BackGo to OmniHuman 2.0
ByteDance · Photo + audio → video

OmniHuman 2.0

Bring a photo to life with any sound. Lips, expressions, gestures and camera follow the audio — speech or a song up to 5 minutes.

Example: a dialogue of two people

One photo and a dialogue voice-over: the prompt says who speaks and who listens silently meanwhile. A 13-second clip, sound on.

Russian voice-over. Prompt: Static camera and background. Woman on left speaks, man on right listens silently; man speaks, woman listens silently; woman speaks, man listens silently; finally, man speaks with a smirk, woman listens silently.
English voice-over. Prompt: Static camera, both stay in frame. First the woman frame-right speaks curiously while the man frame-left listens silently. Then he speaks proudly while she listens silently. Then she speaks suspiciously while he listens silently. Finally he grins, speaks and laughs while she listens silently.
up to 5 min
Duration
Body and camera
Not just the face
from ···
credits
720p
Quality

How it works

1. Upload a photo

A person, an animal or a drawn character with a sharp, large face. This is who comes to life.

2. Upload audio

Speech, a song or a voice-over from 4 seconds to 5 minutes. You can voice a text right here.

3. Get the video

Lips, expressions and gestures follow the sound. Optionally add a prompt: camera, actions, events in the frame.

No recording yet? Voice a text with a neural network — in your own voice or a voice from a description. More about text-to-speech →

Features

Not just a talking head: body movement, gestures, reactions
Singing and rhythmic performances, including high parts
Camera moves and actions in order are set by a short prompt
Several characters in the frame: the prompt picks who speaks
Animals, toys and drawn characters speak and sing too
Up to 5 minutes, 720p quality

How to write a prompt

A prompt is optional: the model understands the meaning of the audio. It is for the camera, actions and events in the frame.

Say how the character sounds: "talks to the camera", "sings", "whispers". Without it the lips come out weaker.
Describe only what changes. The model takes appearance and surroundings from the photo.
Put actions in order: "first… then… finally".
For a camera move, say where it ends: "pushes in and stops on the face".
The speaker should not leave the frame and come back: the face may change.
A dialogue of two people: for every line, write who speaks and who listens silently, naming them by the side of the frame ("the woman on the left", "the man on the right"). Keep the camera static.

Pricing

··· cr/sec

Price = rate × audio duration, seconds are rounded up

4 sec
··· cr
15 sec
··· cr
1 min
··· cr
5 min
··· cr

FAQ

What is OmniHuman 2.0?

It is a ByteDance model: from one photo and one audio file it makes a video where the character speaks or sings, gestures and moves in time with the sound. The video is as long as the audio.

How is OmniHuman 2.0 different from Kling Avatar?

Kling Avatar is an inexpensive talking head in 720p or 1080p. OmniHuman 2.0 does more: body movement and gestures, camera moves, singing, several characters in the frame, animals and drawn characters. The resolution is 720p and a second costs more.

What audio can I upload?

Speech, a song or a voice-over from 4 seconds to 5 minutes. MP3 and WAV work as they are; other formats (M4A, OGG, FLAC, AAC) and files that are too heavy are re-encoded for you.

Do I need a prompt?

No. The model understands the meaning of the audio and acts accordingly. A prompt (up to 300 characters) is for camera moves, actions and events in the frame. It must be in English, Chinese, Japanese, Korean, Spanish or Indonesian: Russian is not supported, so the form has a translate button. If the result with your prompt is worse, try without it.

How much does it cost?

Per-second pricing: rate × audio duration, seconds are rounded up. Credits are charged at launch and returned if generation fails.

How long should the videos be?

Up to 5 minutes is allowed, but the structure holds best up to about 15 seconds; after that it can drift. For a long recording it is safer to make several short videos from the same photo.

Which photo works?

A sharp photo with a large face, preferably frontal or three-quarter. Sides from 300 to 6000 pixels, aspect ratio from 2:5 to 5:2. If the face is small in the frame, the mouth may barely move: crop the photo closer.

Bring a photo to life