Creative Space
Write Better Seedance 2.5 Video Prompts

Konten

Start from an ineffective prompt/result or an idea with references. Diagnose using Seedance's writing method and deliver usable prompt cards. With agreement, compare original/revised trial generations in Flova.

Tools and references

  • Flova: Current parameters/reference requirements and optional trials.
  • Optional image tools/FFmpeg: Padding to supported ratios or relevant crops.
  • , steps 2–4: suitability, four-part structure, references, character anchors, timelines, editing/extensions/transitions, sound/consistency, failures, current Flova limits.
  • , steps 4–6: format, uploads/roles, reference wording, comparisons.

Boundaries

  • Preserve user-approved dialogue/on-screen text verbatim: no translation, polishing, additions.
  • Prompts can improve success probability, not guarantee outcomes. Do not call an untested rewrite definitely better.
  • Use authorized references; do not generate real celebrity likenesses or copy copyrighted characters.
  • Explain quota use and obtain agreement before trial generation.

Workflow

Collect → check capabilities → diagnose → rewrite/confirm → optional comparison → deliver.

1. Collect

Obtain any existing prompt, purpose/main event/style, each reference's intended responsibilities/exclusions, duration/aspect ratio, single continuous or multiple shots, dialogue/music/effects, prior unsatisfactory results (screens/video if available), and operation type: new generation, edit, extend, or transition between videos.

Resolve gaps using model-guide section 2. Check suitability in section 1 first; if unsuitable, say so and suggest another model/method.

2. Current capabilities

Query creation_generation_models for Seedance 2.5, detail="full": duration, aspect ratio, resolution, reference quantity/type/roles. If unavailable, mark parameters unverified and continue prompt work. Check reference compliance, especially ratios (guide section 8). Explain necessary fixes or pad with user agreement.

3. Diagnose

Create a table of issue, writing rule, likely consequence:

  • Four sections present: reference roles, overview, timeline, preservation constraints?
  • References specify only one/two responsibilities each, rather than “use everything”? Any competing identity controls?
  • Actions/shots/dialogue fit duration? A short clip should focus on one core action.
  • Shot instructions match continuous versus cut intent? Orbit/follow specifies direction?
  • Action has start/process/end rather than teleportation?
  • Character appearance anchors repeated after cuts?
  • Dialogue specifies speaker/language/emotion? Explicit silence when no speech?
  • Positive/negative requirements conflict, or excessive negatives?

For supplied failures, first use guide section 7's matching failure/correction.

4. Rewrite — confirmation

Use the four-part structure and . Name each reference ordinally with responsibilities/exclusions matching submission order. Use guide section 3's timeline template. For edit/extension/transition, follow section 4 and specify only changes/additions.

Explain changes against each diagnosis. For long content, optionally provide complete/concise versions and when to use each. Show for confirmation before trial or delivery.

5. Optional comparison

After agreement, Quick-generate original and revised versions once each with identical parameters/references. Show side by side against acceptance criteria. For remaining failures, revise one variable per round, usually stopping within two/three rounds. If still poor, explain why, e.g. overly complex action should split into two clips.

6. Deliver

One/more copy-ready cards, diagnosis/change notes, and, if tested, per-round changes/task IDs/results and recommended version. Remind the user that generation varies and several samples may be needed for selection.

Seedance 2.5 writing method

The duration, scale, resolution and number of reference assets actually supported by the model are subject to the query results of creation_generation_models.

1. When is it appropriate to use it?

Suitable for:

  • Videos that require long narratives, continuous action, or obvious emotional changes.
  • The complex task of simultaneously referencing characters, products, scenes, actions, camera movement, timbre or music.
  • Realistic characters, commercials, short dramas, action dances, game demos and multiple characters in the same frame.
  • Migrate motion, shot tracks, rhythm or atmosphere based on reference videos.
  • Edit existing videos, continue the plot, or design a natural transition between two clips.
  • Multi-language dialogue is required, and images, lip sync, timbre and sound scenes are to be generated collaboratively.

