How to Replace a Video Character Completely with Seedance 2.5 (2026)

Complete step-by-step walkthrough

Original TECH4SSD tutorial and reference media

Tools

• Seedance 2.5
• Luma
• An AI assistant for optional reference-video analysis

Before you begin

• A source/reference video you are allowed to use
• One clear character image or, preferably, a multi-angle character sheet
• Access to Seedance 2.5 through Luma or another supported AI video interface

Watch the original Instagram tutorial

1. 1. Prepare the source video

Choose the exact video you want to recreate and trim it to the duration you actually want to generate. The source video should be treated as the master for movement, choreography, timing, camera, framing, perspective, environment, and scene geometry, not as the authority for the target actor’s appearance. If the source contains several people, decide which one performer is the TARGET_ACTOR before generation.

• Final reference clip is trimmed
• Target performer is identifiable
• Source motion and camera are clear

2. 2. Prepare the replacement character reference

Upload either a single clear image or a character sheet. A character sheet is preferable when the source includes rear, profile, wide, or full-body views. The replacement reference should control the complete new character: face, hair, skin tone, body appearance, clothing, footwear, accessories, and style. Do not ask the model to preserve the source actor’s wardrobe if you want a true full-character replacement.

Tip: For difficult turns, use a sheet containing face close-up, full-body front, side, three-quarter, and back views.

3. 3. Open Seedance 2.5 in Luma

Open Luma and choose Seedance 2.5 for the generation. Luma’s current Seedance 2.5 guidance supports explicit roles for image and video references, so keep the mapping simple: @video1 controls performance/camera/environment; @image1 controls the complete replacement character. Upload both assets before pasting the prompt.

Luma, Seedance 2.5 Complete Guide

4. 4. Optional: analyze complex source footage first

If the scene is complex, give the exact edited source video to a multimodal AI assistant and ask it to identify the lead performer, shot changes, movement path, turns, occlusions, camera behavior, and important action beats. Use that analysis to clarify TARGET_ACTOR in the universal prompt. Do not describe a different or longer cut than the one you actually upload.

Exact prompt

Analyze this exact reference video for a full-character replacement workflow. Identify the ONE primary performer to replace and track that same role across the entire timeline. Describe their starting position, movement path, turns, occlusions, interactions, final position, and distinctive actions. Also describe meaningful camera/framing changes. Focus on tracking the same performer across front, side, rear, wide, and close views. Do not redesign the scene.

5. 5. Paste the universal full-character-replacement prompt

Use the complete prompt below. Its central rule is strict reference separation: @video1 supplies motion, timing, camera and environment; @image1 supplies face, hair, body, clothing, shoes, accessories and style. This prevents the common failure where only the face changes while the source actor’s clothing survives.

Exact prompt

UNIVERSAL FULL CHARACTER REPLACEMENT — IDENTITY + OUTFIT TRANSFER

INPUTS:
Source Video = @video1
Replacement Character Reference = @image1

GOAL:
Replace ONE specific main performer in @video1 with the complete character shown in @image1.

This is a FULL CHARACTER REPLACEMENT.

Do NOT perform a face swap only.
Do NOT preserve the original actor's clothing.
Do NOT blend the source actor with the replacement character.

REFERENCE RESPONSIBILITIES:

@video1 provides ONLY:
- movement
- choreography
- body poses
- timing
- walking path
- gestures
- interactions
- camera movement
- framing
- perspective
- scene geometry
- environment
- lighting reference

@image1 provides:
- facial identity
- hairstyle
- hair color
- skin tone
- body appearance
- body proportions
- clothing
- footwear
- accessories
- styling
- overall character design

If @image1 is a character sheet, use ALL available views to reconstruct the exact same character consistently from every camera angle.

If @image1 is a single image, infer the unseen side, back and full-body appearance consistently from that image without borrowing identity or wardrobe from the original performer.

STEP 1 — IDENTIFY THE TARGET ACTOR:

Analyze @video1 from beginning to end before generating.

Identify the ONE main performer who carries the primary action.

Track this same person temporally using:
- starting position
- movement path
- body position
- interaction with surrounding people
- role in the choreography
- turns
- gestures
- final position
- distinctive actions

Lock this performer as TARGET_ACTOR.

