How to Use OmniHuman 2: From One Photo to a Talking Video

What you need before you start

OmniHuman 2 needs exactly two inputs and accepts one optional extra. Getting these right matters more than any setting, because the model builds the whole performance from them.

• An image: JPG, PNG or WEBP (other common formats work too), between 300 and 6000 pixels on each side, under 30 MB, with an aspect ratio between 0.4 and 2.5.
• An audio clip: MP3 or WAV, between 4 seconds and 5 minutes long, under 15 MB. The final video is exactly as long as your audio.
• Optional: a short text prompt describing gestures, mood, scene details or camera movement. English, Chinese, Japanese, Korean, Spanish and Indonesian prompts are supported.

How to use OmniHuman 2, step by step

Here is the full workflow we use for every video in our showcase.

1. 1Choose a clear photoPick an image where the face is sharp, evenly lit and not covered by hair, hands or a microphone. A half-body or full-body shot gives OmniHuman 2 room for hand gestures; a tight headshot keeps the focus on the face.
2. 2Prepare your audioUpload an MP3 or WAV or record straight from your browser. Dry speech without background music gives the tightest lip sync. For singing, a clean vocal track with the instrumental mixed low works best.
3. 3Write an optional promptDescribe what the person should do in one or two short sentences — for example “smiles, leans toward the camera on the last line, slow push-in”. Keep it about motion and mood, not about what they say; the audio already defines that.
4. 4Generate and watch progressRendering typically takes a few minutes depending on length. You can leave the page; the finished video appears in your library.
5. 5Download and shareDownload the MP4 without a watermark. Videos are rendered in 720p, which suits social feeds, ads and presentations.

Choosing the right photo for OmniHuman 2

The image sets the identity, framing, lighting and style of the final video, so it is worth thirty seconds of thought. OmniHuman 2 is not limited to real people: illustrations, 3D characters, paintings and even animals with a recognisable face can all be animated.

If you want several people talking in one scene — a podcast, an interview or a duet — use a single photo with everyone in frame and describe who is speaking in your prompt. Multi-person performances are one of the clearest upgrades over OmniHuman 1.5.

Writing prompts that OmniHuman 2 follows

Prompts work best when they are concrete and physical. Name the gesture, the emotion and the camera move, and avoid stacking more than three instructions in one sentence. If a result feels too busy, remove instructions rather than adding new ones.

Camera directions such as “slow push-in”, “static medium shot” or “gentle orbit” are understood, as are gestures like “counts on fingers”, “points at the camera” or “holds the product up”.

Tips for better lip sync

Most lip-sync problems come from the inputs rather than the model. Before re-rendering, check these first.

• Remove long silences at the start and end of the audio — the video mirrors them.
• Keep the mouth fully visible in the photo; beards are fine, hands over the mouth are not.
• Avoid heavy reverb or music under speech; it blurs the syllables the model locks onto.
• For long scripts, render a short 5–10 second test first, then the full take.