Luma Digital Human

../luma-shared/SKILL.md

Asset First

luma-cli asset list voice
luma-cli asset list roles

luma-cli voice clone ./voice.wav --name my_voice

roles

luma-cli asset upload avatar.mp4 --group roles

Standard Flow

1. Create or select a project:
   bash

   luma-cli project create demo
   luma-cli project use demo

2. Generate voice:
   bash

   luma-cli --json tts "script text" --voice my_voice --speech-rate 1.1 --output step2_tts.wav

   The

   --json

   flag returns

   audio_object_key

   in the output envelope. Use this key in step 3 to skip a redundant upload.
3. Generate lip-sync video (prefer

   --audio-key

   over

   --audio

   ):
   bash

   luma-cli lipsync --avatar 数字人男 --audio-key <audio_object_key> --output step3_lipsync.mp4

   If

   --audio-key

   is omitted, lipsync falls back to the project's

   latest_tts_key

   , then to

   --audio

   file upload.
4. Add subtitles:
   bash

   luma-cli subtitle step3_lipsync.mp4 --output step5_subtitle.mp4

5. Optionally enhance:
   bash

   luma-cli enhance step5_subtitle.mp4 --scale 2

Agent Notes

• Script must come from research, not imagination. If the script is for a short-video production, the text source must be backed by

  luma-cli research run

  data or a known viral reference. Never invent a script topic without data support. See

  ../luma-workflow-viral-remix/SKILL.md

  for the full research → rewrite flow.
• Use

  voice.clone

  when a user provides a voice sample.
• Uploaded videos are ambiguous. If the user did not clearly say whether the video is for voice clone, avatar/source role, PIP material, ASR/rewrite, or video processing, ask a short confirmation before choosing a workflow.
• Do not assume a user-uploaded video is a digital-human avatar. If the user asks to use "the voice/audio/sound inside this video", extract or use audio for voice clone instead.
• Use

  asset.list voice

  and

  asset.list roles

  when the user asks what is available.
• Use the latest project TTS output for lip-sync unless the user explicitly provides

  --audio

  .
• Keep the script text outside media commands until it is final enough for this generation attempt.
• Do not enhance every draft; enhance only the selected final render.
• Keep script revisions outside the media commands. The CLI should receive the final text for each generation attempt.
• Use advanced backend parameters only when the user asks for them:
  • TTS:

    --trim-long-silence

  • Lip-sync:

    --random-start

    ,

    --guidance-scale

    ,

    --num-inference-steps

    ,

    --no-superres

    ,

    --superres-scale

    ,

    --multi-shot-json