's right index finger beside the repair.

[17-25s - Pack]
Hard cut to a high three-quarter shot. The artist wraps the same cup in brown paper and places it inside one open box. End state: wrapped cup inside box, lid still open, blue string on the left.

[25-30s - Handoff]
Eye-level two-shot. A customer enters only with their hands and forearms visible. The artist closes the box, ties the blue string once, and passes it across the workbench. End on the artist's empty hands.

Natural wheel hum fading out, paper friction, string tightening, birds outside. No dialogue and no music. The cup never returns to an earlier state. Preserve prop count, workbench orientation, window direction, apron, and dawn light.

Why it works: each stage changes the state of the cup once and hands a clear arrangement to the next stage.

5. First-and-last-frame transition

Use two images with the same aspect ratio.

@Image 1 is the first frame. It defines the empty winter orchard, locked wide composition, bare branches, gray sky, and one red chair at center.
@Image 2 is the last frame. It defines the same composition at spring sunrise, with blossoms on the same trees and the same red chair.

Generate one continuous 12-second locked-off transformation from @Image 1 to @Image 2. The season changes through visible physical progression rather than a dissolve: frost recedes from the foreground, small buds form along the nearest branches, pale blossoms open from near to far, and warm sunlight reaches the red chair last. The chair, camera, horizon, tree positions, and branch geometry remain fixed. A faint winter wind gradually gives way to insects and morning birds. No people, no camera movement, no crossfade, no text, and no sudden object replacement.

Why it works: the prompt names both frame roles, protects shared geometry, and describes the transition as one continuous event.

Editing and Extension Prompt Patterns

The following workflows are part of the official BytePlus model, but they are not currently exposed in Melies.

Edit only one part of a clip

[Edit Goal]
Edit @Video 1 only from 6 to 10 seconds. Change the cool blue light on the right wall to warm amber.

[Source Role]
@Video 1 is the sole editing master. It defines the person, room, action, composition, camera movement, dialogue, ambience, and event order.

[Preserve]
Keep identity, clothing, expression, body position, room geometry, camera, dialogue, and timing unchanged. Allow the nearby skin and wall surfaces to respond naturally to the warmer light. No other edit.

Extend a clip forward

Extend @Video 1 forward. The first frame of the extension directly continues its last frame. Preserve the locked medium shot, the paper airplane's position and direction, the classroom window, afternoon light, curtain motion, and room tone. The airplane continues toward frame right, exits, and the curtain settles after its wake. Do not reintroduce the airplane after it leaves.

For editing, the source video's ratio and approximate duration are locked. For extension, the source ratio is locked. BytePlus recommends MOV input and output for better audio-visual continuity in these workflows.

Common Seedance 2.5 Prompt Failures

What Seedance 2.5 Cannot Guarantee

Good prompting reduces ambiguity. It does not turn generation into deterministic rendering.

• Timestamps allocate time, but they are not frame-accurate cut points.
• Storyboard grids guide sequence and composition, but they do not guarantee exact panel reproduction.
• Edits can preserve the source closely, but not necessarily pixel for pixel.
• Exact signs, subtitles, equations, packaging copy, and legal text belong in post-production.
• More references are not automatically better. BytePlus recommends 1 to 8 subjects for more stable image-reference results.
• First-and-last-frame images should share an aspect ratio. Otherwise, the last frame can be stretched to match the first.
• The BytePlus API does not support direct reference-image or reference-video uploads that contain real human faces. Review the provider's current portrait workflow and obtain the necessary rights and consent before using any person's likeness.

These limits are creative constraints, not footnotes. If exact text, identity, product geometry, or timing is essential, split the task into smaller controlled shots and finish the precise layer in editing.

How to Use Seedance 2.5 in Melies

Seedance 2.5 is a premium model available on Melies Pro and Max. It costs 50 credits per generated second at 720p and about half that at 480p, before small aspect-ratio adjustments. See the Seedance 2.5 pricing guideSeedance 2.5 Pricing: Credits, Cost per Second and PlansSee Seedance 2.5 pricing per second in Melies, fal.ai API costs, BytePlus billing, plan capacity, and practical ways to reduce AI video costs in 2026. for plan capacity, direct API prices, and budget examples.

1. Open the Melies AI Video Generator.
2. Select Seedance 2.5.
3. Choose text-to-video, one starting image, multiple image references, or first-and-last-frame control.
4. Set a duration from 4 to 30 seconds.
5. Choose 480p for lower-cost iteration or 720p for higher-quality output.
6. Choose your aspect ratio and enable synchronized audio when the shot needs it.
7. Paste a prompt that fits Melies' current 2,500-character limit.
8. Check the displayed credit quote before generating.

Seedance 2.5 pricing in Melies is duration-based. The current base quote is 50 credits per second at 720p, before small aspect-ratio adjustments. A 4-second clip starts at 200 credits, a 10-second clip at 500, and a 30-second clip at 1,500. The 480p option is about half the 720p base. The quote shown in the tool is the source of truth before generation.

For an expensive 30-second attempt, test the hardest uncertainty first. If identity is the risk, generate a short identity-and-motion test. If the orbit is the risk, test the camera path. If dialogue is the risk, read and time it before spending credits on the full scene.

Seedance 2.5 Prompt Checklist

Before you generate, verify that:

• the subject and main event are clear in the first lines;
• every reference has one named role and explicit exclusions;
• multiple views are identified as one subject or object;
• the camera has a subject, path, and final frame;
• every long-video stage has one main change and an end state;
• identity, wardrobe, prop ownership, screen direction, and light do not contradict each other;
• dialogue fits its time range and has a speaker and language;
• sound instructions distinguish voice, ambience, effects, and music;
• negative instructions do not contradict the desired style;
• exact text and graphics are reserved for post-production;
• the generation settings are selected in the tool, not buried in the prompt;
• the cost quote is acceptable before you generate.

The best Seedance 2.5 prompt is not the longest. It is the one that gives every important element a role, every event enough time, and every stage a clean handoff to the next.

For adjacent workflows, see the AI camera movement guideAI Camera Movements: Dolly, Jib & Cinematic PresetsMaster camera movement in AI video generation. Learn about Dolly, Jib, and Static presets on LTX 2 Pro, plus cinematic camera techniques for all models., consistent character guideHow to Create Consistent Characters with AILearn how to generate consistent AI characters that look the same across multiple images. Step-by-step tutorial using Nano Banana 2 and reference images on Melies., AI video extension guideHow to Extend AI-Generated VideosLearn how to make an AI video longer with final-frame continuation, current model options, honest credit costs, and tips for smoother results., and complete AI filmmaking guideAI Filmmaking: How to Create an AI Film from Idea to ExportLearn the complete AI filmmaking workflow with Melies. Generate ideas, cast AI actors, create storyboards, produce video clips, edit on a timeline, and export your finished AI film..