OmniHuman: Create talking avatars with precise lip sync

ByteDance's OmniHuman turns one image and one audio track into a talking or singing avatar video. OmniHuman 2.0 runs up to 5 minutes and takes an optional prompt for gestures, camera moves and scene details. OmniHuman 1.5 is the quick pick for clips up to 30 seconds at up to 1080p.

Try OmniHuman 2.0
Omnihuman AI
A face emerging from torn paper fragments, illustrating OmniHuman 2.0 avatar generation

What is OmniHuman 2.0?

OmniHuman 2.0 is ByteDance's audio-driven avatar model, available in Artlist's AI tools. Upload one character image and a voice or music track, and it generates a talking or singing video with lip sync, expression and gestures that follow the audio. Clips run from 4 seconds to 5 minutes at 720p, and an optional prompt directs camera and gestures.

Try OmniHuman 2.0

What you can make with OmniHuman 2.0

One image, one audio file, and a prompt if you want to direct it. Here's where the 5-minute ceiling and prompt control change what you can ship.

  • Spokesperson videos that run the full script

    Record a 3- or 4-minute explainer, product walkthrough or course lesson and generate it in one pass. No cutting the script into 30-second pieces and stitching the clips back together.

  • Music performances

    Upload the vocal mix and the character performs it, with mouth, face and gestures following the rhythm and emphasis of the track. Prompt a turn to camera on the chorus or a lean-in for the bridge.

  • UGC and product content with a consistent face

    Reuse the same character image across ads, testimonials and social cuts. Swap the audio and the same face delivers the new script, so your spokesperson looks the same on every placement.

  • Directed talking heads

    Leave the prompt empty for a straight talking head, or add camera moves and gestures when a static frame isn't enough. Either way, the audio drives the performance.

OmniHuman 2.0 vs OmniHuman 1.5 vs HeyGen Avatar 4

Feature
OmniHuman 2.0
OmniHuman 1.5
HeyGen Avatar 4
What it does
Image + audio to a talking or singing avatar, with optional prompt direction
Image + audio to a talking avatar, with optional prompt
Image + audio to a talking avatar
Max clip length
Up to 5 minutes, set by the audio
Up to 30 seconds
Up to 3 minutes
Resolution
720p, shaped to the uploaded image
720p or 1080p
Up to 1080p
Framing
Follows your image (a square image gives 960×960)
Follows your image
16:9 only
Text prompt
Gestures, camera moves, scene details
Scene context, mood, framing
Inputs
1 image (JPG, PNG, WebP) + 1 audio (WAV, MP3)
1 image + 1 audio
1 image + 1 audio
Best for
Long-form spokesperson videos, songs and directed performances
Short talking heads and UGC clips at up to 1080p
Reusable 16:9 avatars up to 3 minutes

Frequently asked questions

OmniHuman 2.0 is ByteDance's audio-driven avatar model. You upload one character image and one audio file, and it generates a video of that character speaking or singing the audio, with lip sync, facial expression and gestures that follow the recording. In Artlist it runs from 4 seconds up to 5 minutes, with an optional text prompt for direction.

Length and direction. OmniHuman 2.0 generates clips up to 5 minutes and takes an optional prompt for gestures, camera moves and scene details. OmniHuman 1.5 stops at 30 seconds but outputs at up to 1080p, against 720p for 2.0. Use 1.5 for short, high-resolution talking heads and 2.0 when the script is long or the performance needs directing.

Between 4 seconds and 5 minutes. The clip is exactly as long as the audio you upload, so there is no duration setting to adjust. Audio shorter than 4 seconds won't generate, and anything longer than 5 minutes, like a full song or a long lesson, needs to be split into separate generations.

No. For a straight talking head, leave the prompt empty and the audio drives the whole performance. Use the prompt for what the audio can't say: "slow push-in", "turns to camera on the chorus", "leans in and gestures with both hands". Prompts can be written in English, Chinese, Japanese, Korean, Spanish (Mexico) or Indonesian.

One image in JPG, PNG or WebP and one audio file in WAV or MP3, plus the optional text prompt. It doesn't take reference images, start or end frames, or video input. Start from a clear image with the face fully visible, and use clean audio without long silences or background noise for the tightest lip sync.

Yes. Upload the vocal mix you want the character to perform, not the instrumental, and the mouth, face and gestures follow the rhythm and emphasis of the track. Because clips run up to 5 minutes, most songs fit in a single generation, and a prompt can add camera moves on the chorus or bridge.

Not in OmniHuman 2.0. Output is 720p and the framing comes from the image you upload, so a square image produces a 960×960 video and a portrait image stays portrait. If you need 1080p, OmniHuman 1.5 outputs at 720p or 1080p for clips up to 30 seconds. HeyGen Avatar 4 goes up to 1080p but only in 16:9.

Still have questions? We're here to help.