Seedance 2.0 Complete Guide: Features, Comparison & How-to Tutorial

What is Seedance 2.0?

Seedance 2.0 is ByteDance's advanced multimodal AI video model, combining images, videos, audio, and text inputs for unprecedented creative control. This complete guide compares Seedance to Kling, Veo, and Sora, and shows professionals how to master multimodal video workflows on Morphic.

Unlike traditional text-to-video models that rely solely on written prompts, Seedance 2.0 enables you to show the AI exactly what you want through visual and audio references. Upload reference images to define style and composition, use video clips to demonstrate desired camera movements or actions, add audio to establish mood and rhythm, and combine everything with detailed text prompts for precise creative direction.

Two companion guides go deeper on specific problems: cinematic Seedance 2.0 videos for the craft side, and what to do when Seedance 2.0 prompts get flagged for generations that come back rejected.

Why Seedance 2.0 for professional video creation

Seedance 2.0 addresses the fundamental limitation of AI video generation: the gap between description and vision. Instead of trying to describe complex camera movements, character details, or visual effects in words, you can provide direct examples. This multimodal approach delivers:

• Precise visual control through image references
• Accurate motion replication via video references
• Rhythm and mood synchronization with audio integration
• Consistent character and style across multiple shots
• Complex scene transitions that maintain continuity

The model excels at understanding and combining multiple reference types simultaneously, making it particularly valuable for commercial production, content creation, and professional video workflows.

Seedance 2.0 vs Kling vs Veo vs Sora: Feature comparison

When evaluating AI video generation tools, understanding the specific capabilities of each platform helps inform the right choice for your workflow. Here's how Seedance 2.0 compares to leading alternatives:

Key differentiators:

Multimodal flexibility: Seedance 2.0 and Kling 3.0 both offer comprehensive multimodal support including direct video and audio file uploads. Veo 3.1 supports image references (up to 3) but audio is generated rather than referenced. Sora remains primarily text and image-based.

Video reference depth: Seedance 2.0 and Kling 3.0 excel at replicating complex camera movements, choreography, and special effects from reference footage. Kling 3.0's "AI Director" feature automates multi-shot scene composition. Veo 3.1 focuses on image-to-video with strong character consistency but less emphasis on video-to-video motion replication.

Audio capabilities: Seedance 2.0 allows direct audio file upload for precise mood control and beat synchronization. Kling 3.0 generates native multi-language audio with accurate lip-sync across 5 languages. Veo 3.1 generates audio natively but doesn't accept audio file references. Sora generates audio from text descriptions only.

Duration and extension: While Sora offers the longest single generations (up to 60 seconds), Veo 3.1's extension feature allows chaining clips beyond 60 seconds. Seedance 2.0 and Kling 3.0 both support 15-second generations with extension capabilities.

Resolution and quality: Seedance 2.0 outputs up to 4K with 10-bit encoding, which preserves colour gradation for HDR and grading work. That puts it level with Kling 3.0 and Veo 3.1 on raw resolution, so the deciding factor is usually reference control rather than pixel count.

Editing existing footage: Seedance 2.0 treats an uploaded video as something you can edit, not just imitate. You can add, remove, or swap elements inside a clip, extend it forward or backward in time, and stitch separate clips into one continuous piece. Editing is a first-class task here rather than a bolt-on, which is what makes it practical to keep footage a client has already approved.

Which Seedance 2.0 model to use

Seedance 2.0 is a family of three models that share the same feature set and differ on quality and cost:

A practical pattern: draft on Mini or Fast until the prompt and references are behaving, then run the final pass on full Seedance 2.0.

Information accurate as of July 2026. Features and availability subject to change.

Try Seedance 2.0 on Morphic →

Key features and capabilities of Seedance 2.0

Multimodal input system

Seedance 2.0 accepts four distinct input types that work in combination:

Image Inputs (Up to 9 images)

• Define visual style and aesthetic direction
• Establish character appearances and maintain consistency
• Set scene composition and framing
• Specify product details for accurate reproduction
• Control lighting, color grading, and atmosphere

Video Inputs (Up to 3 clips, maximum 15 seconds combined)

• Reference specific camera movements and cinematography
• Replicate motion patterns and choreography
• Copy scene transitions and editing rhythms
• Demonstrate special effects and visual techniques
• Show character actions and interactions

Audio Inputs (MP3 or WAV, up to 3 files, each 2-15 seconds, maximum 15 seconds combined)

• Set mood and emotional tone through music
• Control pacing with rhythm and beat structure
• Add specific sound effects or ambient audio
• Match voice characteristics for dialogue
• Synchronize visual changes to audio cues

Text Prompts (Natural language)

• Guide narrative and story progression
• Specify actions and movements not shown in references
• Describe scene transitions and timing
• Clarify how references should be applied
• Add details beyond what visual references show

On limits: there is no combined file cap. The ceilings are per type (up to 9 images, 3 video clips, 3 audio files), and the real constraint is not the ceiling but your restraint: ByteDance recommends using far fewer references than you are allowed. One combination is genuinely unsupported, though. Text plus audio alone, and audio alone, will not generate. Every job needs text, and audio only ever works alongside a visual reference.

