OmniHuman 2.0: Turn a Photo and Audio Into a Talking Video

OmniHuman 2.0 is ByteDance’s audio-driven avatar model. You give it one image and one audio file, and it produces a video where the person or character in the image speaks or sings in sync with the audio. The look comes from the image, the motion comes from the audio, and an optional short text prompt can steer the scene, movement and camera.

Updated October 2026.

Disclosure: VdoBloom is our platform.

Key takeaways

• Input is one photo plus one WAV or MP3 file between 4 seconds and 5 minutes long. The video length follows the audio.
• There is no aspect ratio setting. The output frame follows your photo, so crop the photo first.
• An optional prompt of up to 300 characters lets you direct the scene, motion and camera.
• On VdoBloom, OmniHuman 2.0 costs 5.6 credits per second of audio, about half the per-second price of OmniHuman 1.5 (10.8 credits).
• Only animate a real person’s face with their permission, and label the result as AI-generated.

What OmniHuman 2.0 does

OmniHuman 2.0 animates a still image using sound. The provider documentation describes it simply: the output video matches the appearance of the input image, and its motion and lip sync are driven by the audio.

That makes it different from a general image-to-video model. A general model invents motion from a text prompt. OmniHuman starts from a voice or a song and builds mouth shapes, head movement and body language to match it.

According to launch coverage, ByteDance and BytePlus released version 2.0 around 15 October 2026. Reported improvements include full-body gestures and camera movement that follow the rhythm of the audio, support for several people in one scene, and better handling of animals, anime and 3D characters. These points come from secondary reporting, so treat them as expectations rather than guarantees.

OmniHuman 2.0 specs at a glance

The table separates what the provider’s API documentation states from what has only been reported.

Spec	Value	Source
Inputs	1 image + 1 audio file, optional text prompt	Provider docs
Image formats	JPEG, PNG, WebP (other common formats also listed)	Provider docs
Image size	Under 30 MB, 300 to 6,000 px per side	Provider docs
Image aspect ratio	Width / height between 0.4 and 2.5	Provider docs
Audio formats	WAV or MP3, under 15 MB	Provider docs
Audio length	4 to 300 seconds	Provider docs
Prompt	Optional, up to 300 characters	Provider docs
Prompt languages	Chinese, English, Japanese, Korean, Spanish (Mexico), Indonesian	Provider docs
Output aspect ratio	Follows the image (no setting)	Provider docs
Output length	Follows the audio	Provider docs
Resolution	720p	Reported
Max length change	Raised from 60 seconds to 5 minutes	Reported
Multiple people, animals, anime, 3D	Supported	Reported

OmniHuman 2.0 vs 1.5: which one to pick

Both versions are available on VdoBloom. The quickest way to choose is by audio length: anything over 60 seconds needs 2.0.

	OmniHuman 1.5	OmniHuman 2.0
Max audio on VdoBloom	60 seconds	5 minutes
Price on VdoBloom	10.8 credits per second	5.6 credits per second
10-second clip	108 credits	56 credits
60-second clip	648 credits	336 credits
Prompt direction	Optional text prompt	Optional prompt up to 300 characters for scene, motion and camera

For most new projects, 2.0 is the sensible default: it handles longer audio and costs less per second on VdoBloom. Keep 1.5 in mind if you already have a short clip you liked from it and want a matching follow-up.

How much OmniHuman 2.0 costs on VdoBloom

You pay for the length of your audio at 5.6 credits per second. One credit is $0.05 at the Lite rate ($15 for 300 credits).

Audio length	Credits
10 seconds	56
30 seconds	168
60 seconds	336
2 minutes	672
5 minutes	1,680

Audio shorter than 4 seconds is rejected before any credits are taken. See pricing for credit packs.

How to make an OmniHuman 2.0 video on VdoBloom

There are two ways in, depending on whether you already have audio.

Option 1: Talking Avatar (your photo, your audio)

1. Open the AI Talking Avatar tool and choose OmniHuman 2.0.
2. Upload your prepared photo.
3. Upload a WAV or MP3 file between 4 seconds and 5 minutes.
4. Optionally add a prompt of up to 300 characters to direct the scene or camera.
5. Check the credit cost shown for your audio length, then generate.

Option 2: AI Presenter (type a script)

1. Open the AI Presenter tool.
2. Upload a photo and type your script.
3. A voice is generated from the script, then the avatar video is made with OmniHuman 2.0.

The presenter route suits explainers and announcements when you don’t have a recording. The talking avatar route suits songs, voice-overs you recorded yourself, and dialogue.

Photo prep checklist

The photo decides almost everything about the look and framing of your video, since there is no separate aspect ratio control.

• Frame for the motion you want. A tight headshot gives you a talking head. If you want hand gestures, include the hands and upper body in the frame.
• Keep the mouth visible. Avoid hands, microphones, masks or hair covering the lips.
• Face the camera or close to it. A clear, front-on or three-quarter view gives the model the most to work with.
• Check the aspect ratio. Width divided by height must be between 0.4 and 2.5. A 9:16 portrait (about 0.56) and a 16:9 landscape (about 1.78) both fit.
• Check the size. Each side must be 300 to 6,000 pixels, and the file under 30 MB.
• Use even lighting. Harsh shadows across the face can make mouth movement harder to read.
• Crop to the final frame. Whatever shape the photo is, that is the shape of your video.

