OmniHuman 2.0 guide

OmniHuman 2.0 starts from a single photo and builds the whole performance from your audio. It listens for how each line is said, not only which sounds are in it, so the face carries the mood of the track. Turn on sound to hear it.

Three changes set 2.0 apart, according to BytePlus's launch post.

You give it two things, and a third is optional. Most videos need only the first two.

The photo decides more of the result than anything else. The model keeps what it sees, so a sharp, well-lit face gives you a sharp, well-lit performance.

Clay characters and illustrations work too. Keep the face clearly drawn, with the eyes and mouth easy to read.

The audio is the performance, so how the line is delivered is what you get on screen. Generate or record the voice first, and listen to it before you run the video.

The same model reads very different deliveries. A whisper, a shout and a laugh each come out as a different performance.

A single shot can run up to 5 minutes. Before 2.0, anything over 60 seconds had to be generated in pieces and joined, and the face, light or background could jump at each join. A long take avoids that, as long as the photo and the audio hold up for the full length.

1. Write the full script and generate the voice in one pass, so the tone stays the same from start to finish.
2. Listen through once and fix any mispronounced word before you generate the video. A fix later means running the whole take again.
3. Pick a photo with a calm, neutral expression. It has to work for every moment of the track, not just the opening line.
4. Run a 20-second test from the middle of the script first, then the full take.

The audio carries most of the performance. A short direction line helps when you want a particular kind of movement. Describe the delivery, not the words.

Some early viewers of BytePlus's launch demo noticed the background shifting while the person spoke. Audio-driven models generate every pixel of every frame, not just the face, so anything in the photo can drift: people behind the speaker, shelves, signs, patterns. The more there is to draw, the more there is to move.

We tested this on Morphic with the same face, the same voice line and two backgrounds. In the busy café version, the people behind her turn their heads and move on their own, and the background changes about 2.3 times more from frame to frame than in the plain-wall version.

Five steps keep the background still:

1. Use a plain or soft-focus background. If you only have a busy photo, edit the background into a plain wall first, keeping the person the same. That is how the plain version above was made.
2. Keep one face in the frame. Crop out or remove anyone else, even blurred, even at the edge.
3. Avoid fine detail behind the person: text, tiles, shelves of small objects, string lights.
4. Add "background stays still" to the direction line.
5. If one stretch of a long take still drifts, run that stretch again on its own and cut it in, at a pause in the speech where a cut is least noticeable.

Our test used Veed Fabric on Morphic, because OmniHuman 2.0 is not on Morphic yet. The cause is the same for any model that draws the whole frame from one photo, so the same fixes apply.

Both turn one photo and one audio track into a talking video. Veed Fabric is live on Morphic today, and every example in this guide was made with it. OmniHuman 2.0 is coming soon.

1. Open Copilot and attach a photo with one clear face and a plain background.
2. Attach your voice or song track, or ask Copilot to generate the voice from your script first.
3. Ask for a talking video with Veed Fabric, add a short direction line, and download the result or open it in the editor.