Skip to main content
For an agency localizing an existing speaker video, Magic Hour’s lip-sync workflow takes a video and replacement audio. It animates mouth movement; it does not by itself validate a translation, approve a voice or finish the ad. If your starting asset is a still portrait, use Talking Photo instead. This is a pilot process, not a claim that Magic Hour outperforms another provider.

​
Prepare one representative clip

  1. Obtain permission for the footage, likeness and voice. Approve the translated script and replacement recording before generation. Check the current terms of service: commercial use requires a paid subscription, credit-pack purchase or separately agreed paid usage-based arrangement, subject to applicable rights.
  2. Start with a short, clear single-speaker segment with visible lips and stable lighting. Set the segment with start_seconds and end_seconds and provide enough audio for it. The Lip Sync guide lists input requirements and supported audio formats.
  3. Set a small generation budget and a fixed retry allowance. Check style.generation_mode, max_fps_limit and current pricing before submitting.
  4. Run the lip-sync starter recipe or use the web workflow for a manual pilot. The recipe uploads local files, submits a project, waits and downloads the output.

​
Review the delivered result

Watch the complete output with sound and compare it with the source. Review speech onset, pauses, visible mouth closures, head turns, face texture and identity. Check duration, audio alignment and required export settings. Record a concrete reason when you reject an output. Keep separate cases for multiple speakers, frequent cuts, obscured mouths and larger movements. A result on a clean single-speaker clip does not establish quality for those cases.

​
Compare alternatives on the same client brief

Use the same approved input video and voiceover for each compatible configuration. Fix the clip length, output requirements and retry budget. Record provider, model or generation mode, settings, date and every attempt, including failures. Use the API evaluation scorecard to calculate:
  • Accepted-output rate: accepted outputs divided by submitted jobs.
  • Generation cost per accepted output: all generation charges, including charged retries and rejected outputs, divided by accepted outputs. Leave this undefined if none are accepted.
  • Time to usable output: submission through download and review, with operator time recorded separately.
For Magic Hour, record the project’s credits_charged and convert credits using your own billing arrangement. Do not compare one provider’s credit count directly with another provider’s dollar price or treat a mock response as a quality benchmark.

​
Deliver and repeat safely

Retain the project ID if polling times out and check that job before creating another. Download successful output before its signed URL expires. Keep client media, keys and signed URLs out of public scorecards; use permitted example assets for any public demonstration. For batch delivery, add bounded concurrency, per-client budget tracking and an explicit review step. Follow production integration and webhook signature verification for application workflows. Re-run your matched cases when generation settings or the client brief change.