Omnihuman 1.5

Omnihuman 1.5 is a video generation model by ByteDance designed to create realistic, lip-synced digital avatars from a single portrait image and an audio track. It excels at generating expressive performances where the character's facial expressions, body gestures, and head movements match the rhythm and emotional tone of the speech. It is highly effective for creating virtual presenters, personalized video messages, and cinematic talking-head videos, and it can even animate stylized non-human characters like anime figures or pets.

It utilizes a "cognitive simulation" architecture inspired by human psychology. Instead of mechanically matching lips to audio waveforms, a multimodal large language model analyzes the semantic meaning of the audio to plan appropriate emotional reactions and gestures. Then, a diffusion transformer renders the physical movements. This dual-system approach allows the avatar to appear as if it is thinking and reacting naturally to the context of the speech, reducing the stiff, robotic feel common in older avatar models.

Omnihuman 1.5 was developed by ByteDance's Intelligent Creation team. The model's research paper and core architecture were unveiled on August 26, 2025. It serves as a major architectural upgrade over the original OmniHuman-1, adding text-prompt guidance, unconstrained camera movement, and cognitive reasoning. ByteDance is also the developer behind other notable generative models, including the image generator Dreamina 3.1 and the multimodal video model Seedance 2.0.

According to the official BytePlus documentation, Omnihuman 1.5 cannot use a single audio file to drive a back-and-forth conversation between multiple characters simultaneously. To achieve this, you must use subject detection to isolate each character with a mask, generate individual speaking clips for each person using their specific audio segments, and then stitch the resulting videos together in a video editor.

Yes. A key upgrade in version 1.5 is the addition of text prompt support. You can provide a text prompt alongside your image and audio to explicitly direct the character's emotional state, body language, and even camera movements, such as zooms or pans. This gives you directorial control over the final performance rather than relying entirely on the AI's automatic interpretation of the audio track.