Seedance 2.0 Complete Prompting Guide
A practical guide to multimodal references, video editing, extensions, shot planning, subject consistency, camera direction, sound, and production-ready Seedance 2.0 prompts.
Seedance 2 models can work with text, images, video, and audio in a single generation. That makes them useful for more than text-to-video: you can preserve a character, borrow a camera move, transfer an action, revise an existing clip, continue a scene, or bridge multiple shots.
The key is to give every input a clear job. Tell the model what each asset represents, what should change, what must stay fixed, and how the scene should unfold over time.
The examples use current Seedance 2-series generation modes available on Veo3 AI. Model availability can change, but the prompting principles and reference-packaging methods remain the same. All example inputs and outputs in this guide are synthetic.
Basic Prompt Formulas
Start by identifying the task. Most Seedance 2 workflows fall into four categories: reference, edit, extend, or a combination of them.
1. Reference Media
Use reference language when an image, video, or audio clip should guide a new generation rather than be directly modified.
Formula
Use [asset] as the reference for [subject / action / camera / style / sound].
Create [new scene and action].
Keep [identity or visual rules].
Avoid [unwanted changes].Be precise about the role of each asset:
Image 1 defines the character's face and wardrobe.Video 1 defines the movement rhythm only.Audio 1 defines the narrator's pace and energy.Do not copy the background or composition from Video 1.
Hands-on example: Flower Shop Conversation
Image 1 — Flower-shop cast and interior

Prompt
Use Image 1 to define the two adults, their wardrobe, and the flower-shop interior. Keep both faces and outfits consistent. Create a short cinematic scene during a rainstorm. Shot 1: a wide locked shot as the customer closes his umbrella. Shot 2: make a gentle push-in while the florist says in natural English {You made it before the rain.} He answers {Barely.} Shot 3: she slides a small bouquet across the counter and they smile. <Warm rain, room tone, and subtle shop ambience.> Accurate English lip sync. No subtitles, Chinese characters, logos, or watermark.2. Edit a Video
Use edit language when the original clip should remain recognizable and only selected elements should change.
Formula
Edit Video 1: change [target element] into [new state].
Keep [camera, timing, subject identity, background, lighting, and all other details] unchanged.Define both the edit and the protected areas. “Make the scene better” is too broad; “remove the bicycle and reconstruct the wall behind it while preserving the camera move” is actionable.
Hands-on example: Remove a Bicycle
Video 1 — Original street clip
Prompt
Edit Video 1: remove the blue bicycle leaning against the brick wall. Reconstruct the hidden bricks and pavement naturally. Keep the pedestrian, shop windows, passing traffic, camera motion, timing, lighting, and ambient English street sound unchanged. Do not add any new object, text, logo, watermark, or Chinese character.3. Extend a Video
Use extend language when you want the model to continue the same scene before or after the source clip.
Formula
Extend Video 1 [forward / backward] by [duration].
Continue from [last visible action, framing, lighting, and sound].
Then show [new action or story beat].Describe the transition first, then the new event. This reduces repeated actions, sudden cuts, and character drift.
Hands-on example: After the Umbrella Closes
Video 1 — Original rainy-station clip
Prompt
Extend Video 1 forward. Begin from the exact final pose, rain intensity, lens height, and street lighting. The woman closes the umbrella, gives it to the man, and says in natural English {Take the long way home.} He nods and walks toward the warm station entrance while she remains under the awning. Preserve both identities and wardrobe. <Continuous original rain and street ambience.> No subtitles, Chinese characters, logo, or watermark.4. Combine Tasks
A prompt can assign different roles to several assets. Keep the instructions modular so the model can distinguish identity, motion, camera, and audio.
Formula
Use Image 1 for the character.
Use Video 1 for the camera move only.
Use Audio 1 for pacing only.
Create [new scene].
Keep [protected details].
Do not inherit [unwanted details from the references].Hands-on example: Snow Bridge Transition
Image 1 — Snowy-station design