TARGET_ACTOR must remain the same role across:
- wide shots
- close-ups
- side views
- rear views
- camera changes
- occlusions
- crowd crossings
- turns
- different framing

Do not accidentally switch to another person after a cut or perspective change.

STEP 2 — REMOVE THE ORIGINAL CHARACTER APPEARANCE:

Use the original TARGET_ACTOR only as a motion skeleton and spatial guide.

Do NOT preserve:
- original face
- original hairstyle
- original skin
- original body appearance
- original clothing
- original shoes
- original accessories
- original character styling

The source actor's visual identity must disappear entirely.

Only their motion, pose, timing, position and performance remain.

STEP 3 — REBUILD TARGET_ACTOR FROM @image1:

Replace TARGET_ACTOR with the COMPLETE character shown in @image1.

Preserve from @image1:
- exact facial structure
- eyes
- eyebrows
- nose
- lips
- jawline
- ears
- skin tone
- hairstyle
- hairline
- body build
- visible proportions
- clothing design
- clothing colors
- footwear
- jewelry
- accessories
- hats
- glasses
- masks
- jackets
- shirts
- trousers
- dresses
- uniforms
- distinctive character details

Treat @image1 as the absolute visual authority for the replacement character.

If the character sheet contains multiple angles, use them together as ONE identity.

Do not mix clothing or facial features from @video1 into the replacement character.

STEP 4 — MOTION RETARGETING:

Transfer TARGET_ACTOR's exact performance from @video1 onto the replacement character.

Preserve:
- walking direction
- speed
- stride length
- posture
- turns
- head movement
- shoulder movement
- arm movement
- hand placement
- gestures
- leg movement
- body orientation
- sitting
- standing
- running
- dancing
- bending
- reacting
- interactions
- pauses
- choreography
- action timing

The replacement character must perform the same movement at the same moment.

Do NOT change the choreography to match the reference image pose.

Instead, animate the @image1 character naturally through the poses required by @video1.

FRONT / SIDE / BACK CONSISTENCY:

When TARGET_ACTOR faces away from camera:
generate the correct rear view of the @image1 character and outfit.

When TARGET_ACTOR is in profile:
generate the correct side view of @image1.

When TARGET_ACTOR turns toward camera:
reveal the same @image1 identity.

The same clothing, body, hair and accessories must remain consistent throughout the turn.

There must be no moment where the original actor appears before or after the turn.

CLOTHING — IMPORTANT:

The clothing from @image1 MUST replace the source actor's clothing.

Do NOT retain the source costume.

If @image1 contains:
- a jacket, preserve that jacket
- a hoodie, preserve that hoodie
- a suit, preserve that suit
- a school uniform, preserve that uniform
- jewelry, preserve that jewelry
- glasses, preserve those glasses
- hat, preserve the hat
- shoes, preserve the shoes

Adapt the @image1 outfit naturally to:
- walking
- running
- turning
- sitting
- dancing
- bending
- wind
- gravity
- camera perspective

The wardrobe must remain visually identical throughout all shots.

No wardrobe drift.
No colors changing.
No source-video clothes returning.

BODY ADAPTATION:

The body from @image1 should replace the source actor's visual body appearance.

Retarget the choreography onto the replacement body's proportions while preserving the original motion path.

If the replacement character is:
- taller
- shorter
- slimmer
- broader
- more muscular
- wearing different clothing

adapt the movement naturally while keeping the same spatial action and timing.

Do not force the source actor's exact body shape onto @image1.

OTHER PEOPLE:

Do NOT replace anyone except TARGET_ACTOR unless explicitly instructed.

All other people in @video1 should keep:
- their original identity
- original clothing
- original position
- original choreography
- original timing

Never clone @image1 into the crowd.

Never apply @image1's face or clothes to surrounding people.

TARGET_ACTOR = @image1
EVERYONE ELSE = source-video characters

CAMERA:

Use @video1 as a strict camera reference.

Preserve:
- exact shot sequence
- shot duration
- camera position
- camera height
- camera angle
- camera path
- tracking
- pull-back
- push-in
- zoom behavior
- perspective
- focal feel
- framing
- composition
- subject scale
- depth relationships

Do not invent new camera moves.

Do not reframe the replacement character differently from the source performance.

The new character must occupy the same spatial role in the frame.

