Contents
How OmniHuman worksOmniHuman versionsOmniHuman at a glanceOmniHuman vs other avatar approachesTips for better OmniHuman resultsFrequently asked questionsOmniHuman is ByteDance's audio-driven human video model. Give it a single image of a person or character and an audio track, and it generates a video in which the subject speaks or sings with matching lip-sync, facial expressions, head movement and body gestures. The research model OmniHuman-1 was unveiled in February 2025, and OmniHuman 1.5, unveiled in August 2025, added semantic understanding of the audio so gestures and emotions match what is being said, not just the rhythm.
In Rekreate, OmniHuman powers audio-driven avatar videos: you supply a portrait, a voice track (for example a cloned voice) and a short guiding prompt.
- Input: one image (portrait, half-body or full-body) + audio + optional text guidance.
- Output: lip-synced talking or singing video with natural gestures.
- OmniHuman-1: February 2025 · OmniHuman 1.5: August 2025 with MLLM-planned emotions and gestures, clips up to ~30 s.
- Best for: UGC talking heads, faceless explainers, character dialogue, music clips.
- In Rekreate: image + audio + prompt, with automatic subject detection for the speaking person.
How OmniHuman works
Earlier talking-head tools mostly animated the mouth region of a cropped face. OmniHuman was trained with "omni-conditions" — audio, pose and text signals mixed during training — so it can animate any framing: close-up portraits, half-body shots and full-body images, including hands, gestures and interactions with objects.
OmniHuman 1.5 introduced a dual-system design: a multimodal large language model first interprets the meaning and emotion of the audio (and optional text), plans appropriate reactions and gestures, and then a diffusion transformer renders the motion. The result is a performance that reacts to content — emphasis on key words, pauses, smiles on a punchline — and can follow text instructions for camera and gesture control.
OmniHuman versions
OmniHuman-1
Research paper and demos showing realistic human video from one image plus audio, across portrait, half-body and full-body framings, speaking and singing, including cartoons and stylized characters.
OmniHuman 1.5
Adds MLLM-based semantic planning for context-aware emotions and gestures, optional text guidance for camera and gestures, multi-person scenes and longer, more continuous videos (up to around 30 seconds on common APIs).
Versions available vary by platform; ByteDance offers OmniHuman through BytePlus/Volcano Engine and its Dreamina/CapCut ecosystem.
OmniHuman at a glance
| Developer | ByteDance (Intelligent Creation / OmniHuman Lab) |
|---|---|
| Inputs | Single image, audio track, optional text prompt |
| Framings | Portrait, half-body, full-body; realistic and stylized characters |
| Motion | Lip-sync, facial expressions, head and body gestures, some object interaction |
| Typical length | Up to ~30 s per generation (1.5, provider-dependent) |
| In Rekreate | Image + audio + prompt; automatic detection of the subject to animate |
OmniHuman vs other avatar approaches
- vs studio avatars (HeyGen, Synthesia): those use pre-trained digital twins recorded on video; OmniHuman animates any single image, which is faster for new characters but less controllable for very long scripts.
- vs Seedance 2.x dialogue: Seedance generates whole scenes with audio from prompts and references; OmniHuman is specialized in making a specific image speak a specific audio track with tight lip-sync.
- vs Kling Motion Control: Motion Control copies motion from a reference video; OmniHuman invents motion from audio.
For UGC, a common combo in Rekreate is: recreate a viral script → generate the voice with a cloned voice → animate a portrait with OmniHuman → add b-roll from Seedance or Veo → caption.
Tips for better OmniHuman results
- Use a sharp, front-facing image with the mouth visible and neutral expression.
- Match framing to the content: half-body images allow hand gestures; tight portraits focus on facial expression.
- Clean audio (no music under speech) gives the most accurate lip-sync.
- Keep lines natural — conversational scripts with pauses animate better than rapid monotone reads.
- Use the prompt for direction: "speaks warmly, small nods, occasional hand gesture, static camera."
- Get consent before animating a real person's likeness, and label AI-generated content where platforms require it.
Frequently asked questions
What is OmniHuman?
OmniHuman is ByteDance's AI model that generates a realistic talking or singing video from a single image and an audio track, with lip-sync, facial expressions and body gestures.
What is the difference between OmniHuman-1 and OmniHuman 1.5?
OmniHuman 1.5 (August 2025) adds a multimodal LLM that understands the meaning and emotion of the audio to plan gestures and expressions, supports text guidance for camera and gestures, and produces longer, more continuous videos than OmniHuman-1 (February 2025).
How long can OmniHuman videos be?
Common OmniHuman 1.5 APIs generate videos up to around 30 seconds per run; limits depend on the provider.
Can OmniHuman animate cartoons?
Yes. It works with stylized, cartoon and anime characters as well as realistic portraits.
Is OmniHuman available in Rekreate?
Yes. Rekreate uses OmniHuman for audio-driven avatar videos: provide an image, an audio track and a short prompt.
Can I use a celebrity photo with OmniHuman?
No. Animating real people without consent can violate platform rules and laws on likeness. Use your own image with permission, licensed avatars or original AI characters.
Create your next viral video with Rekreate
Find a proven short-form video, recreate it with your script, avatar and voice, and generate scenes with Seedance, Kling and Veo in one workspace.
Start free trial →