Video 1 — Left-to-right camera-motion reference
Prompt
Use Image 1 for the snowy station architecture, traveler, coat, and color palette. Use Video 1 only for its left-to-right camera motion. Shot 1: show a close-up of a completely blank, unprinted paper ticket with no letters, numbers, barcode, or marks; it breaks into snowflakes. Shot 2: the flakes sweep left to right and fully cover the lens. Shot 3: they clear to reveal the same traveler beside an arriving train in the station from Image 1. <Only original wind, train, and paper sounds.> No speech, visible text, Chinese characters, logo, or watermark.Advanced Prompt Formula
A production-ready prompt usually combines spatial direction with a clear timeline.
[Task and references]
+ [subject definitions]
+ [scene and environment]
+ [shot-by-shot timeline]
+ [body movement and interaction]
+ [camera direction]
+ [lighting, style, and image quality]
+ [dialogue, ambience, music, and sound effects]
+ [hard constraints]Write in the order a director would brief a crew: identify the cast, establish the location, describe what happens, explain how the camera observes it, then lock sound and constraints. Treat exact timestamps as rough guidance: shot labels are usually more reliable than asking the model to hit a precise second.
Complete Formula in Practice
Hands-on example: Mechanical Bird at Dawn
Image 1 — Mechanical-bird design and first frame

Prompt
Use Image 1 as the first frame and design reference for one brass mechanical bird. Salt lake at dawn, distant mountains, mirror reflections. Shot 1: locked wide shot as the bird tilts its head. Shot 2: it opens both wings in one continuous mechanical action; individual brass feathers catch the sunrise. Shot 3: it runs three steps and lifts just above the water while the camera tracks right at the same speed. Photoreal fantasy film, pale blue and rose palette, precise reflections. <Delicate servo sounds and wind.> No speech. Keep one identical bird throughout. No extra wings, duplicated subject, text, Chinese characters, logo, or watermark.1. Define Every Subject
Give each subject a stable label
Name recurring subjects once, then use the same label throughout the prompt.
Subject A: the adult violinist from Image 1, wearing a dark green coat and carrying a silver case.
Subject B: the station attendant from Image 2, wearing a charcoal uniform and red scarf.Avoid switching among “the woman,” “the musician,” and “she” when several people are present. Stable labels reduce identity swaps.
When several assets describe one subject
Explain what each image contributes.
Images 1 and 2 show the same violinist. Image 1 defines her face and hair. Image 2 defines the full outfit and silver instrument case.Hands-on example: Station Violinist
Image 1 — Standalone character photograph

Prompt
Use Image 1 to define Subject A: the same adult violinist, face, short dark hair, dark green coat, violin, and silver case. Keep her identity unchanged in every shot. Shot 1: Subject A tunes one string under the departure board. Shot 2: she begins a gentle original melody while commuters pass behind her. Shot 3: close-up on her calm expression and bow hand. (Original gentle solo violin.) <Natural station ambience.> No speech. Do not create a second violinist, duplicate the case, alter her coat, show readable signs, Chinese characters, logos, or watermarks.2. Use a Shot Timeline
A list of events is not automatically a usable sequence. Divide the action into shots and specify what the camera should see in each one. Use precise seconds only when a rough shot order is not enough; exact timestamp compliance can vary between generations.
Weak
The courier runs through the paper city, crosses a bridge, and escapes the guards.Directed
Shot 1: wide view of the folded-paper city; the courier enters from frame left.
Shot 2: low tracking shot as she crosses the blue paper bridge.
Shot 3: overhead shot; two guards reach the bridge as it folds closed behind her.Keep each beat small enough to complete within its time range. Preserve screen direction unless a deliberate reversal is part of the story.
Hands-on example: Paper City Chase
Image 1 — Storyboard and paper-craft style guide

Prompt
Use Image 1 only as a storyboard plan and paper-craft style reference. Never display the storyboard grid, panel borders, or multiple panels in the generated video. Shot 1: a single full-screen wide view of the folded-paper city; the red-coated courier enters from frame left. Shot 2: cut to one full-screen low lateral tracking shot as she crosses the blue folded bridge. Shot 3: cut to one full-screen overhead shot; exactly two guards stop as the bridge folds shut behind her. Maintain left-to-right movement, one courier, two distinct guards, and the same paper textures. <Crisp fold sounds and light original percussion.> No dialogue, visible text, Chinese characters, panel grid, split screen, logo, or watermark.3. Describe Action Precisely
Name the body part, direction, and intensity
Instead of “she reacts,” write what the viewer can observe:
She raises her right eyebrow, draws one short breath, and turns her shoulders ten degrees toward the door.Prefer connected actions
Break complex movement into linked steps. Explain the end of one action and the start of the next.
He plants his left foot, blocks once above his shoulder, lowers the blade, then takes one controlled step backward.Translate emotion into visible behavior
“Nervous” can become a tightened grip, shallow breathing, delayed eye contact, or one step toward the exit.
Hands-on example: Rooftop Fencing Drill
Image 1 — Fencers, wardrobe, and rooftop layout