Reference capability architecture

The core innovation in Seedance 2.0 is its reference understanding system. Rather than treating inputs as simple style guides, the model analyzes and extracts specific elements from each reference:

From Images: Composition structure, character features, object details, lighting setup, color relationships, spatial arrangement, style characteristics

From Videos: Camera motion paths, movement speed and acceleration, shot framing changes, subject actions and timing, special effect implementation, transition techniques

From Audio: Rhythm and beat patterns, tonal mood and atmosphere, volume dynamics, sound effect timing, voice characteristics

This granular understanding allows you to specify exactly which aspects of each reference should influence the generation, creating precise control over the final output.

Core generation quality improvements

Beyond multimodal capabilities, Seedance 2.0 delivers foundational enhancements:

Realistic Physical Dynamics: Objects and characters move with authentic physics. Clothing drapes naturally, liquids flow convincingly, and interactions between elements follow real-world rules.

Smooth Motion Performance: Continuous action flows without jarring transitions or morphing artifacts. Complex multi-step movements maintain consistency throughout execution.

Precise Prompt Understanding: The model accurately interprets detailed instructions, including spatial relationships ("in the background behind"), ordered shot sequences, and complex multi-subject scenarios. Precise timecodes are the exception, and are covered below.

Consistent Style Retention: Visual characteristics established at the start of a generation remain stable throughout. Character appearances, lighting conditions, and artistic style don't drift as the scene progresses.

Complex Action Execution: Handles challenging sequences like fight choreography, detailed hand movements, facial expressions during speech, and coordinated multi-character interactions.

Ready to experience multimodal control? Start creating with Seedance 2.0 on Morphic →

Technical specifications

Two notes on the output settings. There is no 2.35:1 option, so for a scope look, choose 21:9 rather than asking for 2.35:1 in the prompt. And because 4K uses H.265, some browsers and players will not preview it directly even though the file is fine.

Understanding Seedance 2.0 input specifications

File count and duration limitations

To optimize generation quality while managing computational resources, Seedance 2.0 implements specific input constraints:

Individual File Type Limits:

• Images: Maximum 9 files
• Videos: Maximum 3 clips
• Audio: Maximum 3 files

Combined Duration Limits:

• Video references: 15 seconds total across all clips
• Audio references: 15 seconds total across all files

Per-clip minimums (easy to miss):

• Every reference video clip must be at least 2 seconds long
• Every reference audio file must be at least 2 seconds long
• Reference video must run at 24 to 60 fps

There is no combined file cap. Nine images plus three videos plus three audio files is a legal request. It is also a bad one, for the reasons in the next section.

Unsupported combination: text plus audio, and audio on its own, will not generate. Audio always needs a visual reference alongside it.

Generated output duration: 4-15 seconds.

Do not fill the slots you are given

The limits above are ceilings, not targets. ByteDance's own guidance is explicit that using the full asset allowance degrades results: with too many references, the model struggles to judge which features take priority, and you get style conflicts, blurry subject identification, and output that drifts from what you asked for.

The recommended configuration is 4 to 5 assets in total. Each one should do a distinct job:

• Character anchoring (1-2 images): lock the character's appearance
• Scene tone-setting (1 image): lock the environment and style
• Camera movement (1 video): lock the shot language and rhythm
• Rhythm and atmosphere (1 audio): control emotion and timbre

If two references are competing for the same job, you have one too many. Uploading four images of the same character does not make the model four times more confident about that character. It makes the model less certain about which image to believe.

Practical example: for a 15-second commercial requiring specific product appearance, dynamic camera work, and upbeat music, five assets is a complete brief:

• 2 images: product from two angles
• 1 image: desired colour grading and lighting style
• 1 video: camera movement reference
• 1 audio: music track for pacing

Everything else, including the environment, the mood, and the pacing, belongs in the text prompt.

A hard ceiling on characters: when more than four reference people are involved, output stability drops sharply. You start seeing the wrong number of people in frame, or duplicates. If you need a crowd of six, generate them as two grouped images of three first, then use those images as the reference for the video.

Input quality guidelines

For Image References:

• Use clear, well-lit photographs when accuracy matters
• Higher resolution provides better detail reproduction
• Multiple angles help for products and objects. They actively hurt for faces: see the headshot rule below, because a multi-angle character sheet is read as several different people
• Avoid heavily compressed or low-quality images

For Video References:

• Ensure the specific element you want to reference is clearly visible
• Shorter clips focused on one aspect work better than longer clips with multiple elements
• Higher quality video improves motion understanding
• Trim videos to show only the relevant section

For Audio References:

• Use clean audio files without background noise when possible
• Ensure audio clearly demonstrates the rhythm or mood you want
• Match approximate duration to your target video length
• Consider using audio from video files if it serves multiple purposes

How to use Seedance 2.0 multimodal references

Seedance 2.0 is accessible through Morphic, which provides an interface for uploading references and writing prompts. The system uses an @ mention structure to specify how each uploaded file should be used in generation.

First, know which of the three tasks you are asking for

Before writing anything, decide which job you are giving the model, because the phrasing differs and getting it wrong is the most common way to waste a generation.

• Multimodal reference: pull elements out of your assets (a subject, a style, a camera move, a voice) to make a brand-new video.
• Video editing: change something inside an existing video. Anything you do not mention stays as it was.
• Video extension: continue an existing video forward or backward in time, keeping the style, subject, and narrative intact.

The trap: for editing and extension tasks, refer to the clip as @Video 1, never as "reference @Video 1". The word reference pushes the model to classify the job as a reference task, so instead of editing your clip it generates a new video that merely resembles it. Say "Strictly edit @Video 1" or "Extend @Video 1 backward", not "Reference @Video 1 and edit it".

