---
name: ai-video-generation-api-workflow
description: "Generate video via Kling 3.0, Veo 3.1, and Runway Gen-4.5 APIs — model routing, image-to-video, async polling. Production text/image-to-video across providers: Kling 3.0 (camera movement controls, start/end keyframes, motion brush), Google Veo 3.1 via Gemini/Vertex AI (native audio, 1080p), and Runway Gen-4.5 (Gen-4 References, image_to_video). Covers async job submission and polling (Kling task_id status, Runway task GET, Veo operations.get long-running ops), prompt and negative_prompt structure, cfg_scale/duration/aspect_ratio params, seed control, and per-model cost/latency tradeoffs for model selection by use case."
license: MIT
---

# AI Video API Workflow

Generate video programmatically across the three leading providers — Kuaishou Kling 3.0, Google Veo 3.1, and Runway Gen-4.5 — from one workflow. This skill handles the parts that differ per vendor: authentication, request shape, model routing, and the asynchronous submit-then-poll lifecycle that all three use. It's for developers building video pipelines who don't want to maintain three separate SDK integrations and three polling loops.

## When to use it

- Text-to-video or image-to-video generation from a script, batch, or backend job.
- Animating a still: pass a start frame (and optionally an end frame) for an interpolated clip.
- Routing by capability — e.g. Veo when you need synchronized native audio at 1080p, Kling when you need explicit camera moves or motion-brush control, Runway when you want Gen-4 References for character/style consistency.
- Long-running renders where you submit a job, store the id, and reconcile results later instead of holding a connection open.

## How it works

Each provider exposes an async job model, and the skill normalizes them:

- **Kling 3.0** — POST to the image2video / text2video endpoints returns a `task_id`; poll the task-status endpoint until `task_status` is `succeed` or `failed`. Supports `camera_control`, start/end `image`/`image_tail` keyframes, and motion brush via dynamic masks.
- **Veo 3.1** — call through the Gemini API (`generateVideos` / long-running operation) or Vertex AI; poll the returned operation until done. Produces native audio and 1080p output.
- **Runway Gen-4.5** — create an `image_to_video` task, then poll the task by id until `SUCCEEDED`; use Gen-4 References to lock identity/style across shots.

## Usage example

```
> Animate hero.png with Kling: slow dolly-in, 5s, then poll until done

# Skill submits image2video with camera_control, captures task_id,
# polls task-status on a backoff, and returns the signed video URL.

> Same prompt, render with Veo 3.1 for narration audio at 1080p
```

## Tips & gotchas

- **Result URLs expire.** Kling and Runway hand back time-limited download links — fetch and persist the asset immediately; don't store the URL as permanent.
- **Don't hammer the poll.** Use exponential backoff (renders take 30s–several minutes). Tight polling can trip rate limits before the job finishes.
- **Keyframe pairing is provider-specific.** Kling's end frame is `image_tail`; passing both start and end constrains motion more than most prompts expect.
- **Auth differs.** Kling uses an access-key/secret JWT, Veo uses Gemini API keys or Vertex service-account ADC, Runway uses a bearer key — keep them in separate env vars.

## FAQ

**Can I get a webhook instead of polling?** Runway supports task webhooks; Kling and Veo are poll-based, so the skill defaults to polling with backoff everywhere for consistency.

**Which model for lip-synced talking video?** Veo 3.1 — it's the only one of the three with native synchronized audio.

**How do I keep a character consistent across clips?** Use Runway Gen-4.5 with Gen-4 References, or reuse a fixed start keyframe in Kling.