Prompt
Use Image 1 for the two adult fencers, rooftop layout, and wardrobe. Subject A advances with the right foot, makes one diagonal feint toward Subject B's left shoulder, then withdraws the blade. Subject B turns the wrist, blocks once, pivots clockwise on the left foot, and steps out of range. Repeat no action. One continuous exchange, controlled training speed, realistic balance and blade contact, gentle handheld camera, original metallic impacts and city wind. No dialogue, text, Chinese characters, logos, duplicated limbs, or watermark.4. Direct the Camera
Camera instructions work best when they include a starting frame, movement path, speed, and ending frame.
Begin in a high wide shot above the cliff.
Descend in a slow forward arc toward the glass library.
Pass the upper terrace, rotate twenty degrees clockwise, and end at eye level beside the entrance.Avoid stacking incompatible moves in a short clip. Choose one primary move and one subtle adjustment.
Hands-on example: Island Library Descent
Image 1 — Island-library environment

Prompt
Use Image 1 for the island, glass library, weather, and scale. Begin in a high wide aerial shot. Descend in one slow forward arc toward the roof, pass the terrace, rotate twenty degrees clockwise, and end at eye level beside the entrance. Keep the horizon stable and the movement continuous. Late-afternoon sun, realistic ocean reflections, restrained orchestral swell and waves, no dialogue, no visible text, no Chinese characters, no logo, no watermark.5. Lock Quality, Style, and Constraints
Quality terms should support the scene rather than replace direction.
Quality
Clean facial detail, stable hands, controlled highlights, natural motion blur, consistent exposure.Style
Hand-painted ink animation on textured rice paper, restrained black and indigo palette.Constraints
One subject only. No duplicated fins, no extra objects, no readable text, no logo, no watermark.Put the most important constraints close to the related instruction, then repeat only the critical ones at the end.
Hands-on example: Ink Fish Formation
Image 1 — Ink-wash style and fish-design reference

Prompt
Use Image 1 only for the hand-painted black-and-indigo ink style, rice-paper texture, and design of exactly three fish; do not use it as the first frame. Shot 1: begin with exactly three separate ink drops on blank textured rice paper; each drop spreads into one fish silhouette. Shot 2: the three fish swim in one clockwise circle, leaving soft capillary trails. Shot 3: they merge into one clean wave symbol. Locked overhead camera. <Soft brush, paper, and water sounds.> Exactly three fish before the merge. No photorealism, extra fins, readable text, Chinese characters, border, logo, or watermark.6. Direct Dialogue and Sound
State who speaks, the exact line, the language, vocal character, and what should happen after the line.
The watchmaker says once in natural English {Time always leaves a clue.}
Warm older male voice, measured pace, precise lip sync.
After the final word, he closes his mouth and looks down at the watch.Separate dialogue, ambience, music, and sound effects. If you want silence, say so directly. When a voice reference matters, upload it as its own audio file instead of treating a clip extracted from a video frame as a character reference.
Hands-on example: The Watchmaker's Clue
Image 1 — Watchmaker, customer, and workshop

Audio 1 — Warm older male voice sample
Prompt
Use Image 1 for the same two adults and watch shop. Use Audio 1 only for the watchmaker's warm older male timbre, measured pace, and close recording distance; do not copy background noise. Fixed medium shot with a subtle push-in. The watchmaker lifts a pocket watch and says once in natural English {Time always leaves a clue.} Match Audio 1's vocal character with accurate lip sync. Show the same sentence once as a small cream English subtitle near the bottom, then fade it out before the final frame. <Quiet ticking and room tone.> No music, additional speech, repeated subtitle, Chinese characters, logo, or watermark.Other Practical Techniques
Generate On-Screen Text
For a slogan, subtitle, label, or speech bubble, specify the exact English copy, location, timing, typography, color, and animation.
After the sailboat clears the first bridge, reveal the exact English words "FOLLOW THE LIGHT" in the upper center.
Use a thin ivory sans-serif font with wide tracking.
Fade in over half a second and remain stable until the end.Do not ask the model to invent important brand copy. Provide the exact line.
Hands-on example: Follow the Light
Image 1 — Paper-sailboat first frame

