OmniHuman 2.0 by ByteDance: a talking video from a photo and a voice
OmniHuman 2.0 turns one portrait and an audio recording into a talking video of up to five minutes. What to feed it, three direction prompts and its limits.
Some talking-avatar models come with a panel full of switches. OmniHuman 2.0 from ByteDance has three inputs and nothing else: a portrait, an audio recording and an optional line of direction. No resolution menu, no aspect ratio, no expressiveness slider. That sounds limiting until you notice what it means in practice: almost all of the result is decided by the two files you hand it, so the work is in choosing them well. Here is what I would tell a friend before their first run.
What it is actually good at
Long, natural talking. The audio can run up to five minutes (only the first 300 seconds are used), which is far more than most clip models allow, so a full product explainer, a short lesson or a character monologue fits in one render instead of five stitched pieces. The face follows the voice: pauses, emphasis and the rhythm of a sentence show up in the mouth, the eyes and the head. It is a good pick when you have a finished voice track, from your own microphone or from a text-to-speech model, and want a face to carry it.
Choosing the portrait
The image is converted to JPEG and fitted to the sizes the model supports, and a very wide or very tall picture is cropped from the center. So frame it yourself: a head and shoulders shot, the face roughly toward the camera, mouth visible, nothing in front of the lips. A portrait with room above the head survives the crop better than one where the hair touches the edge. Generated characters work as well as photos; if the skin looks waxy, fix that first with the plastic skin guide, because animation makes it more obvious, not less.
Choosing the audio
- Clean speech beats clever speech. Background music and room echo give the mouth something to react to that is not a word.
- Under four seconds is padded with silence, so a two-second line becomes a clip where the face waits at the end. Either record a full sentence or accept the pause.
- Trim dead air at the start. A long silent intro is a long stare into the camera.
- Anything past five minutes is cut off. Split a longer script into chapters and render each one.
Three directions worth stealing
The prompt guides motion and style. Leave it empty and the voice drives everything; fill it in and you can set a mood. Write it like a note to an actor, not like an image prompt.
Calm and friendly, like explaining something to a colleague across the desk. Small nods on key words, natural blinks, eyes mostly on the camera, a brief smile at the end.An old storyteller by a fire, speaking slowly. The head tilts on questions, eyes drift away while remembering, then come back to the listener. Warm, tired, no sudden moves.Energetic game announcer: raised eyebrows, quick head movements on every exclamation, a grin before the final line.Who should pick something else
If you want to control the output shape or pick a resolution, OmniHuman gives you no handle for it: the frame follows your picture. HeyGen Avatar IV has an expressiveness setting, a resolution choice and an aspect ratio, and its guide walks through them. If you have no portrait at all and want a ready-made presenter, the Avatar V guide is the place to start. The rest of the talking-avatar and lipsync models are on the models page.
The honest limitation
This is a talking portrait, not a scene. The person will not walk across a room, pick something up or change outfits, and the prompt cannot ask for that. It also cannot rescue a bad input: a side profile, a hand over the mouth or a noisy recording will show in the result. And a face you have no right to use stays off limits: the terms forbid deepfakes of a real, identifiable person. Use your own face, a fictional character, or a portrait you have clear permission for.
Questions
What does OmniHuman 2.0 need?
A portrait image and an audio recording. A text prompt for motion and style is optional.
How long can the audio be?
Up to 300 seconds are used; anything longer is cut. Recordings shorter than four seconds are padded with silence.
Can I choose the resolution or aspect ratio?
No. The model fits your image to the sizes it supports and crops very wide or very tall pictures from the center, so frame the portrait before you upload it.
