OmniHuman 2.0 generates 5-minute avatar video at $0.07 a second

ByteDance shipped OmniHuman 2.0 through its enterprise platform BytePlus on October 15, with five-minute single-shot avatar video at $0.07 a second.

BytePlus announced the launch in a post on X on October 15 and lists the model on its OmniHuman product page at $0.07 per second. The same page still carries $0.12 per second for OmniHuman 1.5, and BytePlus describes the new rate as more than 40% cheaper.

The headline change is what the model listens to. OmniHuman 2.0 takes a single image and an audio track and reads the tone, emotion and words in the audio to drive the face and body, rather than matching lip movements alone. BytePlus says it handles speech, singing and dialogue in one model. An optional text prompt steers specific actions, object generation and camera movement, and the product page lists multi-person scenes and non-human or animated characters.

Length is the other change: a single shot runs up to five minutes, up from 60 seconds in OmniHuman 1.5. The API documentation sets the audio input at 4 to 300 seconds, and the audio sets the clip length.

ByteDance has been here before. Its researchers showed OmniHuman-1 in February 2025 with clips of a fictional Taylor Swift performance and an Albert Einstein lecture, which TechCrunch called perhaps the most realistic deepfake videos to date. OmniHuman 1.5 followed and reached third-party clouds: fal.ai hosts it at $0.16 per second with 720p or 1080p output and up to 60 seconds of audio. OmniHuman 2.0 is a BytePlus API so far.

The documented specs are narrow. Output is native 720p. Images run to 30 MB and 6,000 pixels on a side. The prompt is optional and supports Chinese, English, Japanese, Korean, Spanish (Mexican) and Indonesian. The API runs in the ap-singapore-1 region and turns a watermark on by default.

Do the arithmetic and a full five-minute clip costs $21 at the listed rate, 300 seconds times $0.07. The same clip cost $36 at the old $0.12 rate, so each take is $15 cheaper, which adds up across several attempts at one performance.

The impersonation risk is unchanged. The audio-driven approach that makes a digital singer convincing is the one that made the first OmniHuman clips hard to tell from real footage. BytePlus markets 2.0 for advertising, music videos and digital characters, and the documentation shows only a default watermark and a content pre-filter standing between the API and a convincing fake of a real person.

OmniHuman 2.0 is available through the BytePlus API now; the docs give the endpoint and a request key of realman_avatar_picture_omni20_cv.