Prompt
Use Image 1 as the first frame. A small glowing paper sailboat moves slowly through the dark canal while the camera tracks beside it. After the sailboat clears the first bridge, reveal the exact English words "FOLLOW THE LIGHT" in the upper center, thin ivory sans-serif, wide tracking, gentle fade-in, stable until the end. (Original quiet ambient music.) <Water and soft wind.> Do not add any other words, Chinese characters, logos, or watermark.Choose Between Extension and Separate Clips
Use continuous extension for one location, gradual emotion, a single conversation, or movement along one path. It helps preserve immersion and motion continuity.
Use separate clips and editing for location changes, rapid pursuit, complex action, or montage. Generate each segment with a clear entry and exit frame, then cut on movement, occlusion, or an empty shot.
Hands-on example: Before the Coffee Spill
Video 1 — Original café-spill clip
Prompt
Extend Video 1 backward. Generate only the events that happen before the first frame of Video 1; do not replay the spill. The new clip plays forward in time. Shot 1: the same café table is clean and dry, both guests sit calmly, and the server walks toward them carrying two upright cups on a level tray. Shot 2: the man begins sliding his chair backward while the server is still one meter away. Shot 3: the server notices the chair, tightens her grip, and says once in natural English {Wait—}. End at the source clip's opening state, just as the chair touches the tray and one cup begins to tip. Before the final moment there must be no spilled coffee, falling cup, or shocked reaction. Preserve the same three adult identities, wardrobe, table layout, lens, warm lighting, and camera position. <Natural café ambience.> No subtitles, visible text, Chinese characters, logo, or watermark.Plan the Reference Package
Upload only assets that provide useful information. More references are not automatically better.
- Use one clear close portrait plus one separate full-body image for a recurring character.
- Do not use a multi-view character sheet as a single people reference; repeated views can be interpreted as several people.
- Use front, side, and detail views for a product.
- Use a short video when timing, action, camera, or an effect matters more than appearance.
- Tell the model what not to inherit from every reference.
- Put assets in a deliberate order and keep the same numbering in the prompt.
Hands-on example: Vintage Bookstore Package
Image 1 — Bookstore architecture and lighting

