Seedance 2.0 and 2.5: what a clip costs and how to use it

# Seedance prompt card · AIGE

How to write prompts for ByteDance Seedance 2.0 and 2.5 in AIGE, drawn from ByteDance's own prompt guides, with AIGE's settings and credit prices.

Updated 2026-09-27 · By AIGE

---

## Which Seedance to pick

The model menu lists Seedance 2.0 as **Seedance**, with a Pro/Fast quality switch, and Seedance 2.5 as **Seedance 2.5**.

| | Seedance 2.0 Fast | Seedance 2.0 Pro | Seedance 2.5 |
|---|---|---|---|
| Length | 4–15 s | 4–15 s | 4–30 s in one shot |
| Resolution | 480p, 720p | 480p, 720p, 1080p | 480p, 720p |
| References | 9 pictures, 3 clips, 3 sounds (12 in total) | same as Fast | 30 pictures, 10 clips, 10 sounds (50 in total) |
| 5 s at 720p | 121 credits | 152 credits | 232 credits |
| 5 s at 480p | 54 credits | 68 credits | 103 credits |
| How it reads timing | Shot 1, Shot 2… | Shot 1, Shot 2… | Shot 1… or second-by-second timestamps |

- **2.0 Fast** is ByteDance's pick for "a balance of cost and generation speed when top-tier quality is not required" [2]. Use it to test an idea.
- **2.0 Pro** is ByteDance's pick "for the highest generation quality" in the 2.0 series [2]. It is the only Seedance in AIGE that renders 1080p.
- **2.5** makes a single video up to 30 s and takes up to 50 references [4]. It follows integer-second timestamps; 2.0 responds only to shot numbers [4].
- ByteDance says 2.5 and 2.0 have "significantly different aesthetic style preferences" [4], so the same prompt can look different on each. Keep all the shots of one sequence on one version.
- Sound is generated with the picture and costs the same with or without it [6].
- Seedance 2.5 comes with the Pro plan and up. Starter includes fewer Seedance 2.0 settings. If a setting isn't on your plan, AIGE refuses the run before charging any credits.

## The prompt formula

ByteDance's advanced formula for Seedance 2.0 [1]:

> precise subject + action details + scene/environment + lighting & color tone + camera movement + visual style + image quality + constraints

Its order for Seedance 2.5 is subject + action/event + scene and environment + visual style + camera movement/shot cuts + sound. You can leave out any part you don't need [5]. ByteDance sums it up this way: first lock in who is doing what, then where and what atmosphere, then how to shoot, then tighten the result with style, image quality and constraints [1].

1. **Name the subject every time it appears.** Either bind it to its reference each time ("the woman in red from @Image1"), or define a label once ("the tall man in @Video1 is the officer") and use that exact label from then on [1].
2. **Split a complex clip into shots.** Write Shot 1, Shot 2, Shot 3 in the order things happen. Give each shot its camera move or cut, what the subject does, where they are in the space, and the sound [1]. On 2.0, don't give shots durations. ByteDance says support for "precise timing (such as 0–3 seconds) is unstable" [1].
3. **Describe movement physically.** Name the body part and the range, speed and force, for example "slowly raise a hand" or "push hard off the ground". Say how one action flows into the next, and show emotion through physical detail rather than words like "very sad" [1]. For 2.0, ByteDance recommends slow, gentle, continuous movement over sprints, big jumps and violent rolls [1]. For 2.5, describe actions in general terms and detail only a few memorable ones [4].
4. **Use one camera move per shot.** ByteDance warns that asking for push, pull, pan and move at once "will increase image instability" [1].
5. **Finish with quality, style and constraints** [1].
   - Quality: "HD, rich details, cinematic texture, natural colors, soft lighting".
   - Style: "cyberpunk cool blue-purple tone", "retro film".
   - Constraints: "keep it subtitle-free", "do not generate a logo", "do not generate a watermark".
6. **Stay concise.** Avoid contradictions, and don't paste a full script as the prompt [1]. On 2.5, write what you want rather than what you don't; negatives are supported for subtitles and audio [4].

```
[Who, 2–3 fixed traits] [does what, physically] in [where, light, mood].
Shot 1: [one camera move]. [action]. [sound]
Shot 2: Cut to [shot size]. [action]. [sound]
Shot 3: [camera move]. [action]. [sound]
[Style], [image quality]. [Constraints: no subtitles, no logo, no watermark.]
```

