Speaking Portraits
Make a portrait, speech track, and lip-synced result without inventing a 3D avatar.
Secure both likeness and voice
A speaking portrait can be convincing even when the source is only a still image. Obtain permission for the person’s likeness and the speech recording; avoid presenting synthetic speech as a real statement.
- Write a short approved script and confirm the subject agrees to this specific use, audience, and distribution.
- Use a clear frontal or three-quarter portrait with visible mouth; keep an unaltered copy of the authorized source.
Prepare clean narration before animation
Speech timing largely determines the pacing of a lip-synced portrait. Repeated takes with different pauses are easier to evaluate as separate Audio results than as competing references on one node.
- Record or synthesize a short line, listen for clipping and awkward pauses, and leave a little silence at either end.
- Keep one final audio take and an Image or Reference node for the approved portrait.
Wire OmniHuman 1.5 precisely
The Volcengine OmniHuman 1.5 route requires exactly one image and one audio input; it exposes no video reference port. Its ratio follows the source image and its prompt has a 300-character limit.
- Place Video Generation, choose OmniHuman 1.5 in Model, connect portrait to Image and narration to Audio, then select exactly those two @ items.
- Keep audio under 60 seconds; 15 seconds or less is the canvas recommendation. Choose 720p Fast or 1080p in the resolution control if offered.
Understand Digital Human Video
Digital Human Video is another image-plus-audio canvas preset routed through HKGAI Seedance 2.0, not an alias for OmniHuman. Its supplied prompt asks for a fixed camera and restrained expressions.
- For a separate comparison, duplicate the portrait and audio setup and switch the copy to Digital Human Video.
- Do not change both the voice and the model at once; compare lip timing, facial drift, and background stability shot by shot.
Review before publishing
A generated talking head is a rendered video, not an independently rigged avatar. Check the mouth in close-up and the portrait edges at normal playback speed before combining with other clips.
- Compare the spoken words with the approved script, then watch for facial stretching during consonants and unnatural pauses.
- If extra references trigger an exact-input error, deselect them; if the audio stumbles, fix the source narration rather than adding more visual prompts.