Image 2 — Standalone bookseller portrait

Image 3 — Rolling-ladder object reference

Prompt
Use Image 1 for the bookstore architecture and amber lighting. Use Image 2 for the adult bookseller's face and burgundy cardigan. Use Image 3 for the brass rolling ladder. Shot 1: the bookseller rolls the ladder one meter. Shot 2: she climbs two steps. Shot 3: she takes a blue book from the shelf and says in natural English {Found it.} Keep one bookseller, the same ladder, and the same store layout. <Quiet room tone and wood movement.> No subtitles, Chinese characters, brand marks, duplicated people, or watermark.Important Notes
Language
Use direct, literal instructions. Keep dialogue inside quotation marks and state the spoken language explicitly. For this guide's examples, all dialogue and visible copy are English only.
Asset References
Use consistent labels such as Image 1, Video 1, and Audio 1. Do not rename the same asset halfway through a prompt.
Special Characters
Use the model's media markers consistently when they help separate sound and text instructions:
- Put music direction inside parentheses:
(Original restrained piano.) - Put sound effects or ambience inside angle brackets:
<Rain, footsteps, and room tone.> - Put spoken dialogue inside braces:
{Time always leaves a clue.} - Put a required subtitle inside corner brackets:
【LOW TIDE, NEW WORLDS.】
Keep the content inside these markers in English for an English-language result. Avoid decorative symbols that do not have a clear production role. If a name is hard to pronounce, give the visible spelling and a short phonetic cue.
Duration and Density
Match the number of actions to the available duration. A short clip can usually support two or three connected beats, not a full commercial storyboard. Prefer Shot 1, Shot 2, and Shot 3; treat exact seconds as approximate rather than guaranteed edit points.
Text and Watermarks
If text is not required, state no visible text, no logo, no watermark. If text is required, provide the exact English copy and prevent additional writing.
Common Problems and Fixes
Character identity drifts
Symptoms: the face, age, hairstyle, or wardrobe changes between shots.
Fixes: use fewer, clearer character references; provide one close portrait plus one separate full-body image; define the subject once; repeat protected identity details near each shot; reduce action amplitude when consistency matters more than spectacle. Do not use a multi-view sheet of the same person as one input.
Unwanted subtitles appear
Fixes: write no subtitles, captions, lower thirds, or visible text. Remove subtitle-like text from reference images when possible.
Logos or watermarks appear
Fixes: request no logo and no watermark. Avoid using source media with platform marks or visible brand graphics unless they are essential.
The visual style changes mid-clip
Fixes: lock the medium, palette, texture, and rendering method. For animation, explicitly prohibit photorealistic transformation.
An extension jumps or repeats
Fixes: describe the exact source endpoint, the first new action, screen direction, camera motion, and sound continuity. Tell the model not to replay the final source action. For backward extension, state that the new clip plays forward in time and must end at the source opening state. If joining clips manually, try trimming about six unstable ending frames from the first clip and one unstable starting frame from the next before testing the cut.
A character duplicates
Fixes: state the exact number of people, assign stable labels, and add do not create a second version, double, reflection clone, or background duplicate of Subject A. Replace collage or multi-view references with one clean standalone portrait or full-body image.
Quality degrades after several extensions
Fixes: limit chained extensions. Use the cleanest earlier segment as the next reference, or generate a new shot and join it at an intentional cut.
A described effect behaves incorrectly
Fixes: use a short effect reference and specify that only the trajectory, speed, and formation order should transfer—not the original subject or background.
Too many reference characters cause swaps
Fixes: keep primary characters to a manageable number, use distinct wardrobe and silhouettes, and assign each person a separate image and label.
The final frame contains an audio click
Fixes: ask for a clean audio tail and no abrupt cut. Add a short fade at the end of the generated audio and a matching fade at the beginning of the next clip when joining material in editing software.
English pronunciation is inaccurate
Fixes: shorten the line, use familiar words, add a phonetic cue for names, spell out numbers when necessary, and request one delivery only at a measured pace.
The generated voice does not match the reference
Fixes: describe pitch, age range, pace, accent, emotional energy, and recording distance. Ask the model to borrow vocal qualities without copying unrelated background sound.
Appendix: Prompt Examples
The following examples cover text generation, image, video, and audio references, editing, extension, and track completion. Each card includes every reference asset named in the prompt, an English prompt, and the generated result.
Text Generation
English Subtitle: The Tide Remembers

Use Image 1 as the first frame. The marine biologist studies the tide pool while the camera makes a slow push-in. After she lowers the sample jar, show the exact English subtitle 【LOW TIDE, NEW WORLDS.】 in small white documentary type at the lower center. Keep it stable briefly, then fade out. <Ocean ambience and soft wind.> No speech, other text, Chinese characters, logo, or watermark.Speech Bubble: Greenhouse Robot

Use Image 1 as the visual reference for exactly one small white greenhouse robot and the tomato greenhouse. Shot 1: the robot waters one tomato plant; no bubble or text is visible. Shot 2: it stops the water, checks the soil, and looks toward camera. Shot 3: only after the watering action is complete, add one cream speech bubble above the robot containing the exact English words "WATERING COMPLETE" in dark green rounded type. Keep the bubble stable through the ending. <Light greenhouse ambience.> No spoken dialogue, extra text, Chinese characters, logo, or watermark.Speech Bubble: Strawberry Garden

Use Image 1 as the first frame. Shot 1: the farm drone scans the strawberry basket and flashes a green light while the farmer laughs naturally. Shot 2: after the scan finishes, show one white speech bubble beside the drone with the exact English words "READY TO PICK!" in bold red rounded type. <Light garden ambience.> No spoken dialogue, additional text, Chinese characters, brand marks, or watermark.Image References
Multi-angle Product: Brass Desk Lamp



Images 1–3 show the same brass desk lamp from the front, side, and rear. Preserve its proportions, cream shade, braided cord, and rotary switch. Create a slow premium product orbit on a dark walnut desk. After the camera passes the side profile, the lamp turns on once. <A subtle mechanical click and quiet room tone.> No speech, visible text, Chinese characters, logo, or watermark.Home Product: Ceramic Diffuser

