An e-commerce seller wants thirty product videos before the morning rush. There is no studio, no presenter, and no budget for either. What she does have is a single photograph and an audio file — and, since ByteDance opened up its digital human research, that is enough to make the photograph talk.
What OmniHuman actually does
OmniHuman is ByteDance's audio-driven human video model. Feed it one image of a person — portrait, half-body or full-body — plus a voice track, and it returns a video in which the subject speaks or sings with matching lip-sync, facial expressions, head movement and body gestures.
The first version, OmniHuman-1, was released in February 2025. Two details separated it from earlier talking-head tools:
- It animates any framing, not just a tightly cropped face — hands, gestures and object interactions included.
- It handles stylized characters, not only real humans, so anime figures and 3D cartoon characters move as naturally as a photographed person.
The 1.5 upgrade: from mouth movement to performance
In late August 2025, ByteDance's digital human team shipped OmniHuman-1.5. The research framing borrows the "dual-system" idea from cognitive science: a slow system reads what is being said and plans the reaction, while a fast system renders the motion.
In practice, that changes what the output feels like:
- Emotion awareness — the model matches expressions and gestures to the tone and meaning of the audio, not just its rhythm.
- Two-person scenes — a single generation can drive two characters interacting, with accurate lip-sync on both.
- Longer takes — clips can run past one minute while keeping the character consistent.
- Text-direction control — prompts can steer camera style, scene mood and specific gestures.
The team published the technical report as arXiv 2508.19209, and third-party coverage of the release emphasized the same shift: from "speaks correctly" to "performs convincingly."
Where consumers actually touch it
The consumer gateway is Jimeng (即梦), ByteDance's creative platform. On March 7, 2025, Jimeng's digital human feature launched a "Master Mode" (大师模式) powered by OmniHuman-1, letting anyone upload a photo and an audio clip to produce footage of that person speaking, singing, playing an instrument or walking around — through both the web app and mobile client.
ByteDance said it applied a strict review process to the feature and stamps every output video with a watermark that flags it as AI-produced, an attempt to keep the tool from being misused on real people's likenesses.
Developers reach the same technology through BytePlus and Volcano Engine (火山引擎), ByteDance's cloud arm, which positions OmniHuman as a building block for education, marketing and virtual-host products rather than a standalone toy.
The workflow around the model
In practice, the model slots into a simple pipeline that non-professionals can run end to end. A creator drafts a script, records or synthesizes a voice track, generates the talking video from a photo, then edits the result in CapCut (剪映) or another editor alongside b-roll and captions. Because the human never appears on camera, revisions are cheap: change the script, regenerate the voice, and the presenter "re-records" instantly.
That economics explain why the technology spread first among small operators. A solo seller cannot afford retakes; a generated presenter never fluffs a line, never needs a studio booking, and works in as many languages as the voice track supports. The trade-off is texture — audiences increasingly notice when a presenter never blinks at the wrong moment or gestures a half-second late — so experienced users keep clips short and cut away often.
Who is using it
Typical production patterns have settled into a few lanes:
- E-commerce: product explainers and storefront hosts in multiple dialects and languages.
- Education: lecture-style videos where a generated presenter reads prepared scripts.
- Marketing: localized ad variants, where only the voice track changes per market.
- Character content: anime and cartoon hosts for channels that never show a real face.
Key takeaways
- OmniHuman turns a single image plus an audio track into a talking or singing video, including cartoon characters.
- Version 1.5 added emotion-aware performance, two-person scenes, minute-long takes and prompt control.
- The consumer entry point is Jimeng's Master Mode; developers get access via BytePlus and Volcano Engine.
- Every output carries a marking that flags machine-made footage, and the platform reviews submissions.
Honest limitations
Lip-sync still drifts on long, technical scripts, and generated gestures can look theatrical in ways that read as uncanny on a second viewing. The model animates a still image; it cannot re-voice existing footage, which limits repair workflows. Consent is the harder problem: animating a real person's photo is one click away, and platform review plus watermarks reduce but do not eliminate misuse. Availability and features also vary by region and by platform, so what a demo shows is not always what your account can generate today.
What readers can do now
- Try Master Mode with your own photo and a short script first; keep clips under a minute and check mouth shapes on consonants before scaling up.
- If you plan to animate anyone else's likeness, get written consent and keep the AI-produced label intact — stripping watermarks invites both platform bans and legal exposure.
- For business use, test the dialect and language support against your actual audience before committing a campaign to a synthetic presenter.