Not suitable for:

  • A single prompt for a very long story with too many locations, characters, and plot twists.
  • Tasks that require pixel-level certainty for every text, logo, or final frame.
  • For tasks that only have a simple action and require higher generation speed, a lighter video model can be considered.

2. Ask users for clear needs

Before you start writing, ask questions from users:

  • The purpose, subject matter, overall style and core events of the video.
  • Protagonist, scene, key props and identifying characteristics that must be maintained.
  • I hope that one shot will lead to a multi-shot narrative.
  • How actions occur, how the camera moves, and how emotions change.
  • Whether there is dialogue, environmental sounds, action sound effects or music; the dialogue needs to provide the exact original text.
  • What each reference asset wishes to control and what it does not wish to inherit.
  • If editing or extending, clarify which parts of the original video must be kept and where the new content begins.

3. Prompt writing method

Four-stage structure
完整 Prompt = 参考素材职责 + 一句话概述 + 具体时间线 + 全局保持项
  1. Refer to assets responsibilities: Explain that each asset is responsible for characters, costumes, products, scenes, actions, camera movement, timbre or style.
  2. One sentence summary: subject + location + core event + theme style + special shot language.
  3. Specific timeline: Write pictures, scenes, camera movement, actions, dialogue and sounds according to event stages.
  4. Global preservation constraints: Reaffirm character characteristics, props, environment, light, color, camera position and content that cannot be changed.

The number of shots, action density and dialogue length must match the length of the video. Short videos should only retain one core action; longer videos should also revolve around a clear line of events.

Refer to how to write assets responsibilities
第一张参考图负责女主角的脸、发型和服装;第二张负责庭院布局与晨光。
参考视频只提供剑术动作和手持摄影节奏,不继承其中的人物外观和场景。
参考音频负责女主角的说话音色。
  • Don't say "completely refer to all assets", explain what is migrated and what is excluded.
  • Character identities, actions, scenes, timbres and artistic styles should be handled by different assets as much as possible to reduce competition for control.
  • Multiple characters are described one by one with character references and distinguished by name, clothing, body shape or position.
  • The blockout model, line drawing, storyboard diagram or preview video can be responsible for composition, positioning, movement path and shotblocking; the final material and appearance will be explained separately.
Real people and character descriptions

When there is no clear character reference, or the texture of the writing needs to be enhanced, describe:

  1. Age, regional appearance and skin color.
  2. Facial structure, identifying features and natural skin texture.
  3. Eyes, breathing, micro-expressions and emotional changes.
  4. Hair style, hair color and hair condition.
  5. Clothing style, material, new and old condition and accessories.
  6. Body shape, posture, temperament and position in the shot.

Rather than writing a list of isolated adjectives, connect appearance and performance:

一名二十多岁的东亚女性,皮肤保留自然毛孔与柔和高光。
她始终注视对面的人,呼吸克制,下唇轻微颤抖,直到镜头末尾才形成一滴泪。
Timeline template
全局设定:
- 环境与质感:[时间、地点、天气、空间和物理质感]
- 视觉语言:[写实/动画/广告/电影感、景深、光线和色彩]
- 摄影机:[主要景别、视角、连续运镜或切镜方式]
- 人物锚点:[外观、服装、道具和参考职责]
- 表演重点:[最重要的动作或情绪变化]

开场:[建立空间、人物初始状态、镜头起点和声音]
发展:[动作如何承接、摄影机如何跟随、台词或声音如何进入]
高潮:[关键动作、情绪转折或信息揭示]
收束:[最终动作、构图、视线方向和声音结束状态]
  • Only write continuous spatial paths, character movements, and camera movements from one shot to the next. Do not mix in hard-cut shot structures.
  • Each switch between multiple shots re-explains the characters, locations, scenes and scene relationships.
  • Use First / Then / Finally or natural event node organization, there is no need to cram in too dense time points for accuracy.
Complete example
第一张参考图负责女主角的脸、发型、浅蓝色汉服和银色长剑;
参考视频只提供舞剑动作节奏,不继承人物和背景。

