Seedance 2.5

Introduction

Text-to-video invents the motion on every run. That is what you want for a first shot, and the wrong tool when a performance already exists and has to reappear on a new subject. The exact dance you captured, now danced by an animated character. The one clip you shot, now delivered in six languages. A grey pre-vis turntable your director already signed off, now rendered in real materials.

Seedance 2.5 handles those with inputs.referenceVideos. A reference video hands over motion, camera path, pacing, and lip-sync timing, a reference image hands over identity or material, and the prompt names what changed. The example below fuses two inputs, a video of a live-action dance and a still of a 3D fox character, into one clip where the fox performs the dance move for move.

A contemporary dancer in charcoal practice wear performs a short flowing full-body movement on a bare pale-wood stage in a bright studio, wide 16:9 locked camera showing the full body, soft daylight from the left. The dancer reaches one arm overhead, sweeps it down while turning a half-turn, then folds into a low crouch with the head bowed.

A stylized 3D-animated anthropomorphic fox character standing upright in a neutral pose, full body, warm orange fur with a cream chest, an expressive friendly face, wearing a forest-green scarf, on a plain soft neutral-grey studio backdrop, animated-film render.

This guide covers what a reference video actually carries, then the three jobs it does best: transferring motion across subjects, recasting a performance, and turning a clay pre-vis into a finished shot.

Request shape

A call takes inputs.referenceVideos for the motion and inputs.referenceImages for the identity, addressed in the prompt as @Video1 and @Image1.

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'bytedance:seedance@2.5',
  inputs: {
    referenceVideos: [
      'https://example.com/dance.mp4'
    ],
    referenceImages: [
      'https://example.com/character.jpg'
    ]
  },
  positivePrompt: 'Transfer the exact movement and timing from @Video1 onto the character in @Image1. It performs the identical full-body sequence beat for beat, at the same pace, on the same stage. Preserve the character\'s design and proportions throughout.',
  width: 1280,
  height: 720,
  duration: 6,
  deliveryMethod: 'async'
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "bytedance:seedance@2.5",
            "inputs": {
                "referenceVideos": [
                    "https://example.com/dance.mp4"
                ],
                "referenceImages": [
                    "https://example.com/character.jpg"
                ]
            },
            "positivePrompt": "Transfer the exact movement and timing from @Video1 onto the character in @Image1. It performs the identical full-body sequence beat for beat, at the same pace, on the same stage. Preserve the character's design and proportions throughout.",
            "width": 1280,
            "height": 720,
            "duration": 6,
            "deliveryMethod": "async"
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "videoInference",
      "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
      "model": "bytedance:seedance@2.5",
      "inputs": {
        "referenceVideos": [
          "https://example.com/dance.mp4"
        ],
        "referenceImages": [
          "https://example.com/character.jpg"
        ]
      },
      "positivePrompt": "Transfer the exact movement and timing from @Video1 onto the character in @Image1. It performs the identical full-body sequence beat for beat, at the same pace, on the same stage. Preserve the character's design and proportions throughout.",
      "width": 1280,
      "height": 720,
      "duration": 6,
      "deliveryMethod": "async"
    }
  ]'

runware run bytedance:seedance@2.5 \
  inputs.referenceVideos.0=https://example.com/dance.mp4 \
  inputs.referenceImages.0=https://example.com/character.jpg \
  positivePrompt="Transfer the exact movement and timing from @Video1 onto the character in @Image1. It performs the identical full-body sequence beat for beat, at the same pace, on the same stage. Preserve the character's design and proportions throughout." \
  width=1280 \
  height=720 \
  duration=6 \
  deliveryMethod=async

{
  "taskType": "videoInference",
  "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "model": "bytedance:seedance@2.5",
  "inputs": {
    "referenceVideos": [
      "https://example.com/dance.mp4"
    ],
    "referenceImages": [
      "https://example.com/character.jpg"
    ]
  },
  "positivePrompt": "Transfer the exact movement and timing from @Video1 onto the character in @Image1. It performs the identical full-body sequence beat for beat, at the same pace, on the same stage. Preserve the character's design and proportions throughout.",
  "width": 1280,
  "height": 720,
  "duration": 6,
  "deliveryMethod": "async"
}

[
  {
    "taskType": "videoInference",
    "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
    "videoUUID": "9c1b2d3a-4e5f-6789-abcd-ef0123456789",
    "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/9c1b2d3a-4e5f-6789-abcd-ef0123456789.mp4"
  }
]

The fields that matter here:

• inputs.referenceVideos takes up to 10 videos, each a URL or a UUID, with the combined duration under 30 seconds. One video is the common case. Several let the model blend cues, like a follow move from one and an orbit from another.
• inputs.referenceImages supplies identity, up to 30. One image covers a straight character or product swap.
• Match the output size to the reference video. A landscape source paired with a portrait output pulls the model between two structures. Set width and height (or resolution) to the source's shape.