ENVIRONMENT:

Preserve @video1's environment.

Maintain:
- location
- architecture
- furniture
- vehicles
- walls
- ground
- signs
- props
- crowd placement
- foreground
- background
- spatial layout

Do not redesign the environment unless explicitly requested.

Only regenerate pixels necessary to integrate the replacement character naturally.

LIGHTING:

Relight the @image1 character to match @video1.

Preserve the replacement character's identity and outfit while adapting them to:
- source light direction
- exposure
- shadow direction
- color temperature
- highlights
- reflections
- ambient light
- depth of field
- motion blur
- atmosphere

The person must look physically present in the original environment.

Do not make the character look pasted on top.

IDENTITY LOCK:

Once @image1 replaces TARGET_ACTOR, lock this identity for the full timeline.

Maintain the exact same:
- face
- skin
- hair
- body appearance
- clothing
- shoes
- accessories
- age
- styling

through every frame.

Never interpolate between the source actor and @image1.

Never temporarily restore the source actor.

Never use the source actor's appearance to fill missing angles.

ANATOMY:

Maintain realistic human anatomy.

Correct:
- face
- neck
- shoulders
- arms
- elbows
- wrists
- hands
- five fingers
- torso
- hips
- knees
- ankles
- feet

No:
- extra limbs
- missing limbs
- warped hands
- duplicated fingers
- stretched torso
- broken neck
- facial melting
- body morphing

TEMPORAL CONSISTENCY:

The replacement must remain stable across the entire video.

No frame-to-frame identity drift.

No:
- changing face
- changing hairstyle
- changing clothing
- changing skin tone
- changing body type
- disappearing accessories
- changing shoes
- source actor returning

Treat the full video as one continuous performance by the @image1 character.

STRICT NEGATIVE INSTRUCTIONS:

NO face-only replacement.
NO source actor clothing.
NO source actor hairstyle.
NO source actor body appearance.
NO identity blending.
NO face morph between both people.
NO alternating characters.
NO partial replacement.
NO original actor appearing during wide shots.
NO original actor returning during rear views.
NO original actor returning during motion blur.
NO original actor returning after camera cuts.
NO @image1 clones.
NO replacement of background people.
NO wardrobe blending.
NO source wardrobe returning.
NO wardrobe color drift.
NO hairstyle drift.
NO body drift.
NO new actions.
NO missing actions.
NO timing changes.
NO camera changes.
NO unnecessary environment changes.

FINAL RESULT:

The final video should appear as if the complete character from @image1 was physically present during the original filming of @video1.

Keep:
- the same source performance
- same movement
- same timing
- same choreography
- same interactions
- same camera
- same scene
- same environment

Change:
- the entire TARGET_ACTOR visual appearance

The final TARGET_ACTOR must be 100% derived from @image1:
FACE + HAIR + BODY + CLOTHING + SHOES + ACCESSORIES + STYLE.

Use @video1 ONLY for motion, position, timing, camera and environment.

PRIORITY ORDER:

1. Replace TARGET_ACTOR completely with @image1.
2. Preserve @image1 clothing and full appearance.
3. Preserve exact source motion and timing.
4. Never show the original TARGET_ACTOR.
5. Never apply @image1 to other people.
6. Preserve camera and framing.
7. Preserve environment.
8. Maintain photorealistic temporal consistency.

6. 6. Match duration and generate

Set the generation duration to match the edited reference clip as closely as the interface allows, then generate. Seedance 2.5 supports clips up to 30 seconds in a single generation, though available controls can vary by interface. Do not ask the model to invent extra action after the reference ends.

• @video1 uploaded
• @image1 uploaded
• Reference roles are explicit
• Duration matches source
• Universal prompt pasted

7. 7. Diagnose only the part that failed

If the original actor’s clothes remain, reinforce that @image1 is the absolute authority for wardrobe and that @video1 supplies motion only. If the original face returns during turns, strengthen temporal identity lock and use a multi-angle character sheet. If other people are replaced, reinforce that TARGET_ACTOR is the only person mapped to @image1. Avoid rewriting everything when only one failure needs correction.

Tip: Luma’s own Seedance guidance recommends giving each reference a clear role and fixing the specific failure rather than endlessly expanding the prompt.

Try Seedance 2.5 on Luma