一段电影感的中式庭院剑舞,晨光柔和,全程采用连续镜头。
开场为大全景:女主角从画面左侧缓慢进入,长剑低垂,摄影机以步行速度向右跟拍,花瓣从身后飘落。
随后镜头自然推至中景,她拔剑并完成参考动作,脸、服装和长剑设计保持不变。
结尾推至近景,剑刃停在脸侧,几片花瓣划过前景,她平静呼吸并望向镜头。
全片无对白、无音乐,只保留脚步、风声和一次短促金属声。

4. How to describe editing, extension and transition

Video Editing

Write only the goals that need to be changed, and make it clear that the rest remains the same:

将人物手中的透明玻璃杯替换为磨砂银色保温杯。
保持人物身份、动作节奏、构图、运镜、光线、背景、对白和其他细节不变。
新物体要持续匹配原视频的透视、遮挡、阴影、反射、运动模糊和手部接触关系。
Video extension

Extend the description and only write the new clips, do not repeat the plot that has already happened in the source video:

从原视频最后的姿态、摄影机方向、光线、运动惯性和声音状态自然接续。
女主角缓慢放下长剑,望向院门并后退两步;摄影机延续轻微推进。
结尾保持她面向院门。无对白、无音乐,保留风声和衣料摩擦声。

Write down the extension direction, starting continuous state, new actions, camera continuation method, final state and sound. Extending multiple times in a row can easily amplify the character and texture drift, so you should try to plan a complete new event at once.

Transition between two videos
从前一段视频末尾的动作、运动方向、光线和声音开始,
通过[遮挡经过镜头/镜头甩动/相似形状匹配/光线闪白]自然连接到后一段视频的首个状态。
保持主体运动惯性和画面方向连续,不新增无关人物或物体。

5. Sound and dialogue writing

  • The dialogue must be faithful to the confirmed script or storyboard word for word, without translation, polishing, or rewriting.
  • For each dialogue, clearly describe the speaker, language, emotion, speaking speed, and the on-screen or off-screen state.
  • The length of dialogue should match the rhythm of the shot; short shots should not carry long sentences.
  • Make it clear that characters remain quiet when there is no dialogue, and indicate whether to retain ambient sounds and action sound effects.
  • Music is not required; describe style, instrumentation, mood, and entry/exit events when music is required.
  • Instead of describing a mix in abstract percentages, write it as "vocals are clear up front, music is muffled, and ambient sounds are in the distance."
  • When a dialogue spans shots, make it clear that it continues from the previous shot, and explain the speaker's changes inside and outside the shot.

6. How to improve performance and consistency

  • Establish a stable anchor point for the character: face, hairstyle, clothing outline, accessories, body shape and voice should not change with the shot.
  • Each reference only has one or two clear responsibilities, and duplicate or conflicting assets are removed.
  • Complex movements are broken down into starting postures, movement processes, contact relationships and ending postures.
  • Camera movement is written as direction, speed and following object, for example "slowly move sideways to the left while panning right to keep the character centered".
  • Restate characters, props, spatial position and main light direction when switching between multiple shots.
  • The product video repeatedly emphasizes the structure, material, color, logo, packaging text and placement direction.
  • Use events to control sounds, such as "a metallic sound appears the moment the sword is drawn, and the sound disappears after the sword stops."
  • Only the most important negative constraints are retained; conflicting positive and negative requirements significantly reduce stability.