Define your subject before you reference it

An uploaded image usually contains more than one thing. Naming the subject explicitly is the step most people skip, and it is the one that prevents the model from latching onto the wrong element.

Pick 2 to 3 stable, static features (clothing, hairstyle, category) that uniquely identify the subject, then bind them to the asset:

Define [core features] in @Image 1 as [Subject name]

For example: Define the woman in the red dress and straw hat in @Image 1 as Subject 1. From that point on, refer to her only as Subject 1. Use the same label every single time she appears. If you switch between "the woman", "she", and "Subject 1", the model can lose track of who you mean.

For quick jobs you can skip the definition and bind inline instead, repeating the pairing each time you mention her: Subject 1@Image 1.

Two rules that matter here. Keep the description short and free of contradictions, because two conflicting features on the same subject will confuse the binding. And where a spatial relationship is easier to show than to say, show it in a reference image rather than describing it in text.

The @ reference system

After uploading your materials to Morphic, you reference them in your prompt using the @ symbol followed by the file identifier (Image 1, Video 1, Audio 1, etc.). The key is explicitly stating what purpose each reference serves.

Basic Reference Structure:

@[Material Type + Number] as/for [specific purpose], [additional context]

Clear vs Unclear Referencing:

Unclear: "Use @Image 1 and @Video 1 to make a video"

Clear: "@Image 1 as the opening frame showing the character's face, reference the camera push-in movement from @Video 1, use @Audio 1 for background music to establish an upbeat mood"

Put your most important asset first. The earlier an asset appears in the prompt, the more weight it carries. If the character's face has to be exactly right, lead with it.

Punctuation that tells the model what kind of information it is reading

Seedance 2.0 reads four bracket types as type signals. Using them is a cheap way to stop dialogue being rendered as an on-screen caption, or a sound effect being spoken aloud.

If a line of dialogue is in a language other than English, name the language before the braces: says in Japanese {こんにちは}. Keep one language per scene. Mixing languages in the same prompt, proper nouns aside, degrades pronunciation.

Writing effective multimodal prompts: The CRAFTS framework

A good Seedance 2.0 prompt is closer to an engineering instruction than a piece of copywriting. The model reads your text, images, video, and audio at the same time and splits them internally into a spatial layer (what is in the frame) and a temporal layer (how it changes over time). Your job is to feed both layers cleanly.

CRAFTS covers the six things a complete prompt specifies: Context, Reference, Action, Framing, Timeline, Style and constraints.

C - Context: Establish Scene and Environment Set the stage with location, time period, atmosphere, and overall setting. Include references to scene images here.

Example: "In a dimly lit jazz club at night, referencing the interior atmosphere from @Image 1"

R - Reference: Define Your Subjects and Bind Them to Assets Name each subject by 2 to 3 stable features, then state exactly what each asset contributes and what it does not.

Example: "Define the man in the grey suit in @Image 2 as the Pianist. @Image 2 for his face and build only, not his clothing. @Video 1 for the walking pace. @Audio 1 for the background jazz."

A - Action: Describe Character and Object Movements Detail what happens in the scene: character actions, object interactions, and event sequence. Two things separate action descriptions that work from ones that do not.

Go small and slow. Seedance 2.0 handles gentle, continuous, small-scale movement far more reliably than high-energy motion. Sprints, big jumps, and violent rolls are where limbs deform and physics breaks. Prefer "walks slowly", "gently raises a hand", "sits down with the motion". Specify the body part, plus the range, speed, and force: not "he moves his head" but "he turns his head quickly". Where one action leads into another, say how they connect, so the movement carries its own inertia: "uses the momentum of the turn to naturally raise a hand".

Externalise emotion into physical detail. The model cannot render "very sad". It can render what sadness does to a body. Replace the abstract label with the tell.

Example: "The Pianist walks slowly across the room, stops at the bar, and picks up a glass. He takes a sip, and as he looks toward the door his jaw tightens and his fingers close hard around the glass."

F - Framing: Define Camera Work and Cinematography Specify shot types, camera movements, angles, and transitions using cinematic terminology. The model has a strong grasp of standard camera vocabulary, so use it plainly.

Example: "Start with a wide establishing shot, dolly in to a medium close-up as the character reaches the bar, then cut to an over-the-shoulder shot looking toward the door"

One camera move per shot. Asking a single shot to push in, orbit, and pan at once is the fastest way to destabilise the image. If you want three moves, you want three shots.

T - Timeline: Sequence your shots, not your seconds

This is the element most guides get wrong, including earlier versions of this one. The instinct is to timestamp everything: 0-4 seconds this, 4-8 seconds that. Seedance 2.0's support for precise timing is unstable, and forcing exact durations onto segments can actively break the generation.

Sequence the shots instead, and let the model find the pacing. Label them Shot 1, Shot 2, Shot 3, in the order events happen, and describe each one as: camera move, then subject action, then position, then sound.

Weak: "A man runs nervously down the street, and the scene feels very cinematic."

Strong:

• Shot 1: Side view of a street alley. The man starts to run, breathing fast.
• Shot 2: He knocks over a fruit stand. The camera shakes and pushes to a close-up of his frightened face.
• Shot 3: He climbs a low wall and disappears. The camera pulls back slowly and settles on the empty street.

