OmniHuman 1.5

OmniHuman 1.5 turns a static photo and an audio track into a video of that person appearing to speak or perform the audio naturally — with lip sync, facial expression, and body language that respond to the tone and pacing of the voice, not just its phonemes. It's the model to reach for whenever the goal is a talking presenter, avatar, or character rather than a generated scene.

Because output length tracks the audio input rather than a fixed clip duration, it's well suited to voiceover-driven content — narrated explainers, presenter-style videos, or a character singing along to a track — where the video needs to match a pre-existing audio timeline exactly.

OmniHuman 1.5 works fundamentally differently from prompt-driven video models like Veo, Kling, or Seedance. Instead of generating a scene from a text description, it takes two required inputs — a single reference photo or avatar image, and an audio track — and animates the person in the photo to speak and move in sync with that audio. Duration isn't set by a prompt; it's driven by the length of the audio track you provide.

Under the hood, ByteDance's approach fuses a multimodal understanding component with a diffusion-based video generator: the model reads semantic and prosodic cues from the audio (tone, emphasis, pacing) and translates them into lip movement, facial expression, and body language, rather than only matching mouth shapes to phonemes. That's what produces motion that reads as emotionally responsive to the audio rather than mechanically lip-synced.

It offers standard and high quality tiers, trading render time and cost for fidelity in facial detail and motion smoothness. Because it only needs a still image and an audio file as source material, it doesn't require filming a presenter or actor at all.