A production AI video integration is more than a single HTTP request. The application must coordinate model selection, multimodal inputs, asynchronous processing, task recovery, temporary media delivery and usage reconciliation.
Light Speed Future provides a managed API layer for native Seedance generation, standalone Video Enhancement, LSF Seedance2.5 Production, the Asset Library and Seed Audio. The products can share a familiar API surface while retaining different inputs, outputs and billing rules.
Start by choosing the correct workflow
Do not choose a model alias only by its name. Decide what the user provides and what the application must return.
| User provides | Desired result | Workflow |
|---|---|---|
| Prompt and supported references | Native generated video | Native Seedance |
| Existing source video | Enhanced version of that video | Standalone Video Enhancement |
| Prompt and supported references | Generated and enhanced 1080p, 2K or 4K MP4 | LSF Seedance2.5 Production |
| Text and supported references | Speech audio | Seed Audio |
Use the exact tenant-facing alias supplied by Light Speed Future. Direct provider model names and provider-only routing fields are not interchangeable with the LSF contract.
Keep authentication on the server
Send the LSF API key in the Authorization: Bearer header from a trusted backend. Do not embed customer credentials in frontend JavaScript, a public website or a mobile application bundle.
Use the same authorized task-owning key when retrieving a video task. Another key in the same organization should not be assumed to have access to that task or its assets.
Design around asynchronous tasks
Video generation and enhancement are background jobs. A create request can be accepted before the final media exists.
The core video flow is:
- Submit the job through
POST /v1/videos. - Store the returned public task ID.
- Retrieve the task through
GET /v1/videos/{task_id}. - Continue polling every 30 seconds or longer while the task is queued or in progress.
- Stop at completed, failed or expired.
- Read the terminal result for delivery and billing reconciliation.
Production advances automatically between generation, enhancement and optional frame extraction. Polling observes progress; it does not trigger the next internal stage.
Put Seedance controls under metadata
The LSF video API follows a familiar request pattern, but Seedance controls belong under metadata. These can include duration, resolution, aspect ratio, audio generation, ordered content references and supported output options.
For native Seedance 2.5, supported references can include images, videos and audio. The request can contain up to 30 images, 10 videos, 10 audio items and 50 total references. Keep the item order stable so the prompt can refer to @Image1, @Video1 and @Audio1 predictably.
Use documented roles such as reference_image, reference_video, reference_audio, first_frame and last_frame. A last-frame input constrains generation; it is not the same as asking Production to return a JPG from the enhanced result.
Make retries idempotent
A client timeout does not prove that a create request failed. The server may have accepted the job before the connection was interrupted.
Use metadata.client_request_id as a stable identifier for the same logical request. Store it together with the exact payload and returned task ID.
For an identical replay, keep the request ID and payload unchanged. Use a new request ID only for a deliberately new job after changing the prompt, assets or settings, or after confirming a terminal failure.
This prevents accidental duplicate generation and gives the application a reliable path for recovering from uncertain submission outcomes.
Validate remote media before submission
Reference URLs must remain accessible to the processing service. A URL that opens in a browser may still be unsuitable if it depends on cookies, private headers, VPN access, anti-bot checks or a short-lived redirect.
For Production video references, Light Speed Future performs bounded media validation and measures input duration before acceptance. Authorized Asset Library video references can use asset:// identifiers, allowing LSF to resolve access internally.
Do not send local paths, Base64 media or private-network URLs where the contract requires a public media URL.
Separate reservation from final settlement
An initial reservation is an estimate. The completed output may differ in duration or FPS from the request, and automatic-duration generation may require a conservative reservation.
Use the terminal task response and billing record as the authoritative outcome. Repeated polling does not create another job or duplicate the charge.
Remember that the pricing model depends on the workflow:
- Native Seedance follows its native usage contract.
- Standalone enhancement uses output seconds and an FPS band.
- Production uses combined input/output seconds for requests containing video and has no additional FPS multiplier.
- Seed Audio bills from successfully generated audio duration.
Store completed output promptly
Returned video, image and audio URLs are temporary delivery links. Copy media into storage controlled by the application if it must remain available.
For each publishable result, retain the task ID, client request ID, model alias, submitted settings, terminal response and relevant asset lineage. This makes support, cost reconciliation and content governance easier.
Monitor the complete product experience
Track more than endpoint uptime. Useful operating metrics include:
- Task acceptance and completion rate.
- Queue and processing time.
- Failure categories.
- Retry and replay rate.
- Final output duration and resolution.
- Usable-output rate.
- Reconciled cost per usable result.
These metrics show whether the integration is delivering a reliable customer experience rather than simply producing successful HTTP responses.
Build with Light Speed Future
Explore the Seedance 2.5 API, review API documentation or contact Light Speed Future to confirm model aliases, output combinations and Asset Library access for your application.