This page explains where the model comes from, what changed across versions and how ordinary creators can use it. One note up front: omnihuman2.ai is an independent service that runs the model through a licensed provider — we are not ByteDance or BytePlus and are not endorsed by them.
Where ByteDance OmniHuman 2 comes from
The OmniHuman line started as a research project presented in early 2025, notable for generating full-body, gesturing humans from a single picture rather than only animating a cropped face. Subsequent releases improved realism and controllability, and version 1.5 became widely available through model hosting platforms.
Version 2.0 is the current commercial release. BytePlus describes it as film-grade, with camera and body movement, multi-person scenes and gesture control, and offers a separate avatar-focused variant tuned for lip sync and expression.
What the model can do
Based on the published product description and input specifications:
- Animate one image with speech or singing from 4 seconds up to 5 minutes.
- Render multi-person scenes from a single group photo.
- Follow text prompts for camera movement, body movement and gestures.
- Accept prompts in six languages and work with stylised, 3D and animal characters.
- Output native 720p video.
Ways to access it
There are two broad routes. Developers can call the model through official cloud and model-hosting APIs, handling uploads, queues, storage and billing themselves. Creators who simply want videos can use a hosted studio like this one, where the technical pipeline is already built and you pay per second of video.
Whichever route you choose, the inputs are the same: a clear photo, an MP3 or WAV file, and optionally a short direction prompt.
Responsible use
Realistic talking video of real people can be misused. Only animate photos and voices you own or have permission to use, label synthetic content where platforms require it, and never use the technology to impersonate someone or mislead viewers. Our terms prohibit deceptive deepfakes, and we act on reports of misuse.
About this site
omnihuman2.ai is operated by Nexa Labs, LLC. OmniHuman is a trademark of its respective owner; we use the name to describe the model our service runs. For official announcements, research details and enterprise agreements, refer to ByteDance and BytePlus directly.
How the model works at a high level
OmniHuman belongs to the family of diffusion-transformer video models. During training it learns to generate human motion conditioned on several signals at once — audio, reference appearance, pose and text — which is why the research describes it as “omni-conditioned”. Mixing these conditions lets the model learn from much more varied video data than models that only see tightly cropped talking heads.
At inference time you only supply the reference image, the audio and optionally a text prompt; the model infers body pose, facial motion and camera behaviour from them. That is also why the input image matters so much: it defines identity, framing, lighting and what parts of the body are available to move.
Versions at a glance
A quick summary of the public releases:
- OmniHuman-1: research release showing full-body animation from one image.
- OmniHuman 1.5: widely hosted version with prompt guidance, an optional speaker mask and 720p or 1080p output for short clips.
- OmniHuman 2.0: commercial release with clips up to five minutes, multi-person scenes and camera, body and gesture control at 720p.
Why it matters for creators
Before models like this, a talking presenter required a camera, a person and an edit. Now a single image carries a whole recurring series, which changes the economics of explainers, ads and music content for small teams. The trade-offs — 720p output and minutes of render time — are acceptable for most social and web formats.
FAQ
Who makes OmniHuman 2?
OmniHuman is developed by ByteDance and offered commercially through BytePlus and partner model platforms.
Is omnihuman2.ai an official ByteDance site?
No. We are an independent service that runs the model through a licensed provider and are not affiliated with or endorsed by ByteDance or BytePlus.
Is OmniHuman available in TikTok or CapCut?
We can’t speak for ByteDance’s consumer apps. This site gives you direct access to the OmniHuman 2.0 model in a browser.
What does OmniHuman 2 need as input?
One image and one MP3 or WAV audio file between 4 seconds and 5 minutes, plus an optional text prompt.
Can I use the output commercially?
On our paid plans, yes — provided you own the rights to the photo and audio you upload.
Where can I read the research?
Search for the OmniHuman technical reports published by ByteDance’s research team.
Is OmniHuman open source?
The research has been published as papers, but the production 2.0 model is offered as a commercial service through ByteDance’s cloud and partner platforms rather than as downloadable weights.
What is the avatar variant BytePlus mentions?
BytePlus lists a separate avatar-oriented model focused on lip sync and expression alongside the main 2.0 model, which emphasises camera, body and multi-person performance.