rw-integrate-video

name	rw-integrate-video
description	Help users integrate Runway video generation APIs (text-to-video, image-to-video, video-to-video)
user-invocable	false
allowed-tools	Read, Grep, Glob, Edit, Write

Integrate Video Generation

PREREQUISITE: Run +rw-check-compatibility first. Run +rw-fetch-api-reference to load the latest API reference before integrating. Requires +rw-setup-api-key for API credentials. Requires +rw-integrate-uploads when the user has local files to use as input.

Help users add Runway video generation to their server-side code.

Available Models

Model	Best For	Input	Cost	Speed
seedance2	Reference image and video, long duration	Text, Image, and/or Video	36 credits/sec	Standard
gen4.5	High quality, general purpose	Text and/or Image	12 credits/sec	Standard
gen4_turbo	Fast, image-driven	Image required	5 credits/sec	Fast
gen4_aleph	Video editing/transformation	Video + Text/Image	15 credits/sec	Standard
veo3	Premium Google model	Text/Image	40 credits/sec	Standard
veo3.1	High quality Google model	Text/Image	20-40 credits/sec	Standard
veo3.1_fast	Fast Google model	Text/Image	10-15 credits/sec	Fast

Model selection guidance:

• Default recommendation: gen4.5 — best balance of quality and cost
• Product ads / e-commerce: seedance2 — up to 15s, supports reference image and video
• Budget-conscious: gen4_turbo (requires image) or veo3.1_fast
• Highest quality: veo3 (most expensive)
• Video-to-video editing: gen4_aleph or seedance2

Security

promptImage, promptVideo, videoUri, and references[].uri are fetched server-side by the Runway API — treat them like any outbound fetch:

• Prefer runway:// URIs from +rw-integrate-uploads — scoped to your account, no arbitrary web content.
• If accepting URLs from clients, validate first: require https://, allowlist trusted hosts, reject private addresses. See the Express.js example below.
• Never forward req.body.imageUrl (or similar) straight into promptImage / promptVideo. The SDK snippets below use raw URLs for brevity — they aren't production templates.
• Treat generated outputs as untrusted when piping into downstream automations — ingested media influences the result.

Endpoints

Text-to-Video: POST /v1/text_to_video

Generate video from a text prompt only.

Compatible models: seedance2, gen4.5, veo3, veo3.1, veo3.1_fast

// Node.js SDK
import RunwayML from '@runwayml/sdk';

const client = new RunwayML();

const task = await client.textToVideo.create({
  model: 'gen4.5',
  promptText: 'A golden retriever running through a field of wildflowers at sunset',
  ratio: '1280:720',
  duration: 5
}).waitForTaskOutput();

// task.output is an array of signed URLs
const videoUrl = task.output[0];

# Python SDK
from runwayml import RunwayML

client = RunwayML()

task = client.text_to_video.create(
    model='gen4.5',
    prompt_text='A golden retriever running through a field of wildflowers at sunset',
    ratio='1280:720',
    duration=5
).wait_for_task_output()

video_url = task.output[0]

Image-to-Video: POST /v1/image_to_video

Animate a still image into a video.

Compatible models: seedance2, gen4.5, gen4_turbo, veo3, veo3.1, veo3.1_fast

Recommended: upload via +rw-integrate-uploads and pass the returned runway:// URI.

// Node.js SDK — preferred flow
import fs from 'fs';

const upload = await client.uploads.createEphemeral(
  fs.createReadStream('/path/to/image.jpg')
);

const task = await client.imageToVideo.create({
  model: 'gen4.5',
  promptImage: upload.runwayUri,
  promptText: 'The scene comes to life with gentle wind',
  ratio: '1280:720',
  duration: 5
}).waitForTaskOutput();

External URLs also work — only pass origins you control (see Security):

const task = await client.imageToVideo.create({
  model: 'gen4.5',
  promptImage: 'https://cdn.yourapp.com/landscape.jpg',
  promptText: 'Camera slowly pans right revealing a mountain range',
  ratio: '1280:720',
  duration: 5
}).waitForTaskOutput();

# Python SDK
task = client.image_to_video.create(
    model='gen4.5',
    prompt_image='https://cdn.yourapp.com/landscape.jpg',
    prompt_text='Camera slowly pans right revealing a mountain range',
    ratio='1280:720',
    duration=5
).wait_for_task_output()

Video-to-Video: POST /v1/video_to_video

Transform an existing video with a text prompt and/or reference image.

