工作原理

所有 Kie.ai API 都是异步的：

1. 创建任务 → 获取 taskId
2. 轮询状态直到 success 或 fail
3. 下载生成的文件（URL 有效期为 14 天，直接下载链接有效期为 20 分钟）

两种 API 模式

市场统一接口（大多数模型）

• 创建：POST /api/v1/jobs/createTask，附带 {"model":"<model-id>","input":{...}}
• 轮询：GET /api/v1/jobs/recordInfo?taskId=<id>
• 状态：waiting → queuing → generating → success / fail

专用端点（Runway, Veo3, 4o Image, Flux Kontext）

• 每个模型都有其独立的创建/轮询/下载路径
• 轮询使用 successFlag：0=生成中，1=成功，2/3=失败

生成工作流程

步骤 1：询问用户要生成什么

根据任务呈现模型选项：

对于图像，按顺序推荐：

1. Nano Banana 2 (nano-banana-2) — 最佳全能型，4K，支持 Google 搜索基础，14 张参考图像
2. Imagen 4 (google/imagen4) — Google 最新模型，支持负面提示词
3. Flux Kontext（专用端点）— Pro/Max 层级，非常适合编辑
4. 4o Image（专用端点）— GPT-4o 驱动，支持蒙版编辑
5. Seedream 5.0 Lite — 字节跳动推理模型，每个请求最多生成 6 张图像
6. Flux-2, Grok Imagine, Qwen, Ideogram, GPT Image 1.5

对于视频，按顺序推荐：

1. Kling 3.0 (kling-3.0/video) — 多镜头，支持音频，3-15秒，包含元素
2. Veo 3.1（专用）— Google，质量/快速模式，1080p/4K
3. Sora 2 (sora-2-text-to-video) — OpenAI，角色，故事板
4. Runway Gen-4（专用）— 720p/1080p，5-10秒
5. Wan 2.6 (wan/2-6-*) — 阿里巴巴，Flash 变体，支持音频
6. Hailuo, Bytedance Seedance

对于音频：Suno V4（音乐），ElevenLabs（TTS，音效，对话） 对于超分辨率：Topaz（图像/视频），Recraft（背景移除，超分辨率）

步骤 2：收集参数

用户选择模型后，询问模型特定的设置。请阅读 references/models-catalog.md 了解所选模型的确切参数。

各模型间的通用参数：

• prompt（生成时始终必需）
• aspect_ratio / size（因模型而异）
• callBackUrl（可选，轮询时可省略）
• 用于编辑/I2V 模式的图像/视频 URL

步骤 3：执行

使用辅助脚本：

# 市场模型（大多数模型）
bash scripts/kie_api.sh create-market '{
  "model": "nano-banana-2",
  "input": {
    "prompt": "A futuristic cityscape at sunset",
    "aspect_ratio": "16:9",
    "resolution": "2K"
  }
}'

# 专用模型（Runway, Veo3, 4o Image, Flux Kontext）
bash scripts/kie_api.sh create "/api/v1/veo/generate" '{
  "prompt": "A cat exploring a magical forest",
  "model": "veo3_fast",
  "aspect_ratio": "16:9"
}'

步骤 4：轮询和下载

# 市场模型
bash scripts/kie_api.sh poll-market "task_12345" 60 10

# 专用模型
bash scripts/kie_api.sh poll "/api/v1/veo/record-info" "veo_task_abc" 60 10

# 下载结果
bash scripts/kie_api.sh download "https://tempfile..." "./output.png"

步骤 5：呈现结果

向用户显示下载 URL 或本地文件路径。如果有多个 URL，请全部列出。

速率限制

• 每 10 秒最多 20 个新请求
• 约 100 个并发任务
• HTTP 429 = 速率受限，请等待后重试

错误处理

文件上传（当用户有本地文件时）

# 上传本地文件
bash scripts/kie_api.sh upload-stream "/path/to/image.png"

# 从 URL 上传
bash scripts/kie_api.sh upload-url "https://example.com/image.png"

上传响应包含 fileUrl，可用作输入。文件在 3 天后过期。

定价参考（积分）

检查积分：bash scripts/kie_api.sh credits

每周安装数

1

代码仓库

antbald/kie-ai-skill

首次出现

1 天前

安全审计

Gen Agent Trust HubPassSocketPassSnykWarn

安装于

mcpjam1

claude-code1

junie1

windsurf1

zencoder1

crush1

How It Works

All Kie.ai APIs are asynchronous :