Use Image 1 to define the exact diffuser shape and material. Morning light moves slowly across the travertine shelf while a thin ribbon of mist rises and bends toward the window. Gentle five-degree camera slide, premium home-film style, quiet room tone, no speech, text, Chinese characters, logo, or watermark.Logo Reference: Paper Plane Brand Film

Use Image 1 only for the white paper-plane symbol. Place the same symbol on a deep cobalt background. It folds from one sheet, lifts into the air, and leaves a thin silver trail that forms the English tagline "IDEAS TAKE FLIGHT" beneath it. Clean motion-design film, precise edges, no extra text, no Chinese characters, no watermark.Multiple Subjects: Fox and Crow


Image 1 defines Subject A, the red fox mechanic. Image 2 defines Subject B, the black crow navigator. In a miniature forest workshop, Subject A tightens one brass bolt while Subject B lowers a map onto the table. They exchange one glance and continue working. Preserve one fox and one crow, distinct silhouettes, stop-motion felt style, no speech, text, Chinese characters, logo, or watermark.Individual Storyboard Images: Moon Kitchen



Images 1–3 are consecutive storyboard frames. Follow their composition in order. Shot 1: wide shot of the lunar kitchen as the adult astronaut enters. Shot 2: medium two-shot; the older cook offers a floating dumpling and says in natural English {Dinner still counts up here.} Shot 3: close-up as the astronaut catches it and laughs. Accurate English lip sync. <Soft kitchen ambience.> No subtitles, Chinese characters, logos, or watermark.Video References
Marketing Motion: Dew to Perfume

Use Image 1 for the unbranded bottle design. Use Video 1 only for the liquid transformation timing. A dew drop rolls along a leaf, falls in slow motion, and becomes the same glass perfume bottle when it touches a black stone pedestal. Macro camera, cool morning light, precise liquid sound, no speech, text, Chinese characters, logo, or watermark.Game Effect: Glass Butterfly Wings

Use Image 1 to define exactly one adult game character, her braided hair, black tactical coat, rooftop environment, and camera framing. Use Video 1 only for the glass-particle formation order and timing; do not copy its mannequin or black studio. Shot 1: the character stands alone with no wings. Shot 2: fine glass particles gather behind both shoulders and form two symmetrical transparent wings. Shot 3: the wings lock into place with one blue pulse; immediately show the exact English HUD text "WINGS ONLINE" once, then remove it before the final frame. Preserve one identity and one pair of wings. <Original glass chimes and a low energy pulse.> No speech, additional text, Chinese characters, duplicated person, extra wings, body distortion, logo, or watermark.Video Editing
Add an Element: Flowers on a Bench
Edit Video 1: add one terracotta pot of small yellow flowers to the right side of the bench. Match the perspective, contact shadow, wind, and afternoon light. Keep the seated reader, leaves, camera movement, timing, and all other objects unchanged. No text, Chinese characters, logo, or watermark.Replace an Element: Cello Case
Edit Video 1: replace the traveler's brown suitcase with a full-size black cello case. The hand must grip the cello-case handle naturally and the case must move with realistic weight. Preserve the traveler, face, coat, station, camera, timing, lighting, and sound. No other changes, no text, Chinese characters, logos, or watermark.Video Transition and Track Completion
Fill the Gap: Snowflake Transition
Unlike the earlier image-plus-motion-reference example, this workflow supplies two boundary clips. The model's job is to generate only the missing section between their endpoints.
Fill the gap between Video 1 and Video 2. This is a track-completion workflow: generate only the missing bridge, not a replay of either source clip. Begin with the blank ticket in Video 1 breaking into snowflakes. Continue the same left-to-right motion as the flakes cover the lens. Clear them to reveal the snowy platform in Video 2, matching its first-frame camera height, traveler position, and train sound. <Original wind, paper, and train sounds only.> No repeated source frames, speech, visible text, Chinese characters, logos, or watermark.Start Creating
Open the Seedance video generator, choose text-to-video or reference-to-video, then begin with one clearly defined example from this guide. Change one variable at a time between generations so you can see which instruction improved the result.
Seedance 2.5 Prompting Guide
Build clearer Seedance 2.5 prompts with a five-part production brief, continuous timelines, multimodal reference roles, sound direction, and failure-focused constraints.
MiniMax H3 Prompting Guide
Master MiniMax H3 video prompting through prompt structure, reference-asset roles, camera, sound, complete examples, and generated results.