What a reference video carries

A reference video hands five things to the output:

• Camera movement: the exact orbit, dolly, pan, or handheld drift arrives on the render.
• Pacing: how long each beat holds and how fast the subject moves between them.
• Framing: a head-and-shoulders shot stays head-and-shoulders.
• Motion trajectory: an arm lift lands at the same moment, a turn reaches the same angle on the same beat.
• Lip-sync duration: every phrase occupies the same span the source phrase did.

That last one is the load-bearing signal for recasting and localization. It is what stops a dubbed line from finishing before the mouth closes, or a gesture timed to English syllables from cutting a Japanese line short.

Transferring motion across subjects

The clearest case is a performance that has to move to a completely different body. A motion-capture shoot is expensive. A reference video is not. In the hero above, the live-action dance drives the motion while the fox reference supplies the body.

The fox lands every beat of the dance at the same moment the dancer did, with the weight and follow-through carried across. The transfer works because the two share a body plan: an upright figure with two arms maps cleanly onto another upright figure. Push it further, onto a very different anatomy, and the model has to invent how the movement would translate, so keep the target's structure close to the source's when the motion has to read exactly.

Recasting and localizing at scale

The pattern that only inputs.referenceVideos delivers is one performance, many faces and languages. The reference video holds the gestures and the exact lip-sync timing, a reference image swaps the face, and the prompt swaps the language. Below, a single English recipe-tip clip becomes three regional creators, each in a different language, off the same source.

Every variant lifts the herb board on the same beat and holds each phrase for the same span, so three different languages all fit the timing the source left for them. Each output is one call with a new reference image and a new language line, which is what turns a single shoot into a market-by-market rollout: five markets and twenty creator cuts share one motion source, and each variant is a fresh request. The multilingual guide covers the language and lip-sync side in depth.

Clay pre-vis to finished shot

A 3D commercial often locks the camera and blocking before the materials exist. The director signs off a grey clay turntable, and the CG team spends days on the material pass. Feeding the clay render to Seedance as the reference video collapses that gap: the clay carries the motion and pacing, a packshot carries the material, and the output is the same move in production quality.

A grey untextured 3D clay pre-visualisation render of a mid-century lounge armchair on a matte grey turntable in a plain grey stage, all flat matte grey clay with no colour or texture. A slow smooth 360-degree orbital camera path around the armchair at a three-quarter angle, even soft key light from above.

A high-end furniture packshot of a mid-century lounge armchair with warm tan leather upholstery, softly tufted cushions, tapered walnut legs, and a low curved backrest, three-quarter angle on a warm off-white studio backdrop, soft warm directional light.

The rendered chair follows the clay frame for frame: the same arc at the same speed, the same three-quarter angle reached on the same beat. Every material property comes from the packshot, the leather grain and the walnut all lifted from the still. The pattern generalises past turntables to any pre-vis that carries an approved camera move, an architectural walkthrough or a car reveal locked to a specific arc, so the material pass moves out of the render queue and into a prompt.

Tips

1. Match the output size to the source. A reference video sets the motion structure, so a landscape source with a portrait output fights itself. Keep the aspect and size of the source unless there is a reason to differ.

2. Send a reference image whenever identity matters. A reference video alone preserves motion but reinvents the face, product, and wardrobe each run. Character swaps, product locks, and material renders all need at least one reference image.

3. Keep the target's body plan close to the source's. Motion maps cleanly between similar structures. The further the target anatomy is from the source, the more the model has to invent, so an upright character reads a human performance better than a very different body would.

4. Pin the pacing explicitly. Adding "preserve every gesture and the exact lip-sync duration of each phrase" reinforces what the reference video already carries, and heads off the subtle timing drift a dubbed line can pick up.

5. Feed a clean, well-lit source. Motion blur, low light, and shake all degrade the motion signal, and the render inherits the noise. Recasting is the most sensitive, since the model reads small mouth movements to hold lip-sync.

6. Keep localization sources short and gesture-driven. A five-to-eight-second clip with a clean opening gesture, a delivered line, and a closing beat relocalizes cleanly across a dozen variants. Long dialogue-heavy sources have more phonemes to re-match.

7. Reach for inputs.video when the goal is to keep the source pixels. Reference videos regenerate the shot around the motion. To swap one element while everything else stays put, that is an in-place edit, covered in the editing guide.