## Long single shots (Seedance 2.5, up to 30 s)

In AIGE, Seedance 2.5 runs any whole length from 4 to 30 seconds. ByteDance's structure for a 2.5 prompt [4]:

1. **References:** what each one is for. Skip this if you are working from text only.
2. **Summary:** one sentence.
3. **Plot:** a timeline using timestamps or "Shot N". For each segment, give the visuals, camera, action, dialogue and sound.
4. **Constants:** what stays the same throughout, such as camera angle, movement, setting, sound and atmosphere.

Timestamps on 2.5 [4]:
- Use whole seconds: "0-3 seconds… 3-7 seconds… 7-15 seconds", "[1s-4s]…", "At the 2-second mark…", or "After 3 seconds…".
- Keep the timeline continuous. A gap such as "0-3s… 5-6s…" breaks it.
- Put a sensible amount in each segment. Too little and the model improvises; too much and you get extra cuts or lose parts of the plot.
- Don't use timestamps for fast repeated actions such as "shake your head three times per second".
- For a transition, give the moment and the method: "At the 5-second mark, the camera quickly transitions leftward using a left wipe combined with a natural dissolve."
- For one unbroken take, say so in the constants. ByteDance's one-shot example asks for camera movement that is "continuous and smooth, presenting a one-shot flow" [4].

A full 30-second shot costs 1,387 credits at 720p and 617 at 480p, so check the timing at 480p or on a shorter clip first.

**Longer than 30 s:** cut several clips together on the Canvas timeline. ByteDance suggests one continuous take for dialogue within a single scene, and separate clips edited together for plot turns and fast action such as chases, fights and montages [1].

In the Canvas video editor:
- **Storyboard** turns each line into its own clip, up to 8.
- **Between two shots** makes a clip that carries one shot into the next.
- To keep several shots inside *one* clip, write them as a single prompt (in From words or the studio prompt bar). Storyboard would split them into separate clips.

## Using reference images

Where AIGE takes Seedance references:
- **Canvas video editor → Make → With references.** Pick Seedance 2.5 or Seedance (2.0), then add pictures, clips and sounds.
  - Each one gets a token (@Image1, @Video1, @Audio1), numbered in the order you add it. Click a token to put it in the prompt.
  - Canvas makes these clips at 720p, and runs Seedance 2.0 on the Fast tier.
- **Claude or ChatGPT:** use `referenceImages`, `referenceVideos` and `referenceAudios` (see the last section).
- **Studio prompt bar:** there is no Seedance reference mode. A picture there is used as the start frame, with an optional end frame.

The fal Seedance endpoints AIGE uses name references @Image1, @Video1, @Audio1 [6][7]. ByteDance's guides write "@Image 1" or "Image 1" [1][5]. Use the tokens AIGE inserts.

- **Say what each reference provides**, for example "Refer to the camera movement in @Video1" or "Refer to @Image2 for lighting and filters". If a reference already shows exactly what you want, write "Strictly refer to the actions and camera movements in @Video1" and don't describe the moves again [4].
- **Bind every character in the words**, as in "the knight in @Image1". A name written inside the picture is not enough [4].
- **For a person, use a headshot plus a full-body photo** and say which is which. For example: "facial features refer to @Image1 (headshot), styling refers to @Image2 (full-body)". The headshot should show only the face, ideally with no expression and little background [1]. On 2.0, avoid multi-view character sheets, which the model can read as several people [1]. 2.5 supports them [4].
- **Order matters.** Put the reference that matters most first [1]. On 2.5, add characters in the order they first appear [4].
- **Don't fill every slot.** ByteDance recommends 4–5 references in total: 1–2 character pictures (face close-up and full body), 1 scene picture, 1 camera-movement clip and 1 sound [3]. Too many make it hard for the model to tell which features matter most [1].
- **A sound reference needs at least one picture or clip** [6]. AIGE checks this before it charges.