Timestamps are not forbidden and you will see them in ByteDance's own sample prompts. Treat them as a hint the model may honour, not an instruction it must obey. If the pacing genuinely matters, the reliable lever is to generate fewer shots per clip, not to write tighter timecodes.

S - Style and constraints: Close the door on failure modes

The last block of your prompt is the one people leave out, and it is doing more work than they realise. It has three jobs.

Image quality: name the finish you want. "High definition, rich detail, filmic texture, natural colour, soft light."

Style: name the look, and name it even when you think the reference image already implies it. If your reference photo is realistic but you want animation, and you never say so, the output will drift back toward live action. Say "2D Japanese animation style" or "3D stylised CG" explicitly.

Constraints: these are the lines that stop the model producing artefacts nobody asked for.

• Keep it subtitle-free. Avoid generating any text or subtitles.
• Do not generate a logo.
• Do not generate a watermark.

Worth being straight about the limits here: constraint lines reduce the probability of stray subtitles and watermarks, they do not eliminate it. Two things that genuinely help. If a reference image or clip has text in it that you do not need, strip the text out of the asset before you upload it. And if your project allows it, generate in landscape, because spurious subtitles appear significantly less often in landscape than in portrait. You can always crop to vertical afterwards.

CRAFTS Example Prompt:

Image reference techniques

Setting visual style and aesthetic direction

Images establish the overall look and feel of your generation. Use them to define color palettes, lighting approaches, compositional style, and artistic treatment.

Maintaining character consistency across shots

When generating multiple videos featuring the same character, reference the same character image in each prompt to maintain appearance consistency.

Product showcase with accurate details

For commercial or product-focused content, use multiple angles and detail shots as references to ensure accurate reproduction.

Video reference techniques

Replicating camera movements and cinematography

Video references excel at demonstrating specific camera techniques that are difficult to describe in text alone.

Copying motion patterns and choreography

For dance, fight sequences, or specific movement patterns, video references provide frame-by-frame motion guidance.

Replicating special effects and visual techniques

Video references can demonstrate particle effects, transitions, compositing techniques, and other visual effects for accurate reproduction.

Audio reference techniques

Background Music Integration and Mood Setting

Audio references establish the emotional tone and pacing of your video through music selection.

Beat Synchronization for Visual Changes

Sync scene transitions, cuts, or visual changes to specific musical beats for polished, professional pacing.

Voice Timbre and Dialogue Matching

When specific voice characteristics matter, reference audio or video files containing the desired voice quality.

Complex multi-reference examples

Combining All Input Types for Commercial Production

Example: Product Commercial

Multi-Character Scene with Dialogue

Example: Narrative Scene

Advanced Seedance 2.0 features

Video extension for continuous narratives

Seedance 2.0 can extend an existing video with new content that continues the story. It works in both directions: you can generate what happens after a clip, and you can generate what happened before it, which is useful when you have a strong shot but no way into it.

How Video Extension Works:

1. Upload your existing video
2. Say which direction you want, and describe what happens
3. The model extracts the transition frames itself and generates the continuation

Refer to the clip as @Video 1, not "reference @Video 1". This is an editing task, and the word reference can push the model into generating a lookalike video instead of continuing yours.

Example: Extending a Coffee Shop Scene

Existing Video: 10-second clip of a person sitting at a cafe table, looking at a laptop

The original footage is not regenerated, so the part you already like stays exactly as it is. By default the output carries only the tail of the source clip, not the whole thing. If you want the original included in the returned video, ask for it: "Extend @Video 1 backward, [description], and then end with @Video 1."

Two limitations to plan around

Extension is genuinely useful, but it is not invisible, and you will save yourself a bad afternoon by knowing this up front.

Joins can jump. At the point where the new segment meets the original, you may see a small jump or a frame of rollback. There is no prompt that reliably fixes this today. The practical repair is in your editor: trim about 6 frames from the end of the outgoing clip and 1 frame from the start of the incoming one, at every join. Better still, plan your clips so the join lands on a cut rather than mid-motion, where a discontinuity is invisible by design.

Quality decays if you chain extensions. Feeding a generated video back in as the source for another extension degrades the image, and the damage compounds with each pass. It shows up first as mottled colour blotches on faces. Keep the chain short and feed in the highest-quality source you have.

If you genuinely need to extend repeatedly, ByteDance documents an odd but effective workaround: convert the clip into a white-model video first, then continue from that. The prompt is roughly "Convert the video into a white 3D model. All characters become pure white 3D models, no colour, no texture, no shadows, pure white background, stable structure, smooth motion." Stripping the colour and texture gives the next pass a clean structural base to build on instead of compounding the artefacts of the last one.

Extension Best Practices:

• Keep extensions relatively short (5-8 seconds) for best continuity
• Clearly describe the connecting action between the original end and new content
• Mention elements that should remain consistent (camera angle, lighting, character position)
• If the original video has audio, reference that audio style for the extension

Video fusion and multi-clip transitions

Create seamless transitions between multiple existing video clips by generating bridging content.

Example: Connecting Two Locations

Existing Videos:

• @Video 1: Character walking in urban street (ends with character approaching corner)
• @Video 2: Same character entering apartment (starts with door opening)

This generates a bridge clip that smoothly connects two separate shoots, maintaining character and narrative continuity.

