Omni Human

Generate professional‑quality animated human videos from a single image and audio clip

Model Overview

OmniHuman turns a still photograph of a person and an accompanying audio clip into a realistic, motion‑based video. By conditioning on the image’s pose and the audio’s rhythm, it creates a high‑fidelity animation that captures facial expressions, lip‑sync and body movement.

Best At

• Quickly produce polished avatar videos for social media, branded content or educational videos.
• Works well with short, high‑quality audio (≤15 s) – the model keeps sync and motion realistic.
• Handles any aspect ratio image (portrait, half‑body, full‑body) and adapts video output accordingly.
• Supports a wide range of styles, from realistic portraits to cartoon‑like characters.

Limitations / Not Good At

• Audio longer than 15 s begins to degrade video quality, and the model is not designed for long‑form content.
• Requires a clear, high‑resolution reference image – low‑quality or heavily occluded images produce blurry or mismatched results.
• Complex lighting or extreme poses may not translate perfectly, as the model relies heavily on learned motion patterns.
• Not suited for driving the video with external video input; it does not accept a video as a locator of the driving sequence.

Ideal Use Cases

• Short TikTok or Reels animations with a brand avatar or influencer.
• Product showcase videos where a spokesperson appears animated.
• Educational or training clips featuring an animated presenter.
• Marketing promos that need a quick, polished video without a lot of editing.

Input & Output Format

• Inputs:
  • image – URL or file path to a human image (any aspect ratio).
  • audio – URL or file path to an MP3/WAV clip (best quality under 15 s).
• Output: a URI pointing to the generated MP4 video.

Performance Notes

• Generates a single clip relatively quickly (minutes for a 15 s video on typical GPU nodes).
• Video quality scales with audio length; shorter clips produce cleaner results.
• The process is GPU‑intensive due to rendering; single prompts are fast, batch processing is more resource‑heavy.