**Keep a saved AIGE character consistent**
- **Studio:** type @ and pick the character. With Seedance, its picture becomes the start frame, unless you have already set one.
- **Canvas:** put the character's headshot and full-body picture on the board (or upload them), add them under With references, then name them in the prompt.
- **Claude or ChatGPT:** pass `characterId`. The character's latest gallery image is used as the start frame. If you also send `referenceImages`, it becomes @Image1 instead.

**Reference clips cost extra.** fal bills the length of reference videos too [6][7], so AIGE multiplies the price by 0.6 × (clip seconds + reference seconds) ÷ clip seconds.
- A 10 s clip with a 5 s reference costs 0.9× the plain price; with a 10 s reference it costs 1.2×.
- Pictures and sounds add nothing.
- Reference clips can add up to 15 s on 2.0 and 30.2 s on 2.5.

fal's file limits:
- 2.0 reference clips: MP4 or MOV, 2–15 s combined, under 50 MB in total, between about 480p and 720p [6].
- 2.5 reference clips: 1.8–30.2 s each, 30.2 s combined [7].
- Sounds: MP3 or WAV [6][7].

ByteDance's own API documentation says Seedance 2.0 and 2.5 do not accept directly uploaded reference images or videos that contain real human faces [2][5]. A run that fails gives its credits back.

## Camera vocabulary

The 2.0 guide says standard terms work: "medium shot, close-up, wide shot, slow push-in, smooth lateral tracking, fixed shot" [1]. It links to a longer lens-language list [8]:
- Moves: tracking shot, pan left/right, truck left/right.
- Shot sizes: long shot, full shot, medium shot, close-up.
- Angles: aerial, high-angle, low-angle, underwater, macro, and "shots with xx as the foreground".

- **Cuts:** name them, as in "cut to…", "slowly push in from a wide shot" or "fixed camera position" [1].
- **Unusual terms:** add what they look like. ByteDance's example explains "rack focus" as the focus sliding from the foreground trees to the character behind them [4].
- **Transitions:** give both when and how [4].
- **Moves in ByteDance's examples:** handheld tracking "with slight realistic camera shake", a drone view looking down, circling halfway around a subject, over-the-shoulder and first-person POV [1][2][4].

## Sound

Seedance generates synchronized sound: "sound effects, ambient sounds, and lip-synced speech", at the same price with or without it [6].

In AIGE, sound is on by default:
- **Studio:** each version has an Audio switch.
- **Canvas:** clips always have sound.
- **Claude or ChatGPT:** pass `audio: false` to turn it off.

ByteDance's symbols for each kind of sound [1][5]:

| Write | For | Example |
|---|---|---|
| ( ) | music | (fast-paced rock music is playing in the background) |
| < > | sound effect | < dog barking can be heard in the distance > |
| { } | dialogue | The girl says {How did the exam go?} |
| 【 】 | on-screen subtitle | 【Chapter One: Departure】 |

- **Languages:** keep dialogue in one language, and don't mix Chinese and English except for proper nouns [1]. For a language other than Chinese or English, name it first: says in Japanese {こんにちは} [1]. 2.5 natively handles Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese and Korean [5].
- **Voice from a reference sound:** write the line and point to the voice, as in "Character" says: "line", voice timbre references "Audio n" [2]. Describe the voice as well, such as "low, warm, middle-aged male voice", and keep the lines close to the reference's tone [1].
- **Switching sound off on 2.5:** write "No BGM; generate only environmental sounds and action sounds", or "No audio" [4].
- AIGE can't lip-sync a finished Seedance clip to a voice track you upload.