Character replacement in existing videos

Swap characters or subjects in videos while preserving all other elements including camera work, motion, and scene details.

Example: Music Performance Replacement

Use Cases for Character Replacement:

• Testing different talent in commercial concepts
• Creating variations of the same scene with different actors
• Updating existing footage with new brand ambassadors
• Producing content for different regional markets with localized talent

Storyline subversion and narrative alteration

Completely change the narrative direction or outcome of existing video while maintaining the visual and technical elements.

Example: Relationship Drama Reversal

Original Video (@Video 1): Romantic scene where man proposes to woman on a bridge, she says yes, they embrace

This technique allows complete narrative redirection while preserving the cinematography and production value of existing footage.

One-take continuous long shots

Create seamless long-take sequences that follow subjects through multiple environments without cuts.

Example: Urban Chase Sequence

Creative template replication

Copy the structure, style, and techniques from reference videos while substituting your own subjects and branding.

Example: Adapting Commercial Style

Reference: @Video 1 shows a high-end perfume commercial with specific camera techniques, transitions, and pacing

Seedance 2.0 use cases and examples

This section demonstrates Seedance 2.0 applications across different industries and complexity levels. Each industry includes basic, intermediate, and advanced examples showing progressive skill development.

Commercial and advertising production

Basic: Single Product Static Showcase

Scenario: Simple product display for e-commerce

Complexity Level: Single image reference, one camera move, one event

Intermediate: Multi-Angle Product Demo

Scenario: Tech product demonstration showing multiple features

Complexity Level: Multiple images, several shots, varied camera angles, audio sync

Advanced: Full Commercial with Scene Transitions

Scenario: 15-second lifestyle commercial showing product in use across multiple settings

Complexity Level: 5 images, 1 video, 1 audio. Layered audio, three locations, professional cinematography.

This example sits at the top of the recommended asset range, and the home office is described in words rather than given its own reference image. That is a deliberate trade. A three-location commercial is exactly where the temptation to upload a picture of everything is strongest, and exactly where doing so starts to blur the subject.

Social media content creation

Basic: Trending Style Quick Cut Video

Scenario: Simple social media content with popular transition effect

Complexity Level: Multiple image references, beat synchronization, simple transition effect

Intermediate: Multi-Location Story Sequence

Scenario: Day-in-the-life vlog style content

Complexity Level: Multiple locations, handheld style reference, audio layering, personality-driven content

Advanced: Viral-Style Complex Visual Effects

Scenario: High-production social media content with trending effects

Complexity Level: Multiple complex references, precise choreography matching, special effects replication, advanced audio sync, trending style integration

Film and entertainment production

Basic: Atmospheric Establishing Shot

Scenario: Scene-setting shot for narrative content

Complexity Level: Single image, basic camera movement, atmosphere building

Intermediate: Dialogue Scene with Shot Reverse Shot

Scenario: Two-character conversation with professional coverage

Complexity Level: Two character images, specific camera technique reference, dialogue pacing, psychological tension

Advanced: Action Sequence with Complex Choreography

Scenario: Fight scene with specific martial arts choreography

Complexity Level: Four image references, two video references (choreography and camera style), audio reference, complex action choreography, slow motion.

Set your expectations honestly here. Fast, high-impact action is the hardest thing to ask of this model, and it is where limbs deform and physics gives way. That is exactly why this prompt leans on a choreography reference video rather than trying to describe the kick in words: showing the motion is far more reliable than describing it. Even so, expect to generate this one several times. If a shot keeps failing, the fix is usually to slow the action down or break it into two shots, not to add more adjectives.

Professional workflow applications

Video Extension for Project Continuity

Scenario: Extending previously shot footage with additional content

Existing Video: 8-second shot of CEO walking through modern office, ending at conference room door

Use Case: Adding to existing professional video assets without reshoots

Template-Based Bulk Content Creation

Scenario: Creating multiple social media videos with consistent style

Master Template Prompt (Video 1):

Variation Prompts: Replace @Image 1 with different products while maintaining @Image 2, @Image 3, and @Audio 1 for brand consistency

Use Case: Scalable content production for product catalogs, maintaining brand identity across multiple assets

Multi-Language Adaptation

Scenario: Creating regional variations of the same commercial

Base Prompt:

Use Case: International marketing campaigns requiring localized versions with consistent visual branding

Best practices for Seedance 2.0

The CRAFTS prompting framework (detailed)

Professional results in Seedance 2.0 require structured prompt engineering. The CRAFTS framework provides a systematic approach that ensures all critical elements are specified:

C - Context: Establish Scene and Environment

Define where and when the action takes place. This includes:

• Physical location and setting
• Time of day or historical period
• Atmospheric conditions (weather, lighting quality)
• Overall mood and tone
• Environmental details that matter to the story

Example: "In an underground nightclub at 2 AM, with the moody atmosphere from @Image 1. Hazy air from smoke machines, a low ceiling lit by the glow of the bar, packed dance floor in the background."

R - Reference: Define Subjects, Then Specify @ Mentions and Exact Purpose

This is where multimodal power lives. Name the subject first, then be explicit about what each reference contributes:

• Define each subject by 2 to 3 stable features, and reuse that label every time
• State the @ mention clearly
• Specify exactly what aspect of that reference to use
• Clarify what NOT to use if the reference contains multiple elements
• Lead with the asset that has to be most accurate

Example: "Define the woman with the shaved head and leather jacket in @Image 1 as Ada. @Image 1 for Ada's facial features and hair only, not the clothing. @Image 2 for the leather jacket costume. @Video 1 for the walking pace and confident stride. @Audio 1 for the electronic background music."

