decart（decart 前端）

Decart

Decart is an AI video/image platform centered on realtime transformation, with three APIs: Realtime (WebRTC, sub-500ms), Queue (async batch video generation/editing), and Process (sync image generation/editing). SDKs for JavaScript, Python, Android, and Swift. Auth via API key from platform.decart.ai — set DECART_API_KEY env var. Docs: https://docs.platform.decart.ai

When to use

• Realtime video transformation — camera effects, video conferencing filters, AR/VR overlays, photo booths, live streaming
• Batch video generation — marketing clips, social content, product demos, and animations from text or images
• Batch video editing — edit existing videos with text prompts, reference images, or both
• Video restyling — apply artistic styles to live streams or existing videos
• Image generation/editing — mockups, thumbnails, creative assets, and image-to-image transformation
• Avatar animation — virtual presenters, AI customer agents, talking heads animated with audio
• Character transformation — transform user's live camera feed into characters (anime, fantasy, etc.)
• Lip sync (Python only) — use the dedicated decart.lipsync.RealtimeLipsyncClient WebSocket client to synchronize lip movements to audio in realtime
• Motion control — animate objects along precise trajectory paths
• Framework integration — Vercel AI SDK, TanStack AI, LangChain.js all supported

API selection

Model selection

Quick start patterns

Realtime (connect + transform)

const model = models.realtime("lucy-2.1");
const stream = await navigator.mediaDevices.getUserMedia({
  video: { frameRate: model.fps, width: model.width, height: model.height }
});
const rt = await client.realtime.connect(stream, {
  model,
  onRemoteStream: (s) => { videoEl.srcObject = s; }
});
await rt.set({ prompt: "Transform into anime character", image: refImage, enhance: true });

Queue (video editing)

const result = await client.queue.submitAndPoll({
  model: models.video("lucy-2.1"),
  data: videoFile,
  prompt: "Transform into anime style",
  resolution: "720p",
});

Process (image editing)

const image = await client.process({
  model: models.image("lucy-image-2"),
  data: imageFile,
  prompt: "Change the background to a beach",
  resolution: "720p",
});

Authentication

• Server-side: Set DECART_API_KEY env var. The SDK picks it up automatically.
• Client-side: Generate short-lived tokens on your backend via client.tokens.create(), then send token.apiKey to the frontend.
• Token options: expiresIn (1–3600s, default 60s), allowedModels (restrict model access), constraints (e.g. realtime.maxSessionDuration).
• NEVER expose permanent API keys in browser or mobile code.

Camera constraints

Always use model.fps, model.width, model.height from the SDK to avoid scaling artifacts.

Boundaries

What agents CAN do

• Generate integration code for all three APIs (Realtime, Queue, Process)
• Help select models, configure SDK, set up authentication
• Write WebRTC realtime streaming code
• Build framework integrations (Vercel AI SDK, TanStack AI, LangChain.js)
• Troubleshoot connection, moderation, and billing issues

What agents CANNOT do

• Create or manage Decart accounts, API keys, or billing
• Access the Decart dashboard or platform settings
• Test realtime connections (requires actual WebRTC + camera)
• Verify credit balance or usage
• Bypass content moderation policies

Common gotchas

• NEVER expose permanent API keys in client code. Use client tokens for browser/mobile.
• ALWAYS use model.fps, model.width, model.height for camera constraints. Mismatched resolution causes latency and artifacts.
• ALWAYS call disconnect() and stop media tracks when done. Leaving connections open causes memory leaks and billing.
• Catch errors from set(), setPrompt(), setImage(). Moderation rejections throw here.
• Listen to connectionChange events. The SDK auto-reconnects, but your UI must reflect the state.
• Keep enhance: true (default) unless you need exact prompt control.
• Use submitAndPoll() for queue jobs. Don't manually poll faster than every 2 seconds.
• Image editing (i2i) only works with Process API, not Queue API. Queue is for video only.
• Realtime sessions end silently when credits run out. Listen for disconnect events.
• Lip sync is not available through models.realtime("lipsync") in the main JavaScript, Android, or Swift SDKs. Use Python's decart.lipsync.RealtimeLipsyncClient instead.
• For Lucy 2 character transform, use clear portrait photos with good lighting.
• Lucy Edit v1 (lucy-clip) is legacy with a 5-second clip limit. Use lucy-2.1 for new projects (unlimited duration).
• lucy-2.1 accepts a reference image, a text prompt, or both together for maximum control.

Verification checklist

• API key stored in environment variable, never hardcoded
• Client tokens generated on backend for any browser/mobile realtime usage
• Camera resolution matches model's fps, width, height constraints
• Error handlers on set(), setPrompt(), setImage() calls
• connectionChange listener updates UI on disconnect/reconnect
• disconnect() called and media streams stopped on cleanup
• enhance: true enabled unless exact prompt control needed
• Queue API for video editing, Process API for images
• Moderation errors caught and shown to user gracefully
• Tested on real device for mobile (simulator lacks WebRTC)

Resources

• Full docs navigation: https://docs.platform.decart.ai/llms.txt
• Models overview: https://docs.platform.decart.ai/getting-started/models
• Realtime best practices: https://docs.platform.decart.ai/models/realtime/streaming-best-practices
• JavaScript SDK: https://docs.platform.decart.ai/sdks/javascript-realtime
• Android SDK: https://docs.platform.decart.ai/sdks/android