## Common problems and fixes

| Problem | What ByteDance recommends |
|---|---|
| Face drifts or changes mid-clip | Use a headshot and a full-body reference and say which is which. Put the most important reference first [1] |
| Subtitles you didn't ask for | Add "Keep it subtitle-free" and "Avoid generating any text or subtitles". Landscape gets fewer than portrait [1]. On 2.5, don't repeat a line's words, write `Name's line (emotion): "…"`, and use reference clips without subtitles [4] |
| Logo or watermark | Add "Do not generate watermarks" and "Do not generate logos" [1] |
| An anime style turns photoreal | Add explicit style words such as "2D Japanese anime style", or convert the reference picture to that style first [1] |
| Two copies of one character | Bind each character to its picture, add an end constraint forbidding identical duplicates, and use single-person photos [1] |
| More than 4 people go wrong | Make pictures with 4 people or fewer, then use them as references [1] |
| Click at the end of a narrated clip | Regenerate, or fade the audio out at the end in an editor [1] |
| Special effect doesn't animate as written | Give a reference clip that shows the effect [1] |
| Voice doesn't match the reference sound | Describe the voice in words and keep the lines in the reference's tone [1] |
| Timestamps ignored on 2.0 | 2.0 follows shot numbers only. Use Shot N, or switch to 2.5 [4] |
| Music despite "no BGM" (2.5) | List every word for it (music, background music, BGM, score, instrumental, melody) and repeat the rule at the start and the end [4] |
| Glowing eyes (2.5) | Add "normal human eyes; no glowing eyes" and use milder words such as "amazed" instead of "extremely shocked" [4] |
| Misspelled on-screen text (2.5) | Have the word appear letter by letter, or supply the text as a reference picture [4] |
| Fingerprint-like texture in grass or leaves (2.5) | Resize the reference picture so it is no larger than the output resolution [4] |

## 8 example prompts

Prices use AIGE's price table; the exact cost shows before you run. "cr" means credits.

**1. 30-second single-take tracking shot**
Seedance 2.5 · studio · 30 s · 720p · 16:9 · 1,387 cr (a 480p test costs 617 cr)
```
One continuous shot, no cuts. A cyclist in a yellow rain jacket rides through an old-town street at dawn.
0-6s: Smooth rear tracking shot at shoulder height. She rides over wet cobblestones past closed shutters.
6-14s: The camera drifts to her left side, still tracking. A baker opens a shop door and waves; she lifts one hand from the handlebar and waves back.
14-22s: The street opens onto a riverside road. The camera rises slowly to reveal a stone bridge in soft golden light.
22-30s: She rides onto the bridge and grows small in the frame; the camera settles into a wide shot and holds.
Throughout: camera movement continuous and smooth, a one-shot flow; soft morning light, natural colors.
<bicycle chain ticking>, <distant church bells>. No BGM. No subtitles.
```

**2. Dance and choreography**
Seedance (2.0) · Canvas → With references · 10 s · 720p, Fast tier · 242 cr base; 1.2× that with a 10 s reference clip
```
Refer to the dance moves and rhythm of the dancer in @Video1 to generate a new scene, keeping the motion details consistent: the dancer from @Image1, in a loose white shirt and black trousers, dances on an empty rooftop at sunset. Fixed camera, full shot. Warm backlight, cinematic texture. Her face and body proportions stay stable; movements are continuous and natural. (upbeat house music) Keep it subtitle-free. Do not generate a logo.
```
This follows ByteDance's motion-reference template, "Refer to the [motion] from [Video N] to generate [scene], keeping the motion details consistent" [1]. The clip carries the choreography, so the words don't have to describe big, fast moves, which ByteDance advises against on 2.0 [1].

**3. Product story in 3 beats**
Seedance 2.0 Pro · studio · product photo as the start frame · 12 s · 1080p · 817 cr (a 480p test costs 162 cr)
```
Shot 1: Medium shot of the matte green thermos from the first frame on a sunlit kitchen counter; the camera slowly pushes in.
Shot 2: Cut to a close-up. A hand enters the frame, unscrews the lid, and steam rises from the tea inside. Fixed camera. <lid unscrewing>
Shot 3: Cut to a wide shot. The thermos sits in the side pocket of a backpack on a mountain trail at sunrise; the camera slowly pulls back. <soft wind>
Warm natural light, rich details, cinematic texture. Keep it subtitle-free. Do not generate a logo. Do not generate a watermark.
```

**4. Short character scene without dialogue**
Seedance 2.5 · studio · type @ and pick your saved character (its picture becomes the start frame) · 10 s · 720p · 463 cr
```
The woman from the first frame sits alone at a window table in a quiet café on a rainy afternoon. She slowly turns a paper cup in her hands, glances at the door, then looks down at her phone. Her tense shoulders drop, she lets out a long breath, and a small smile appears as she types a reply. Medium close-up, fixed camera, then a slow push-in to her face. Soft window light, muted colors, fine film grain. Sound: rain on the glass, <cup set down on the table>, quiet café murmur. No dialogue. No BGM. No subtitles.
```