Compatible models: gen4_aleph, seedance2

// Node.js SDK — gen4_aleph
const task = await client.videoToVideo.create({
  model: 'gen4_aleph',
  videoUri: 'https://cdn.yourapp.com/source.mp4',
  promptText: 'Transform into an animated cartoon style',
}).waitForTaskOutput();

// Node.js SDK — seedance2 video-to-video (with optional image reference)
const task = await client.videoToVideo.create({
  model: 'seedance2',
  promptVideo: 'https://cdn.yourapp.com/input.mp4',
  promptText: 'Transform into a warm golden sunset scene',
  references: [{ type: 'image', uri: 'https://cdn.yourapp.com/style_ref.jpg' }]
}).waitForTaskOutput();

seedance2 VTV input requirements: max 15 seconds, max 32 MB, min 720p resolution, MP4 recommended.

Seedance 2

Seedance 2 supports text-to-video, image-to-video (two modes), and video-to-video. It uses pixel-based ratios: 1280:720, 720:1280, 960:960, 1112:834, 834:1112, 1470:630, 992:432, 864:496, 752:560, 640:640, 560:752, 496:864.

Text-to-Video

const task = await client.textToVideo.create({
  model: 'seedance2',
  promptText: 'A calm ocean wave gently crashing on a sandy beach at sunset',
  duration: 5,
  ratio: '1280:720'
}).waitForTaskOutput();

Image-to-Video — Mode 1: First / Last Frame

Use a specific image as the first and/or last frame. The references field cannot be used in this mode.

const task = await client.imageToVideo.create({
  model: 'seedance2',
  promptText: 'Smooth transition from day to night in a cozy mountain cabin',
  promptImage: [
    { uri: 'https://cdn.yourapp.com/image.jpg', position: 'first' },
    { uri: 'https://cdn.yourapp.com/image2.jpg', position: 'last' }
  ],
  duration: 4,
  ratio: '1280:720'
}).waitForTaskOutput();

promptImage is an array of objects with uri (required) and position ("first" or "last", defaults to first).

Image-to-Video — Mode 2: Image Reference

Use an image as a stylistic/content reference rather than a literal frame. promptImage is still required (as a URI string or single-item array).

const task = await client.imageToVideo.create({
  model: 'seedance2',
  promptText: 'Smooth transition from day to night in a cozy mountain cabin',
  promptImage: 'https://cdn.yourapp.com/image.jpg',
  references: [{ type: 'image', uri: 'https://cdn.yourapp.com/reference.jpg' }],
  duration: 4,
  ratio: '1280:720'
}).waitForTaskOutput();

These two ITV modes are mutually exclusive — you cannot use position in promptImage and references in the same request.

Video-to-Video

Transform an existing video guided by a text prompt, optionally with an image reference.

task = client.video_to_video.create(
    model='seedance2',
    prompt_video='https://cdn.yourapp.com/input.mp4',
    prompt_text='Transform into a warm golden sunset scene',
    references=[{'type': 'image', 'uri': 'https://cdn.yourapp.com/style_ref.jpg'}]
).wait_for_task_output()

VTV input requirements: max 15 seconds, max 32 MB, min 720p resolution, MP4 recommended.

Seedance 2 Parameters

Parameter	Type	Required	Description
model	string	Yes	Must be "seedance2"
promptText	string	Yes	Text description of the desired video
duration	number	Yes (TTV/ITV)	Duration in seconds
ratio	string	Yes (TTV/ITV)	1280:720, 720:1280, 960:960, 1112:834, 834:1112, 1470:630
promptImage	string or array	Yes (ITV)	URI string or array of { uri, position? } objects
promptVideo	string	Yes (seedance2 VTV)	Input video URI (seedance2 only)
videoUri	string	Yes (gen4_aleph VTV)	Input video URI (gen4_aleph only)
references	array	No	Image references — [{ type: "image", uri: "..." }] (ITV Mode 2 and VTV only)

Character Performance: POST /v1/character_performance

Animate a character with facial/body performance.