A - Action: Describe Character and Object Movements

Detail what happens in the scene: the verbs of your video:

• Character movements and gestures
• Object interactions (picking up, setting down, throwing)
• Facial expressions and emotional reactions
• Interactions between multiple subjects
• Physics-based events (things falling, liquids pouring, smoke rising)

Example: "Character enters from frame left, walking with the confident stride from @Video 1. Eyes scan the crowd briefly, then lock onto someone off-screen. Slight smile forms. Character adjusts jacket collar with right hand, then begins moving forward through the crowd with purpose."

F - Framing: Define Camera Work and Cinematography

Use proper cinematography terminology to specify shot composition:

• Shot types: Wide shot, medium shot, close-up, extreme close-up, over-the-shoulder, point-of-view
• Camera movements: Dolly in/out, tracking shot, pan left/right, tilt up/down, crane up/down, handheld, steadicam
• Angles: Low angle, high angle, eye level, dutch angle
• Special techniques: Hitchcock zoom, whip pan, rack focus, shallow depth of field

Example: "Open with wide shot establishing the full nightclub environment. As character enters, camera picks up and begins tracking alongside in medium shot. When character stops to scan crowd, push in slowly to medium close-up. Cut to character's POV shot looking through crowd. Cut back to close-up of character's face as smile forms. Resume tracking shot as character moves through crowd, camera following from behind."

T - Timeline: Order Your Shots and Let the Model Find the Pacing

Break your sequence into numbered shots, in the order events happen:

• Label them Shot 1, Shot 2, Shot 3, most important first
• Give each shot one camera move, one beat of action, one place
• Put the sound for that shot inside that shot
• Do not assign each shot a duration

Precise timing is the model's weakest control surface. Pinning segments to exact second counts can produce worse output than leaving them alone, so describe the order of events and let the model breathe.

Example: "Shot 1: Wide shot of the club. Ada enters and starts walking. Shot 2: The camera tracks alongside her in a medium shot as she scans the crowd. Shot 3: Close-up as a small smile forms. Shot 4: Her POV, looking through the crowd. Shot 5: The camera resumes tracking from behind as she moves forward. Throughout: (electronic music from @Audio 1 at moderate volume)."

S - Style and Constraints: Lock the Look and Rule Out the Artefacts

Close every prompt with the finish and the guardrails:

• Image quality: "high definition, rich detail, filmic texture, soft light"
• Style: name it explicitly, even if a reference image implies it
• Constraints: "Keep it subtitle-free." "Do not generate a watermark." "Do not generate a logo."
• Stability: "The character's face stays consistent, with no deformation. Motion is smooth, with no stutter or flicker."

Complete CRAFTS Example: Corporate Training Video

Input preparation strategy

Image Reference Optimization

Quality input creates quality output. Prepare image references strategically:

For Character Consistency:

• Use a face-only headshot, shot straight-on, neutral expression, minimal background
• Add one full-body shot for costume and proportion. That is the whole set
• Do not add more angles. A front/side/three-quarter turnaround makes identity drift worse, not better, because the model reads the angles as different people
• Avoid heavy filters or effects that might confuse the model
• If the costume matters, let the full-body shot carry it rather than adding detail crops

For Style and Aesthetic:

• Select images that clearly demonstrate the desired visual treatment
• Ensure color grading is consistent with final vision
• Include images showing the specific lighting approach you want
• Consider texture and detail level: high detail references produce high detail outputs

For Products and Objects:

• Photograph against simple backgrounds for focus
• Show multiple angles to ensure accurate reproduction
• Include close-ups of important details (logos, textures, specific features)
• Ensure lighting shows form and dimension clearly

Video Reference Optimization

For Camera Movement:

• Trim videos to show only the specific camera move you want to replicate
• Ensure the movement is clearly visible and not obscured by action
• Shorter clips (3-5 seconds) focused on one technique work better than longer clips with multiple techniques
• Use highest quality video available: compression artifacts affect understanding

For Motion and Choreography:

• The action should be clearly visible without obstruction
• Ensure lighting adequately shows body position and movement
• Multiple angles of the same action can help if available
• Consider slowing down fast movements when creating reference clips

For Special Effects:

• Isolate the specific effect you want to replicate
• Ensure effect is clearly visible against background
• If effect has specific timing, include that timing in reference

Audio Reference Optimization

For Music and Rhythm:

• Use high-quality audio files (avoid low-bitrate compressed audio)
• Trim audio to the section with the most relevant rhythm or mood
• Ensure audio clearly demonstrates what you want (beat, pace, mood)
• Consider starting audio at a strong beat for easier synchronization

For Voice and Dialogue:

• Use clear recordings with minimal background noise
• Ensure the specific vocal characteristic you want is prominent
• Keep reference clips short and focused on the relevant vocal quality

Choosing your four or five assets

There is no combined file cap, so nothing stops you uploading nine images. The discipline has to come from you. Use this to decide what earns a slot.