**5. Nature documentary**
Seedance 2.0 Pro · studio · text only · 10 s · 1080p · 16:9 · 681 cr
```
Shot 1: Aerial wide shot of a misty pine forest at dawn; the camera glides slowly forward over the treetops.
Shot 2: Cut to a low-angle close-up at a forest stream. A red fox steps onto a mossy rock, lowers its head and drinks.
Shot 3: Macro close-up of water drops falling from a fern leaf. Fixed camera.
Nature documentary style, natural colors, soft morning light, HD, rich details. A calm, deep male narrator says: {At first light, the forest belongs to those who move quietly.} <stream water>, birdsong. Keep it subtitle-free.
```

**6. Anime and stylized**
Seedance 2.0 Fast · studio · 8 s · 720p · 194 cr
```
2D Japanese anime style. A girl with short silver hair and a red scarf waits on a train platform in heavy snow.
Shot 1: Medium shot, fixed camera. Snowflakes drift past; her scarf flutters in the wind.
Shot 2: Close-up. She looks up as the train lights appear, her eyes widen slightly, and she smiles.
Shot 3: Wide shot from behind as she steps toward the opening train doors; the camera slowly pulls back.
Hand-drawn line art, cel shading, soft pastel colors. <train brakes hissing>, (gentle piano music). No subtitles.
```
The style words come first, following ByteDance's fix for a style drifting to live action [1].

**7. Fashion walk with a reference image**
Seedance 2.5 · Canvas → With references (two pictures) · 10 s · 720p · 9:16 timeline · 463 cr
```
@Image1 is the model's headshot; @Image2 is her full outfit, a camel wool coat over a black turtleneck. The model from @Image1, wearing the outfit from @Image2, walks toward the camera down a concrete runway lit by one row of overhead lights.
0-4s: Full shot, fixed camera. She walks at an even pace; the coat swings with each step.
4-8s: The camera tracks backward in front of her at the same speed; she turns her head slightly to the left.
8-10s: She stops at the end of the runway, rests one hand on her hip and holds the pose.
Clean editorial look, cool neutral tones, sharp fabric texture. Her face stays consistent with @Image1 throughout. <footsteps echoing>, (minimal electronic beat). No subtitles. Do not generate a logo.
```

**8. Social ad**
Seedance 2.0 Pro · studio · 8 s · 720p · 9:16 · 242 cr
```
Vertical social ad.
Shot 1: Close-up of iced coffee being poured over ice in a clear glass on a bright pink table. Fixed camera; condensation beads on the glass. <ice cracking>
Shot 2: The camera pulls back to a medium shot as a hand slides the glass toward the viewer. Bold white text "COLD BREW DAY" appears at the top center and stays to the end.
Bright saturated colors, crisp commercial lighting. (upbeat pop beat). No other text. Do not generate a logo. Do not generate a watermark.
```
Seedance 2.0 can render ad slogans. You can set the text's color, style, position and when it appears. Stick to common characters and avoid special symbols [1].

## Settings and prices in AIGE

| Setting | Studio prompt bar | Canvas video editor | Claude or ChatGPT (MCP) |
|---|---|---|---|
| Versions | 2.0 Pro, 2.0 Fast, 2.5 | Seedance (runs 2.0 Fast), Seedance 2.5 | `seedance` (2.0 Pro), `seedance-2-5` |
| Resolution | 2.0 Pro: 480p, 720p, 1080p. 2.0 Fast and 2.5: 480p, 720p | 720p | `seedance`: 480p, 720p, 1080p. `seedance-2-5`: 480p, 720p |
| Length | 2.0: 4–15 s. 2.5: 4–30 s | same | same |
| Aspect ratio | auto, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16. On 2.5 with a start picture, the picture sets it | follows the timeline | `aspectRatio` |
| Start and end picture | yes | From a picture, Between two shots | `sourceImage`, `endImage` |
| References | no | With references | `referenceImages`, `referenceVideos`, `referenceAudios` |
| Sound | Audio switch, on by default | always on | `audio`, on by default |

