Seedance 2.5

Introduction

Regenerating a shot to change one thing throws away everything else. When a clip is already right and only the light is wrong, or one object has to go, or the whole look needs to shift, you want to keep the frame and edit it in place.

Seedance 2.5 edits an existing clip through inputs.video. You pass the source, and the prompt says what to change and what to hold. The comparison below kept every pixel of the performance and the camera, and rewrote only the lighting, turning an overcast afternoon into a warm dusk.

The woman, her movement, the rain on the glass, and the locked camera are untouched. Only the light changed. This guide covers the edit request, how the prompt drives it, and the jobs it does: restyling, removing and replacing, swapping backgrounds, and editing a single stretch of time.

The request

An edit is a videoInference call with a source clip in inputs.video and edit instructions in the prompt.

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: {
    video: 'https://example.com/source.mp4'
  },
  positivePrompt: 'Edit @Video1: remove the scooter parked at the curb and fill the space with the cobblestone street behind it. Keep the storefront, the lighting, and the camera move exactly unchanged.',
  resolution: '720p',
  duration: 'auto',
  outputFormat: 'MOV',
  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": {
                "video": "https://example.com/source.mp4"
            },
            "positivePrompt": "Edit @Video1: remove the scooter parked at the curb and fill the space with the cobblestone street behind it. Keep the storefront, the lighting, and the camera move exactly unchanged.",
            "resolution": "720p",
            "duration": "auto",
            "outputFormat": "MOV",
            "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": {
        "video": "https://example.com/source.mp4"
      },
      "positivePrompt": "Edit @Video1: remove the scooter parked at the curb and fill the space with the cobblestone street behind it. Keep the storefront, the lighting, and the camera move exactly unchanged.",
      "resolution": "720p",
      "duration": "auto",
      "outputFormat": "MOV",
      "deliveryMethod": "async"
    }
  ]'

runware run bytedance:seedance@2.5 \
  inputs.video=https://example.com/source.mp4 \
  positivePrompt="Edit @Video1: remove the scooter parked at the curb and fill the space with the cobblestone street behind it. Keep the storefront, the lighting, and the camera move exactly unchanged." \
  resolution=720p \
  duration=auto \
  outputFormat=MOV \
  deliveryMethod=async

{
  "taskType": "videoInference",
  "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "model": "bytedance:seedance@2.5",
  "inputs": {
    "video": "https://example.com/source.mp4"
  },
  "positivePrompt": "Edit @Video1: remove the scooter parked at the curb and fill the space with the cobblestone street behind it. Keep the storefront, the lighting, and the camera move exactly unchanged.",
  "resolution": "720p",
  "duration": "auto",
  "outputFormat": "MOV",
  "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"
  }
]

Four things shape an edit request:

• inputs.video is the single source clip to edit, a URL or a UUID. The source addresses as @Video1 in the prompt, and extra inputs.referenceImages or inputs.referenceAudios can guide the change.
• duration on an edit is "auto", which holds the source's length. An integer duration is rejected here.
• The output size comes from the source. Aspect is inherited, so width and height are not accepted with inputs.video. Use resolution to pick the 480p or 720p tier.
• Set outputFormat to MOV for the cleanest audiovisual continuity, especially if the edit will be chained into further edits or an extension.

How the prompt drives an edit

Seedance reads an edit from trigger words: edit, remove, add, replace, change to, restyle, relight. Leading with one tells the model to keep the source and modify it, rather than regenerate the shot around its motion the way reference videos do.

Two habits make edits land:

• Name what stays, not just what changes. "Keep the storefront, the lighting, and the camera move unchanged" fences the edit. Without it, the model tends to touch more than you asked.
• Describe the change from A to B. "Relight from a cool overcast afternoon to a warm amber dusk" gives the model both ends. A bare "make it warmer" leaves the target open.

Editing operations

Every edit keeps the source and changes one thing. These are the jobs Seedance handles, from a whole-frame restyle down to a single stretch of time.

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: {
    video: 'https://example.com/source.mp4'
  },
  positivePrompt: 'Edit @Video1: between 0:02 and 0:04, add two quick forks of lightning that flash across the storm clouds and briefly light the whole scene, then return to the steady dim stormlight. Keep the lighthouse, the cliff, the waves, the camera move, and everything before and after that window exactly unchanged.',
  resolution: '720p',
  duration: 'auto',
  outputFormat: 'MOV',
  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": {
                "video": "https://example.com/source.mp4"
            },
            "positivePrompt": "Edit @Video1: between 0:02 and 0:04, add two quick forks of lightning that flash across the storm clouds and briefly light the whole scene, then return to the steady dim stormlight. Keep the lighthouse, the cliff, the waves, the camera move, and everything before and after that window exactly unchanged.",
            "resolution": "720p",
            "duration": "auto",
            "outputFormat": "MOV",
            "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": {
        "video": "https://example.com/source.mp4"
      },
      "positivePrompt": "Edit @Video1: between 0:02 and 0:04, add two quick forks of lightning that flash across the storm clouds and briefly light the whole scene, then return to the steady dim stormlight. Keep the lighthouse, the cliff, the waves, the camera move, and everything before and after that window exactly unchanged.",
      "resolution": "720p",
      "duration": "auto",
      "outputFormat": "MOV",
      "deliveryMethod": "async"
    }
  ]'

runware run bytedance:seedance@2.5 \
  inputs.video=https://example.com/source.mp4 \
  positivePrompt="Edit @Video1: between 0:02 and 0:04, add two quick forks of lightning that flash across the storm clouds and briefly light the whole scene, then return to the steady dim stormlight. Keep the lighthouse, the cliff, the waves, the camera move, and everything before and after that window exactly unchanged." \
  resolution=720p \
  duration=auto \
  outputFormat=MOV \
  deliveryMethod=async

{
  "taskType": "videoInference",
  "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "model": "bytedance:seedance@2.5",
  "inputs": {
    "video": "https://example.com/source.mp4"
  },
  "positivePrompt": "Edit @Video1: between 0:02 and 0:04, add two quick forks of lightning that flash across the storm clouds and briefly light the whole scene, then return to the steady dim stormlight. Keep the lighthouse, the cliff, the waves, the camera move, and everything before and after that window exactly unchanged.",
  "resolution": "720p",
  "duration": "auto",
  "outputFormat": "MOV",
  "deliveryMethod": "async"
}

Tips

1. Lead with an edit verb. remove, replace, relight, restyle, change to. The trigger word tells Seedance to keep the source and modify it rather than regenerate the shot.

2. Keep edits on duration: "auto". An integer duration is an extension instruction and is rejected on an edit. Auto holds the source's length.

3. Fence the edit by naming what stays. "Keep the camera, the lighting, and everyone else unchanged" stops the model from drifting into parts you meant to leave alone.

4. Describe the change end to end. "From a cool overcast afternoon to a warm amber dusk" beats "make it warmer". Give the model the target, not just the direction.

5. Feed a clean source. Edits inherit the source's quality, and a well-lit, stable plate separates and fills more cleanly than a noisy or shaky one, especially for removals and background swaps.

6. Reach for a reference for a specific target. A precise replacement, a particular product, wardrobe, or face, rides in through inputs.referenceImages and is named with @Image1, rather than described in words alone.

7. Deliver edits as MOV. It holds audiovisual continuity best, which matters most when an edit feeds another edit or an extension.