Compatible models: act_two

const task = await client.characterPerformance.create({
  model: 'act_two',
  promptImage: 'https://cdn.yourapp.com/character.jpg',
  promptPerformance: 'https://cdn.yourapp.com/performance.mp4',
  ratio: '1280:720',
  duration: 5
}).waitForTaskOutput();

Common Parameters

Parameter	Type	Description
model	string	Model ID (required)
promptText	string	Text prompt describing the video
promptImage	string	URL, data URI, or runway:// URI of input image
ratio	string	Aspect ratio, e.g. '1280:720', '720:1280'
duration	number	Video length in seconds (2-15, model-dependent)

Integration Pattern

When helping the user integrate, follow this pattern:

1. Determine the use case — What type of video generation? (text-to-video, image-to-video, etc.)
2. Prefer uploads over URLs — Default to +rw-integrate-uploads so inputs are runway:// URIs. External URLs only from origins you control (see Security).
3. Select the model — Recommend based on quality/cost/speed needs
4. Write the server-side handler — Create an API route or server function
5. Handle the output — Download and store the video, don't serve signed URLs to clients
6. Add error handling — Wrap in try/catch, handle TaskFailedError

Example: Express.js API Route

import RunwayML from '@runwayml/sdk';
import express from 'express';

const client = new RunwayML();
const app = express();
app.use(express.json());

// `runway://` URIs bypass this check; external URLs must match the allowlist.
const ALLOWED_MEDIA_HOSTS = new Set(['cdn.yourapp.com', 'uploads.yourapp.com']);

function assertTrustedMediaUrl(raw) {
  const u = new URL(raw);
  if (u.protocol !== 'https:') throw new Error('https required');
  if (!ALLOWED_MEDIA_HOSTS.has(u.hostname)) throw new Error('untrusted media host');
  return u.toString();
}

app.post('/api/generate-video', async (req, res) => {
  try {
    const { prompt, imageUrl, model = 'gen4.5', duration = 5 } = req.body;

    const params = {
      model,
      promptText: prompt,
      ratio: '1280:720',
      duration
    };

    let task;
    if (imageUrl) {
      task = await client.imageToVideo.create({
        ...params,
        promptImage: assertTrustedMediaUrl(imageUrl)
      }).waitForTaskOutput();
    } else {
      task = await client.textToVideo.create(params).waitForTaskOutput();
    }

    res.json({ videoUrl: task.output[0] });
  } catch (error) {
    console.error('Video generation failed:', error);
    res.status(400).json({ error: error.message });
  }
});

For browser uploads: POST files to your server, upload via +rw-integrate-uploads, and pass the runway:// URI. Don't accept raw URLs from the browser.

Example: Next.js API Route

// app/api/generate-video/route.ts
import RunwayML from '@runwayml/sdk';
import { NextRequest, NextResponse } from 'next/server';

const client = new RunwayML();

export async function POST(request: NextRequest) {
  const { prompt, imageUrl } = await request.json();

  try {
    const task = imageUrl
      ? await client.imageToVideo.create({
          model: 'gen4.5',
          promptImage: imageUrl,
          promptText: prompt,
          ratio: '1280:720',
          duration: 5
        }).waitForTaskOutput()
      : await client.textToVideo.create({
          model: 'gen4.5',
          promptText: prompt,
          ratio: '1280:720',
          duration: 5
        }).waitForTaskOutput();

    return NextResponse.json({ videoUrl: task.output[0] });
  } catch (error) {
    return NextResponse.json(
      { error: error instanceof Error ? error.message : 'Generation failed' },
      { status: 500 }
    );
  }
}

Example: FastAPI Route

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from runwayml import RunwayML

app = FastAPI()
client = RunwayML()

class VideoRequest(BaseModel):
    prompt: str
    image_url: str | None = None
    model: str = "gen4.5"
    duration: int = 5

@app.post("/api/generate-video")
async def generate_video(req: VideoRequest):
    try:
        if req.image_url:
            task = client.image_to_video.create(
                model=req.model,
                prompt_image=req.image_url,
                prompt_text=req.prompt,
                ratio="1280:720",
                duration=req.duration
            ).wait_for_task_output()
        else:
            task = client.text_to_video.create(
                model=req.model,
                prompt_text=req.prompt,
                ratio="1280:720",
                duration=req.duration
            ).wait_for_task_output()

        return {"video_url": task.output[0]}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

Tips

• Output URLs expire in 24-48 hours. Download videos to your own storage (S3, GCS, local filesystem) immediately after generation.
• gen4_turbo requires an image — it cannot do text-only generation.
• Video-to-video models: gen4_aleph and seedance2 — use for editing/transforming existing videos.
• Duration varies by model. Most models support 2-10 seconds; seedance2 supports up to 15 seconds.
• waitForTaskOutput() has a default 10-minute timeout. For long-running generations, you may want to implement your own polling loop or increase the timeout.
• For local files, always use +rw-integrate-uploads to upload first, then pass the runway:// URI.