Video generation

Video models are task-based. A request returns a task id immediately; the render happens upstream and you poll until it finishes. No long-lived HTTP connection is held open.

Submit a task

curl https://router.triw.io/v1/videos \
  -H "Authorization: Bearer $TRIW_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kling-v2",
    "prompt": "A slow dolly shot through a rain-soaked alley at night, neon reflections",
    "duration": 5,
    "aspect_ratio": "16:9"
  }'

The response is a task object:

{
  "id": "task_8f3c1a...",
  "status": "queued"
}

Keep id. It is the only handle for the job.

Poll for the result

curl https://router.triw.io/v1/videos/task_8f3c1a... \
  -H "Authorization: Bearer $TRIW_API_KEY"

Poll every 5–15 seconds. Treat these status values as terminal: a finished state that carries the asset, or a failure state that carries a reason. Anything else means keep polling.

import time, requests

BASE = "https://router.triw.io"
H = {"Authorization": f"Bearer {API_KEY}"}

task = requests.post(f"{BASE}/v1/videos", headers=H, json={
    "model": "kling-v2",
    "prompt": "A paper boat drifting down a rainy gutter, macro shot",
    "duration": 5,
}).json()

while True:
    r = requests.get(f"{BASE}/v1/videos/{task['id']}", headers=H).json()
    if r.get("status") in ("succeeded", "success", "failed", "error"):
        print(r)
        break
    time.sleep(10)

Image-to-video

Pass a reference image alongside the prompt; the field name follows the model’s own API (commonly image, image_url, or input_reference). Triw forwards unknown fields to the channel, so the provider’s parameter names are the ones that work.

curl https://router.triw.io/v1/videos \
  -H "Authorization: Bearer $TRIW_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kling-v2",
    "prompt": "slow push-in, dust particles in the light",
    "image": "https://example.com/still.jpg",
    "duration": 5
  }'

Other operations

Method	Path	Purpose
`POST`	`/v1/videos`	Create a task
`GET`	`/v1/videos/{task_id}`	Task status / result
`POST`	`/v1/videos/{task_id}/remix`	Re-run a task with changed parameters

Provider-native paths are also proxied when the matching plugin is enabled — for example Kling’s /v1/videos/text2video, Alibaba’s /api/v1/services/aigc/video-generation/video-synthesis, or Doubao’s /ent/v2/text2video. Use those only if you already have code written against that provider’s SDK.

Practical notes

Topic	Detail
Latency	Renders take 1–10 minutes depending on model, duration and resolution
Idempotency	Create one task per user action; retrying the POST creates a second billable render
Storage	Result URLs are provider-hosted and can expire — download the asset after success
Pricing	Billed per task, scaling with duration and resolution; check the model price first
Availability	Only models enabled for your key’s group are callable; confirm with `GET /v1/models`

Troubleshooting

Symptom	Cause	Fix
`model_not_found`	Model absent from the key’s group	Create a key in a group that includes video models
`task_not_exist`	Wrong id, or task belongs to another account	Poll with the same key that submitted it
`not_implemented` on fetch	Channel has no OpenAI-video converter	Read the task through the provider-native path instead
Task stuck in progress	Upstream queue congested	Wait; if it exceeds 30 minutes, treat as failed and resubmit
400 on extra fields	Parameter not accepted by that model	Remove it and follow the provider’s own parameter list

Getting started

Account

API

SDKs

Tools