@apicity/kie

v0.1.0

Published

7 days ago

Kie provider for video and image generation (Kling 3.0, Grok Imagine, Nano Banana Pro).

0High
0Medium
0Low

justintanner

ai kei kling grok video-generation image-generation apicity

@apicity/kie

Kie provider for video and image generation (Kling 3.0, Grok Imagine, Nano Banana Pro).

Installation

npm install @apicity/kie
# or
pnpm add @apicity/kie

Quick Start

import { kie as createKie } from "@apicity/kie";

const kie = createKie({ apiKey: process.env.KIE_API_KEY! });

Real-world example: Kling 3.0 video with named element references

Kling 3.0/video has a feature most upstream video models lack — named element references. You upload reference images for each subject, give the subject a name and description, and refer back to it from the prompt with [name] placeholders. The result preserves identity across frames without prompt-engineering gymnastics.

The flow below is taken verbatim from tests/integration/kie-kling-30-reference-bakeoff.test.ts and replays against tests/recordings/kie_2079838932/kling-30-reference-bakeoff_875607413/recording.har, so the response shapes match what Kie actually returns.

import { readFileSync } from "node:fs";
import { kie as createKie } from "@apicity/kie";

const kie = createKie({ apiKey: process.env.KIE_API_KEY! });

// 1. Upload each reference image. Kie returns a CDN-hosted
//    `downloadUrl` you'll thread through the job request — the bytes
//    themselves never live in the createTask payload.
async function upload(filename: string, mimeType: string): Promise<string> {
  const blob = new Blob([readFileSync(filename)], { type: mimeType });
  const res = await kie.post.api.fileStreamUpload({
    file: blob,
    filename,
    uploadPath: "images/test-uploads",
  });
  if (!res.data?.downloadUrl) throw new Error("upload failed");
  return res.data.downloadUrl;
}

// Two angles of the cat satisfy Kling's "2-4 images per element"
// minimum. The man only has one fixture, so we pass it twice.
const cat1  = await upload("cat1.jpg",  "image/jpeg");
const cat2  = await upload("cat2.jpg",  "image/jpeg");
const man   = await upload("man.jpg",   "image/jpeg");
const beach = await upload("beach.png", "image/png");

// 2. Submit the job. Each `kling_elements` entry binds a `name` to a
//    set of reference images; the prompt then refers to that subject
//    via `[name]`. `image_urls` carries general/setting references —
//    the beach plate here — that aren't tied to a named subject.
const task = await kie.post.api.v1.jobs.createTask({
  model: "kling-3.0/video",
  input: {
    prompt:
      "On a sandy beach with the ocean behind, [blue_suit_man] sits " +
      "cross-legged on the sand. [white_cat] climbs onto his lap, " +
      "purrs, and lifts a paw to bat playfully at his blue tie. " +
      "[blue_suit_man] smiles and waves at the camera with his free hand.",
    image_urls: [beach],
    kling_elements: [
      {
        name: "white_cat",
        description: "A white cat with mismatched yellow and blue eyes",
        element_input_urls: [cat1, cat2],
      },
      {
        name: "blue_suit_man",
        description: "A man wearing a blue suit and a blue tie",
        element_input_urls: [man, man],
      },
    ],
    sound: false,
    duration: "5",
    aspect_ratio: "16:9",
    mode: "std",
    multi_shots: false,
  },
});

if (!task.data?.taskId) throw new Error("createTask returned no taskId");
const taskId = task.data.taskId;
// → "56074f12319c68b246e5a03e05608f31"

// 3. Poll recordInfo until the job leaves the `generating` state.
//    Kie's terminal states are `success` and `fail` — anything else
//    (`waiting`, `queuing`, `generating`) means keep waiting.
let state: string = "waiting";
let resultJson: string | undefined;
for (let i = 0; i < 200; i++) {
  const info = await kie.get.api.v1.jobs.recordInfo(taskId);
  state = info.data?.state ?? "waiting";
  if (state === "success" || state === "fail") {
    resultJson = info.data?.resultJson;
    break;
  }
  await new Promise((r) => setTimeout(r, 10_000));
}
if (state !== "success" || !resultJson) throw new Error(`job ${state}`);

// 4. `resultJson` is a JSON-encoded string — parse it to get the
//    delivered media URLs. The shape is consistent across every kie
//    media model (single `resultUrls: string[]`).
const result = JSON.parse(resultJson) as { resultUrls: string[] };
console.log(result.resultUrls[0]);
// → "https://tempfile.aiquickdraw.com/k/56074f12319c68b246e5a03e05608f31_1_1777540551_7969.mp4"

Notes

kling_elements accepts at most 3 named subjects, and each element_input_urls array must hold 2–4 images. If you only have one reference per subject, repeat it (as the example does for man) to satisfy the minimum.
The poll loop watches for the terminal states success and fail — anything else (waiting, queuing, generating) means keep going. There's no check_after_secs hint as on some other providers, so a 10s cadence is conservative; in the recorded fixture the job completed after ~127s of generating (14 polls).
kie.post.api.v1.jobs.createTask is the unified entry point for every media model in this provider — Kling, Wan, Seedance, Grok Imagine, GPT-Image-2, Qwen2, and others. Swap the model and input block; the rest of the flow (upload → createTask → poll → parse resultJson) is identical.
For convenience, submitMediaJob(provider, request) and uploadFile(provider, blob, filename, uploadPath) re-export the same calls but throw KieError directly when the upstream envelope is missing the expected fields.
Errors surface as KieError with status, body, and an upstream code attached, so try { ... } catch (e) { if (e instanceof KieError) ... } gives you the upstream error directly.

API Reference

26 endpoints across 15 groups. Each method mirrors an upstream URL path.

chat

GET https://api.kie.ai/api/v1/chat/credit

Published

Vulnerabilities

Links

Maintainers

Keywords

Readme

@apicity/kie

Installation

Quick Start

Real-world example: Kling 3.0 video with named element references

API Reference

chat

claude

common

fileBase64Upload

fileStreamUpload

fileUrlUpload

generate

jobs

lyrics

midi

mp4

style

veo

vocalRemoval

wav

Middleware

License