ByteDance OmniHuman 2.0: Pricing, Limits, and Deepfake Risk Explained
ByteDance's OmniHuman 2.0 turns a photo and audio clip into a 5-minute talking video at 7 cents a second. Here's how it works and its limits.
What is OmniHuman 2.0?
OmniHuman 2.0 is ByteDance’s AI model for generating talking or singing avatar videos from a single photo and an audio track. Feed it one image and a voice clip, and it outputs a video of that person (or character) speaking, singing, or performing the audio in a continuous shot. It’s available now through BytePlus and its API, priced per second of generated video rather than per project.
TL;DR
- OmniHuman 2.0 generates a video of a person or character talking or singing from just one still image plus an audio file.
- The model now supports clips up to five minutes long in a single continuous shot, up from the 60-second cap on the previous version.
- Output resolution is capped at 720p, with no 1080p or 4K option, and some early demos show background elements shifting oddly during longer clips.
- Pricing dropped to 7 cents per second, down from 12 cents on the prior version, which puts a full five-minute video at roughly 21 dollars.
- The model isn’t limited to humans; it also works on cartoon characters and animals, and you can add text prompts to control camera moves or introduce elements not in the original photo.
- Turning one photo into a full video of someone speaking raises obvious deepfake concerns, even though current lip-sync quality still has visible tells.
- It joins a wider field of image-to-video and lip-sync tools, where the main competitive axes right now are clip length, price per second, and how convincing the mouth movement actually looks.
How does OmniHuman 2.0 actually work?
The pipeline is simple to describe even if the model underneath isn’t: you supply a single reference image and an audio track, and OmniHuman 2.0 animates that image to match the audio. It handles speech and singing, syncing mouth movement, expression, and some body motion to the performance. You can layer in text prompts on top of that to request specific camera movement or add visual elements that weren’t present in the source photo, giving you a bit of directorial control beyond just “make this person talk.”
The subject doesn’t have to be a realistic human photo. It also works with cartoon characters and animal images, which widens its use case beyond talking-head content into stylized or animated short-form video.
What changed from the first version of OmniHuman?
The headline upgrade is duration. The original OmniHuman topped out at 60 seconds of generated video, which forced anyone making longer content to generate multiple short clips and stitch them together, hoping the cuts weren’t obvious. OmniHuman 2.0 extends that ceiling to five minutes in one continuous generation, removing the need to hide seams between clips.
Pricing also moved in the user’s favor. The previous model charged 12 cents per second of output; OmniHuman 2.0 drops that to 7 cents per second. That’s a meaningful cut, though longer clips still add up: a full five-minute video runs about 21 dollars at that rate, since cost scales directly with duration rather than dropping at volume.
How good is the output quality?
Resolution is capped at 720p. There’s no option for 1080p or 4K output, which matters if you’re planning to use generated footage in any context where upscaling artifacts would be obvious, like a larger screen or a professional production pipeline.
Beyond resolution, the bigger open question is lip-sync realism. ByteDance’s own demo reel shows clean results, but some early viewers flagged background elements shifting or drifting during clips, an artifact common to AI video generation when a scene has to stay consistent over a longer duration. The core lip-sync technology, while improved over earlier generations of this type of tool, still tends to show tells on close inspection, things like mouth shapes that don’t quite match certain phonemes or expressions that read as slightly off. The technology has clearly advanced compared to where AI lip-syncing was a couple of years ago, but “good enough to fool a careful viewer” and “good enough to not notice in a scroll” are different bars, and OmniHuman 2.0 sits closer to the second.
Is OmniHuman 2.0 worth it compared to other options?
For a specific use case, animating a single photo into an extended talking or singing performance, it’s hard to find a direct substitute. Most lip-sync and talking-avatar tools still cap out at much shorter durations, and stitching multiple generations together to simulate a longer video introduces visible seams. The jump to five minutes in one shot is a real capability gain, not just a marketing number.
Nothing gets lost.
Whether it’s “worth it” depends on what you’re building. At 7 cents per second, short clips are cheap (a 15-second clip costs about a dollar), but longer-form content scales linearly, so a five-minute video at roughly 21 dollars isn’t trivial if you’re producing at volume. And since output is locked to 720p, anyone needing higher-resolution footage for broadcast or large-format use will need an upscaling step afterward, adding time and cost.
For social content, avatar-based explainer videos, or prototyping animated dialogue, the combination of long duration and low per-second cost is competitive. For anything requiring pixel-perfect realism or higher resolution, it’s not there yet.
What are the deepfake risks with a tool like this?
The core mechanic, one photo plus an audio track producing a full video of that person speaking, is exactly the recipe that makes deepfake detection and misuse a live concern. Unlike tools that require a video reference or multiple images to build a likeness, OmniHuman 2.0’s low input bar (a single still photo) lowers the effort needed to generate a fabricated video of someone saying or singing something they never said.
That said, current output quality works against convincing misuse right now. The lip-sync artifacts and occasional background drift visible even in ByteDance’s own curated demos suggest that a careful viewer, or any basic forensic review, would likely catch an OmniHuman-generated fake today. The risk isn’t that the current model is already indistinguishable from real footage. It’s that the gap closes with each iteration, and the input requirements (one photo, one audio clip) are trivial to obtain for almost anyone with a public presence. Tools in this category are worth watching less for what they can do today and more for the trajectory they’re on.
FAQ
How long can a video be with OmniHuman 2.0?
Up to five minutes in a single continuous shot, a major jump from the 60-second limit on the previous OmniHuman version.
How much does OmniHuman 2.0 cost?
It’s priced at 7 cents per second of generated video, down from 12 cents on the prior model. A full five-minute video costs roughly 21 dollars.
What resolution does OmniHuman 2.0 output?
720p. There’s no 1080p or 4K option, so higher-resolution use cases would require a separate upscaling step.
Can OmniHuman 2.0 animate anything besides real people?
Yes. It works with cartoon characters and animal images in addition to realistic human photos, and supports text prompts for camera control or added visual elements.
Where can I access OmniHuman 2.0?
It’s available now through ByteDance’s BytePlus platform and its API.
