Skip to main content
An avatar workflow combines a visual identity with audio and performance direction. You can brief the agent in a Space or run the model directly from the Video tool.

​
Create with the agent

1

Prepare the inputs

Use a clear character image and clean speech or music. A front-facing image with an unobstructed face is the most reliable starting point.
2

Start from a presenter or UGC Skill

On the home screen, select presenter or ugc, or describe the intended performance in a new Space.
3

Add the references

Attach the character image and audio. Explain who should speak, the framing, the tone, and any movement or background constraints.
4

Review and refine

Check lip sync, identity, framing, and gesture. Ask for a bounded change while preserving the references that already work.

​
Use the manual Avatar mode

Open Manual tools → Video, then choose Avatar.
  1. Select a model from the filtered catalog.
  2. Add the image and audio references required by that model.
  3. Describe the desired performance.
  4. Choose the available aspect ratio, resolution, duration, and batch size.
  5. Select Generate.

​
Better inputs produce better performances

  • Use a sharp, well-lit face at a useful scale.
  • Avoid heavy occlusion over the mouth and eyes.
  • Use clean audio without clipping or excessive background noise.
  • Start with a short test before generating a long performance.
  • For multiple people, make the intended speaker unambiguous in the prompt and use speaker controls when the selected model exposes them.
Avatar models accept different input combinations and durations. The live model selector shows the requirements for the model you choose.