The four functional roles. A complete brief usually needs one of each, and rarely more:

1. Character anchoring (1-2 images): locks who the subject is
2. Scene tone-setting (1 image): locks the environment and style
3. Camera movement (1 video): locks the shot language and rhythm
4. Rhythm and atmosphere (1 audio): locks emotion and timbre

Four questions to cut the rest:

1. Will removing this reference actually change the result? If not, cut it.
2. Can this be said in text instead? Environments, moods, and colour grades usually can. Faces, exact camera moves, and specific choreography usually cannot. Upload what is hard to describe. Describe everything else.
3. Is another asset already doing this job? Two references competing for the same role is worse than one.
4. Is this a must-have or a nice-to-have? Cut the nice-to-haves first.

Example decision process

You are making a music video and have fifteen candidate references:

• 4 images: the artist from different angles
• 3 images: the performance venue
• 2 images: lighting setups
• 2 videos: dance choreography and camera movement
• 2 audio files: music track and ambient sound
• 2 images: costume details

Applying the framework:

• Keep (character anchoring): 1 headshot of the artist, plus 1 full-body shot. Not the four-angle set. A multi-angle sheet is the single most reliable way to make the model think it is looking at several different people
• Keep (scene): 1 venue image, the most representative
• Keep (camera): 1 video. Choose between the choreography and the camera-movement clip based on which one you could not write down in words
• Keep (rhythm): the music track
• Describe in text: both lighting setups, the costume details, the ambient audio, and the other two venue images

Result: 5 assets. The other ten pieces of information still reach the model, they just arrive as words instead of files, which is exactly where they belong.

Consistency techniques for multi-shot projects

Character Consistency Across Generations

Maintaining the same character appearance across multiple video generations requires systematic reference management:

Use a headshot plus a full-body shot. Nothing more.

This is the single most counter-intuitive rule in the model, and getting it wrong causes the exact problem you were trying to solve.

The instinct is to give the model a character turnaround: front, side, three-quarter, back. Do not. Seedance 2.0 frequently reads a multi-view sheet as several different people. That makes identity drift worse, not better, and it is a leading cause of the "twins" bug, where two identical copies of your character show up in the same frame.

What works is two images with clearly separated jobs:

• A face-only headshot. Head only, neutral expression, minimal shoulders, neck, or background. This is the image the model leans on for identity, so the face needs to dominate the frame. If the face is a small region inside a busy full-body shot, the model under-weights it and the surrounding clutter bleeds in.
• A full-body shot for styling, costume, and proportion.

Then split the roles explicitly in the prompt, and put the face first:

Note that this applies to faces. For products and objects, multiple angles are genuinely useful and you should supply them.

The mixed-reference mistake: do not hand the model one composite image that crams the face, the pose, the outfit, and a detail crop into a single picture. The face ends up occupying a small fraction of the pixels, gets a correspondingly small share of the model's attention, and identity drifts.

If your character drifts anyway, there is a consequence worth knowing about: a drifting face can wander toward resembling a real public figure, and that can get the generation blocked at review even though you never asked for a celebrity. So identity drift is not only a quality problem. It is one of the quieter reasons a clean-looking prompt comes back rejected. Our guide to Seedance 2.0 prompts getting flagged covers the rejection side in detail.

Keeping a consistent label Use the same name for the character in every prompt, and specify "maintaining exact appearance from @Image [X]".

Style Consistency Across Scenes

For projects requiring multiple shots with consistent visual treatment:

Technique 1: Style Reference Template Select one image that perfectly captures your desired visual style:

• Color grading
• Lighting approach
• Composition style
• Texture and detail level

Include this same style reference in every generation prompt:

Technique 2: Previous Output as Reference Use earlier successful generations as references for later shots:

Temporal Continuity for Sequential Shots

When creating shots that connect sequentially:

Technique 1: Overlap Description Describe how the new shot connects to the previous:

Technique 2: Transition Specification Clearly state the connection point:

Common pitfalls to avoid

Pitfall 1: Vague Reference Usage

Problem: "@Image 1 as reference" without specifying what aspect to reference

Solution: Always state exactly what the reference provides: "@Image 1 for character's facial features and expression, not the background or lighting"

Pitfall 2: Contradictory Instructions

Problem: "Fast-paced action scene with slow, contemplative camera movements and calm ambient music"

Solution: Align all elements: action pace, camera energy, music tempo, editing rhythm: toward a consistent goal

Pitfall 3: Over-Complicating Prompts

Problem: Uploading a dozen barely-differentiated references and writing a 500-word prompt full of conflicting details. ByteDance's guidance is blunt about this: do not hand the model your whole script, and do not use the full asset allowance

Solution: Use fewer, higher-impact references with clear, structured prompts following the CRAFTS framework

Pitfall 4: Ignoring Duration Limitations

Problem: Trying to fit 30 seconds of detailed action into 15-second generation

Solution: Break complex sequences into multiple generations or simplify action to fit time constraints

Pitfall 5: Under-Specifying Camera Work

Problem: "Camera moves around" without specific direction

Solution: Use precise cinematography terms: "Camera dollies in from wide shot to medium close-up over 5 seconds, maintaining eye-level perspective"

Pitfall 6: Neglecting Audio Integration

Problem: Treating audio as afterthought or only mentioning "add music"

Solution: State what the audio is for, and put the sound inside the shot it belongs to: "(driving rhythm from @Audio 1 throughout). Every cut lands on a beat."