| Version, resolution | Credit key (N = seconds) | 5 s | 10 s | 15 s | 30 s |
|---|---|---|---|---|---|
| 2.0 Fast, 480p | `generateVideoSeedanceFast480p_{N}s` | 54 | 108 | 162 | – |
| 2.0 Fast, 720p | `generateVideoSeedanceFast{N}s` | 121 | 242 | 363 | – |
| 2.0 Pro, 480p | `generateVideoSeedance480p_{N}s` | 68 | 135 | 202 | – |
| 2.0 Pro, 720p | `generateVideoSeedance{N}s` | 152 | 303 | 454 | – |
| 2.0 Pro, 1080p | `generateVideoSeedance1080p_{N}s` | 341 | 681 | 1,021 | – |
| 2.5, 480p | `generateVideoSeedance25_480p_{N}s` | 103 | 206 | 309 | 617 |
| 2.5, 720p | `generateVideoSeedance25_{N}s` | 232 | 463 | 694 | 1,387 |

Every whole second in range has its own price, and sound is included. A clip with reference videos costs more (see "Reference clips cost extra"). A generation that fails gives its credits back.

## Use it from Claude or ChatGPT

Once AIGE is connected over MCP, the tool is `antix_generate_video`. It takes these fields:
- `model`: `"seedance"` (Seedance 2.0 Pro) or `"seedance-2-5"`.
- `prompt`: up to 2,000 characters.
- `duration`: whole seconds.
- `resolution`, `aspectRatio`, `audio`.
- `sourceImage` or `characterId` for the start frame, and `endImage` for the end frame.
- `referenceImages`, `referenceVideos`, `referenceAudios`. Videos and sounds must be HTTPS links; you can upload at aige.ws/r and paste the link.
- `count`: 1–4 separate videos.

References are numbered in this order: the start picture or character comes first, then `referenceImages`, then `endImage`. Poll the result with `antix_check_task`. If you're unsure which model to use, `antix_pick_model` is free.

A realistic ask:

> Use AIGE to make a 20-second Seedance 2.5 video at 480p, 16:9, sound on: "One continuous shot. A lighthouse keeper climbs a spiral staircase at night. 0-8s: Low-angle tracking shot following her up the stairs, lantern swinging. 8-15s: She reaches the top and opens the lamp room door; warm light floods the frame. 15-20s: The camera moves past her to the window and holds on the dark sea. <footsteps on iron stairs>, <wind against glass>. No BGM. No subtitles."

It becomes this call, charged at 412 credits (`generateVideoSeedance25_480p_20s`):

```json
{ "model": "seedance-2-5", "duration": 20, "resolution": "480p", "aspectRatio": "16:9", "audio": true, "prompt": "One continuous shot. A lighthouse keeper climbs…" }
```

## Sources

1. ByteDance / BytePlus ModelArk, "Dreamina Seedance 2.0 series prompt guide": https://docs.byteplus.com/en/docs/modelark/2222480
2. ByteDance / BytePlus ModelArk, "Dreamina Seedance 2.0 series tutorial": https://docs.byteplus.com/en/docs/modelark/2291680
3. ByteDance / Volcengine, "Doubao Seedance 2.0 系列提示词指南" (Chinese original of [1], translated): https://www.volcengine.com/docs/82379/2222480
4. ByteDance / BytePlus ModelArk, "Dreamina Seedance 2.5 prompt guide": https://docs.byteplus.com/en/docs/ModelArk/2607689
5. ByteDance / BytePlus ModelArk, "Dreamina Seedance 2.5 tutorial": https://docs.byteplus.com/en/docs/ModelArk/2607688
6. fal, Seedance 2.0 endpoints AIGE calls: https://fal.ai/models/bytedance/seedance-2.0/reference-to-video (also /text-to-video, /image-to-video and /fast/…)
7. fal, Seedance 2.5 endpoints AIGE calls: https://fal.ai/models/bytedance/seedance-2.5/reference-to-video (also /text-to-video, /image-to-video)
8. ByteDance / BytePlus ModelArk, "Seedance-1.0-pro&pro-fast prompt guide", the lens-language list linked from [1]: https://docs.byteplus.com/en/docs/ModelArk/1631633
9. ByteDance Seed, "Introducing Seedance 2.5" (2026-07-31): https://seed.bytedance.com/en/blog/one-take-creation-flexible-referencing-introducing-seedance-2-5