7. Common failures and corrections

  • Character changes: Reduce conflicting references, fix one main identity map, and recap appearance anchors after shot switches.
  • Action like teleportation: Complete the start, process and end states, specify the speed, center of gravity, contact and inertia.
  • A cut occurs at the end of a shot: Delete shot, cut and multi-shot jumps, and only retain continuous action and continuous camera movement.
  • Wrong shot direction: Split abstract words such as "surround, follow" into camera movement direction, panning direction and subject position.
  • Editing irrelevant areas: Shorten the modification target and write out the content that must be maintained item by item.
  • The extended clip cannot be connected: Supplement the posture, eyeline, motion inertia, light, camera position and sound status at the end of the source video.
  • Multiple character face swapping or crosstalk: Fix names, appearances, positions and voice references one by one to avoid the same assets taking on multiple roles.
  • The dialogue is too fast or the lip sync is unstable: shorten the dialogue, reduce the number of same-shot events, and clarify the speaker and the precise original text.
  • Music or subtitles appear unexpectedly: Use positive states to explain whether characters speak, how the sound scene is composed, and review non-deterministic results after generation.

8. Known limitations on Flova

The following is from the actual call. It may change after the model is updated. The query result of creation_generation_models shall prevail before submission:

  • Reference image aspect ratio: The aspect ratio of each reference image should be between 0.39 and 2.50 (width ÷ height). When exceeded, the model will directly reject the entire request. For example, a three-view of 1338×520 (2.57) will be rejected. For images that are too wide or too tall, first fill the edges with solid white space to within the range, and do not crop to avoid cutting off people or products.

Prompt Cards and Flova Handoff

These inspirations deliver prompt cards, not images/videos: one per image or shot. Cards can be used directly for generation; this reference covers writing, Quick generation, and validation.

Card format

Include all fields; write “none” where absent:

FieldContent
IDE.g. P01, mapped to the requested image/source shot.
Model/capabilityE.g. Seedance 2.5 multimodal video; GPT Image 2 editing.
PurposeOne-sentence intended use.
PromptComplete copy-ready text in one code block.
ReferencesOne row each: filename, submission order, role below, responsibilities, exclusions.
ParametersDuration, aspect ratio, resolution, etc., from creation_generation_models; otherwise “unverified.”
InvariantsUnchanged people, products, text, composition.
RationaleTwo/three sentences linking choices to model-writing rules.
Acceptance checksE.g. identity, verbatim text, absence of teleporting motion.

+Write in the target language. Chinese-user requests default to Chinese unless the model guide/user calls for English. Preserve dialogue/on-screen text verbatim; never translate/polish it.

Quick generation

+Follow Flova Skill's Quick/reference guidance:

  1. Query the model through creation_generation_models, detail="full"; verify parameters, reference slots/roles.
  2. Prepare references: supported aspect ratio/padding, crops, formats.
  3. creation_upload_ticket → transfer bytes to its URL → creation_upload_complete → file_ref. It expires in roughly one hour; upload again before submission when necessary.
  4. Assign each inputs role: first_frame starts matching that image; last_frame ends on it; reference guides appearance/style/content without exact matching; driving_audio drives lips/action; base_video is edited/continued.
  5. Submit with client_request_id; retain card/reference/file_ref/task-ID mappings.

Referencing inputs in prompts

  • Use ordinal natural language, e.g. first image controls face/clothing, second controls product. Match inputs order within each media type; images/videos/audio each count from first.

Other tools

+Remove Flova-specific markers/role instructions and use the destination tool's upload mechanism and parameter controls. Explain that these prompts target a particular model; changing models may change outcomes.

Optional comparison

+Explain quota use before proceeding. Generate original and revised prompts once each with identical parameters/references; show side by side. Change one variable per round (e.g. reference roles or timeline), record edits/task IDs/results, and deliver the record. Generation is stochastic: one comparison shows a tendency; repeat important findings once or twice before concluding.

Proses kreatif

  1. Flova

    Flova

    Generate a test clip

    Test the revised prompt and inspect action, camera movement, and transitions.