Pitfall 7: Inconsistent Reference Quality

Problem: Mixing high-resolution professional photos with low-quality compressed images

Solution: Maintain consistent quality across all references: don't let one poor-quality reference compromise the generation

Pitfall 8: Assuming Model Inference

Problem: "Make it look good" or "you know what I mean"

Solution: Be explicit about every important detail: the model executes your instructions, it doesn't interpret vague intent

Quick Troubleshooting Guide

Issue: Character appearance changes between generations Solution: Use identical character reference image in each prompt, explicitly state "maintain exact appearance from @Image X"

Issue: Camera movement isn't matching reference Solution: Add more specific description of the camera movement in text, break complex movements into stages

Issue: Style doesn't match reference Solution: Describe the specific style elements in text alongside the reference: "Match @Image 1's color grading: desaturated blues, high contrast, crushed blacks"

Issue: Timing feels off Solution: Resist the urge to add timecodes, which tends to make things worse. Reduce the number of shots you are asking for in one clip, and give the important beat its own shot.

Issue: Audio doesn't match mood Solution: Describe the audio's role explicitly rather than pointing at the file: not "@Audio 1" but "(tense, building suspense from @Audio 1)".

Issue: The voice does not sound like my reference audio Solution: Audio references alone under-deliver on timbre. Describe the voice in words as well as referencing it: "Use the low, warm, slightly grainy middle-aged male voice of @Audio 1 to say...". Keeping the written dialogue close in tone to the reference recording also helps.

Issue: Two identical copies of my character appear in one frame Solution: The "twins" bug. Stop using multi-view character sheets, use single-person reference photos, name each character and pair them with their image ("Ada (@Image 1) hands the file to Ben (@Image 2)"), and add a closing constraint: "Throughout the video, do not generate duplicate characters or a twin effect. Keep only one of each character in frame."

Issue: Subtitles or a watermark appear that I never asked for Solution: Add the constraint lines ("Keep it subtitle-free", "Do not generate a watermark"), strip any text out of your reference assets before uploading, and generate in landscape if you can, since spurious subtitles are markedly less common there than in portrait.

Issue: My animation style drifted back to live action Solution: A realistic reference image will drag the output toward realism unless you say otherwise. Name the style explicitly in the prompt ("2D Japanese animation style"), or convert the reference image into the target style before you generate from it.

Issue: A described effect comes out garbled, like a countdown that scrambles its digits Solution: Stop describing the effect and show it. Feed in a reference video of the effect and point at it: "the way the number appears should follow @Video 1." Motion logic is one of the things the model reads far better from footage than from text.

Issue: The video ends with a click or an abrupt cut-off noise Solution: A known artefact on clips with narration. Regenerate, or fix it in your editor by applying a short volume fade to the end of the audio track. It is a 20-second fix and not worth burning a generation on.

Issue: The image jumps or stretches, and my reference photo looks distorted Solution: Usually an aspect-ratio mismatch. If your reference image's ratio does not match the output ratio you selected, the model has to force a fit. Crop the image to match your target ratio before uploading, or set the output to the adaptive ratio so it follows the reference instead. Reference images also need to sit within a 0.4 to 2.5 aspect ratio and 300 to 6000px per side.

Conclusion

Seedance 2.0 represents a fundamental advancement in AI video generation through its comprehensive multimodal approach. By accepting images, videos, audio, and text as inputs, it provides professionals with unprecedented control over the creative process: moving beyond text-only prompts to true show-and-tell direction.

Seedance 2.0's position in the AI video landscape

The multimodal capability distinguishes Seedance 2.0 from competing platforms. While Kling, Veo, and Sora offer impressive text-to-video capabilities, Seedance's integration of direct video and audio references enables precise reproduction of camera work, motion patterns, and rhythm synchronization that would be difficult or impossible to achieve through text description alone. This positions Seedance as the tool of choice for professionals who need exacting control over visual style, character consistency, and cinematic execution.

The platform continues to evolve with regular capability enhancements and expanded feature support. Mastering the multimodal reference system and CRAFTS prompting framework provides a foundation for increasingly sophisticated video creation as the platform develops.

Key takeaways

Multimodal Control: Seedance 2.0's combination of image, video, audio, and text inputs enables showing the AI exactly what you want rather than attempting to describe it entirely in words. This fundamental approach shift makes previously difficult specifications: exact camera movements, specific choreography, beat-synchronized editing: straightforward to achieve.

Strategic Comparison Advantages: Compared to Kling, Veo, and Sora, Seedance 2.0 offers unique capabilities in audio integration and video reference depth. The direct audio file upload and reference system enables precise mood control and beat synchronization. The video reference capability extends beyond style transfer to full motion and camera replication.

CRAFTS Professional Framework: The six-part CRAFTS methodology (Context, Reference, Action, Framing, Timeline, Style and constraints) covers everything a complete prompt has to specify. The two parts people most often skip are the ones that decide whether the generation works: defining the subject before referencing it, and closing the prompt with the constraints that rule out artefacts.

Sequence your shots, do not time them: the instinct to write second-by-second timecodes is the most common way to make Seedance output worse. Order your shots and let the model find the pacing.

Restraint beats volume: four or five well-chosen assets beat a dozen. Every additional reference makes it harder for the model to work out which features matter, and character turnarounds in particular actively cause identity drift.

Available on Morphic: Professional creators can access the full Seedance 2.0 family through Morphic, including Fast and Mini for cheap iteration and full Seedance 2.0 for 4K finals.

Ready to create? Access Seedance 2.0 on Morphic →