1. Create task → get taskId
2. Poll status until success or fail
3. Download the generated file (URLs valid 14 days, direct downloads 20 min)

Two API Patterns

Market Unified (most models)

• Create: POST /api/v1/jobs/createTask with {"model":"<model-id>","input":{...}}
• Poll: GET /api/v1/jobs/recordInfo?taskId=<id>
• States: waiting → queuing → generating → success / fail

Dedicated Endpoints (Runway, Veo3, 4o Image, Flux Kontext)

• Each has its own create/poll/download paths
• Poll uses successFlag: 0=generating, 1=success, 2/3=failed

Generation Workflow

Step 1: Ask user what to generate

Present model options based on the task:

For images , recommend in order:

1. Nano Banana 2 (nano-banana-2) — Best all-round, 4K, Google Search grounding, 14 ref images
2. Imagen 4 (google/imagen4) — Google's latest, negative prompts
3. Flux Kontext (dedicated endpoint) — Pro/Max tiers, great for editing
4. 4o Image (dedicated endpoint) — GPT-4o powered, mask editing
5. Seedream 5.0 Lite — ByteDance reasoning model, up to 6 images per request
6. Flux-2 , Grok Imagine , Qwen , Ideogram , GPT Image 1.5

For videos , recommend in order:

1. Kling 3.0 (kling-3.0/video) — Multi-shot, audio, 3-15s, elements
2. Veo 3.1 (dedicated) — Google, Quality/Fast modes, 1080p/4K
3. Sora 2 (sora-2-text-to-video) — OpenAI, characters, storyboard
4. Runway Gen-4 (dedicated) — 720p/1080p, 5-10s
5. Wan 2.6 (wan/2-6-*) — Alibaba, Flash variants, audio
6. Hailuo , Bytedance Seedance

For audio : Suno V4 (music), ElevenLabs (TTS, SFX, dialogue) For upscale : Topaz (image/video), Recraft (background removal, upscale)

Step 2: Collect parameters

After user picks a model, ask for model-specific settings. Read references/models-catalog.md for the exact parameters of the chosen model.

Common parameters across models:

• prompt (always required for generation)
• aspect_ratio / size (varies per model)
• callBackUrl (optional, omit for polling)
• Image/video URLs for edit/I2V modes

Step 3: Execute

Use the helper script:

# Market models (most models)
bash scripts/kie_api.sh create-market '{
  "model": "nano-banana-2",
  "input": {
    "prompt": "A futuristic cityscape at sunset",
    "aspect_ratio": "16:9",
    "resolution": "2K"
  }
}'

# Dedicated models (Runway, Veo3, 4o Image, Flux Kontext)
bash scripts/kie_api.sh create "/api/v1/veo/generate" '{
  "prompt": "A cat exploring a magical forest",
  "model": "veo3_fast",
  "aspect_ratio": "16:9"
}'

Step 4: Poll and download

# Market models
bash scripts/kie_api.sh poll-market "task_12345" 60 10

# Dedicated models
bash scripts/kie_api.sh poll "/api/v1/veo/record-info" "veo_task_abc" 60 10

# Download result
bash scripts/kie_api.sh download "https://tempfile..." "./output.png"

Step 5: Present result

Show the user the download URL or local file path. If multiple URLs, list them all.

Rate Limits

• Max 20 new requests per 10 seconds
• ~100 concurrent tasks
• HTTP 429 = rate limited, wait and retry

Error Handling

File Upload (when user has local files)

# Upload local file
bash scripts/kie_api.sh upload-stream "/path/to/image.png"

# Upload from URL
bash scripts/kie_api.sh upload-url "https://example.com/image.png"

Upload response contains fileUrl to use as input. Files expire in 3 days.

Pricing Reference (credits)

Check credits: bash scripts/kie_api.sh credits

Weekly Installs

1

Repository

antbald/kie-ai-skill

First Seen

1 day ago

Security Audits

Gen Agent Trust HubPassSocketPassSnykWarn

Installed on

mcpjam1

claude-code1

junie1

windsurf1

zencoder1

crush1