OmniHuman 2
OmniHuman 2 is a video model by ByteDance. In Vini it has one operation, Avatar by voice (Аватар with the driver голос): the picture you connect as the first frame decides who appears, the background and the framing; the sound you connect decides the mouth, the head and the gestures. It does not write a scene from text, does not edit a finished clip and does not take movement from a clip. To animate a hero by the movement of a source clip, choose the other driver of the same operation (ролик): DreamActor 2 or Kling.
When to choose it
Section titled “When to choose it”Choose it when the sound already exists (speech, a voice-over, a vocal) and the hero from a picture must say or sing it, with a body and gestures rather than only a moving mouth. In the owner’s trial it moved the hands and head more than the Avatar address of Kling. It also takes a long sound in one pass, up to five minutes. For a cheaper second the Avatar address of Kling is nearly half the price, with a more restrained mouth.
What goes in and what comes out
Section titled “What goes in and what comes out”Connect one picture as the first frame and one sound as sound 1. The sound must be between 4 and 300 seconds: a shorter sound is refused by the vendor before generation, and the studio refuses it before any charge. The sound is a WAV or MP3 file up to 15 MB; the studio converts what it can. Text is optional and is not needed to move the lips. The vendor lists English, Chinese, Japanese, Korean, Spanish and Indonesian for the text; Russian is not on the list, so write the text in English if you write it at all.
The output is as long as the sound, and the sound of the input arrives in the output unchanged; the model does not synthesize speech. The proportions of the output follow the picture. Connect one picture of the hero, not a sheet of several views.
The cost is shown on the node and in the model window before the run and is counted by the length of the sound.
| Setting | Value |
|---|---|
| Model ID | omnihuman |
| Name | OmniHuman 2 |
| Availability | available |
| Default for new nodes | no |
Common declared limits
Section titled “Common declared limits”| Setting | Value |
|---|---|
| Capabilities | image to video |
| Prompt required | no |
| Maximum prompt length | no fixed maximum declared |
| Duration | same as source video; fallback same as source video |
| Audio | origin; fallback origin |
| Aspect ratio | 9:16, 3:4, 1:1, 16:9; fallback 9:16 |
| Declared language support | prompt: English, Chinese, Japanese, Korean, Spanish, Indonesian; speech: none |
| Content constraints | people: restriction level: ok |
| Negative prompt strategy | instruction in prompt text |
Mode: avatar_audio
Section titled “Mode: avatar_audio”| Setting | Value |
|---|---|
| Aspect ratio | not declared |
| Combined reference limits | images: 1; videos: 0; audio inputs: 1; combined maximum: 2; audio without visual input: no; audio reference seconds: minimum: 4; maximum: 300 |
| First frame | required |
Unavailable modes: edit_video, motion_transfer, recast, text_to_video.
Limits apply together. initial is the studio choice for a new node; default is the fallback for an existing node without that value. Declared language or input support does not guarantee a successful result.
Prompting playbook
Section titled “Prompting playbook”- Give a picture with the face readable and the mouth uncovered; a waist-up shot works better than a full-length one.
- Give a clean voice rather than a mix. In the trial the vocal was separated from the music first; put the full track back in the montage.
- If you write text, say who is in the shot, what they do and the camera, in English: “sings to camera, static shot”.
Known failure modes
Section titled “Known failure modes”The model animates whatever picture it gets. A character sheet with several views makes a sheet that moves, so connect a single view of the hero. A sound under four seconds is refused. Waiting takes minutes: about two to three minutes for five to ten seconds of sound, so the node shows progress.
Evidence and limits
Section titled “Evidence and limits”On 22 October 2026 the owner gave one cartoon tomato picture and 8.5 seconds of a Russian vocal to this model and to the Avatar address of Kling. Both moved the mouth and the head; OmniHuman 2 added more gestures and body. An earlier OmniHuman version scored better on a coarse lip-sync metric, and the owner chose version 2. Two further runs of five and ten seconds confirmed that the output length equals the sound length, that the sound arrives in the output and that the price grows in proportion to the sound. These are observations of one scene type: they do not cover real faces, plain speech instead of singing, two heroes in one picture, fractional lengths, or other pictures.
The vendor’s input schema was read on the same date. See the model–technique applicability matrix for the connected operations and Choosing a model for how availability differs from a rendered outcome.
Reviewed 22 October 2026 against the current studio contract and dated studio research. Vendor documentation checked on that date. Recheck this page when the vendor changes the address, limits or output rules.