Audio prep checklist

• Format: WAV or MP3 only.
• Length: 4 to 300 seconds. Remember that every second is billed.
• File size: under 15 MB. A long WAV can exceed this, so export longer clips as MP3.
• Clean voice: remove background music, echo and noise where you can. The model reads the audio to drive the mouth, so a clear voice helps.
• Trim silence: cut dead air at the start and end. It adds cost and gives you a frozen-looking face.
• Songs: a version with a strong, clear vocal works better than a dense mix.

OmniHuman 2.0 prompt templates

The prompt is optional. Without one, the model works only from the image and audio. With one, you can describe the setting, movement and camera. Each template below is under 300 characters. Swap the bracketed parts for your own details. These are starting points to try, not tested results.

Talking to camera

A friendly [presenter] speaks directly to the camera in a bright [home office], relaxed posture, natural blinking, small nods on key words, steady medium shot.

Singing

The [singer] performs the song with emotion, eyes closing on long notes, gentle head sway to the rhythm, warm stage lighting, soft background glow.

Two people

Two people sit side by side in a [podcast studio]. The person on the left speaks while the person on the right listens and reacts with small nods. Static wide shot.

Anime character

The anime character talks cheerfully with expressive eyebrows and lively eye movement, hair moving slightly in a light breeze, keep the original 2D art style.

Slow push-in

The [speaker] delivers a calm, serious message. The camera slowly pushes in from a medium shot to a close-up over the whole clip. Soft, even lighting.

Hand gestures

The [teacher] explains an idea with open hand gestures, counting points on fingers, leaning slightly forward, upper body and both hands visible in frame.

The two-person and anime templates rely on reported capabilities, so results may vary. Hand gestures only work if the hands are in the photo.

Limitations to know before you start

• One image only. You can’t supply a second reference angle or a separate background.
• No aspect ratio control. Change the frame by cropping the photo.
• Audio sets the length. You can’t ask for a 20-second clip from 60 seconds of audio. Trim the audio instead.
• Short prompts. 300 characters is enough for direction, not a full shot list.
• Listed prompt languages. The docs name six prompt languages. Other languages may not steer the scene as reliably.
• Resolution. Output is reported as 720p, which may need upscaling for large screens.
• Reported features are not guaranteed. Multi-person scenes and camera movement depend on the input and may not always match the prompt.

Consent and disclosure: using a real person’s face

Before you animate a photo of a real person, get their permission. A talking or singing video can make someone appear to say things they never said. That can harm them, and it may break laws on likeness, privacy or deceptive content where you live.

• Use your own photo, a photo of someone who has agreed, or an original character.
• Don’t put words in the mouth of a public figure or a private person without consent.
• Label the result as AI-generated when you share it, especially on social platforms that require it.
• Only use voice recordings and songs you have the right to use.

Where OmniHuman fits among other video models

OmniHuman is a specialist. If you need a scene that isn’t built around a speaking face, a general text-to-video model is usually a better fit. For a wider comparison, see the best AI video models of 2026 and our roundup of AI video generators.

FAQ

Is OmniHuman 2.0 free?

Not on VdoBloom. It costs 5.6 credits per second of audio, so a 10-second clip is 56 credits and a 60-second clip is 336 credits. One credit is $0.05 at the Lite rate ($15 for 300 credits). Audio under 4 seconds is rejected before any charge.

How long can an OmniHuman 2.0 video be?

The video is as long as your audio, which must be between 4 seconds and 5 minutes (300 seconds). The provider documentation sets the same 4 to 300 second range. The earlier 60-second limit was reportedly raised to 5 minutes with version 2.0.

What resolution does OmniHuman 2.0 output?

Output is reported to be 720p. The provider documentation used for this guide does not list a resolution setting, and the shape of the video follows the shape of your photo, so crop the photo to the frame you want before generating.

Do I need a prompt?

No. The prompt is optional. The model can work from the image and audio alone. Add a prompt of up to 300 characters when you want to direct the setting, the kind of movement, or the camera, such as a slow push-in or visible hand gestures.

Can a photo sing with OmniHuman 2.0?

Yes. Upload a song as WAV or MP3 and the motion follows the audio, including lip sync. A track with a clear, prominent vocal tends to give the model more to follow than a dense mix. Use only music you have the right to use.

Does it work with anime or cartoon characters?

Support for anime, 3D characters and animals is reported in launch coverage, not stated in the provider documentation used here. A clear front-facing character with a visible mouth gives the best chance. Ask in the prompt to keep the original art style.

Can I put two people in one video?

Multi-person scenes are a reported feature of 2.0. Use a photo that shows both people clearly, and describe in the prompt who speaks and who listens. Results may vary, since this capability comes from secondary reporting rather than the provider documentation.

Should I use OmniHuman 1.5 or 2.0?

For most projects, use 2.0. On VdoBloom it accepts audio up to 5 minutes, against 60 seconds for 1.5, and it costs 5.6 credits per second against 10.8. Choose 1.5 mainly if you want to match an earlier short clip made with it.