Give OmniHuman 1.5 one portrait and a voice track, and it returns a video of that person saying your exact words — lips, expressions and gestures follow the sound.
OmniHuman 1.5 is ByteDance's audio-driven avatar model: give it one portrait and a voice track, and it returns a video of that person speaking your audio — lips, expressions, head motion and gestures all driven by the sound. Unlike prompt-driven video models, you supply the exact words. The research paper pairs a multimodal language model with a diffusion transformer, so the performance reacts to what is being said rather than just tracking the mouth.
The honest part: it needs a frontal portrait. Our three-quarter-profile test below shows lips that barely articulate and a head that drifts away. Audio is capped at 35 seconds per run, so longer scripts mean splitting and stitching, and multi-person dialogue — a headline feature of the paper — is not available here. It is speech video, not scene video: for camera movement or action, use Seedance 2.5 or Kling 3.0 Turbo.
Architecture, positioning and user-study numbers are from the ByteDance research paper: OmniHuman-1.5: Instilling an Active Mind in Avatars via Cognitive Simulation (arXiv:2508.19209)
“Version 1.5 demonstrates improved robustness for challenging inputs.”
“A waist-up portrait with clear hands, shoulders, and natural lighting usually gives better body movement.”
“Heavy shadows, sunglasses, strange occlusions, and tiny faces make the job harder.”
The model is ByteDance's; the reason to use it here is that the other two inputs — the face and the voice — come from the same workspace.
A talking-avatar clip needs three things: a face, a voice, and the model. Here all three come from the same workspace — generate the portrait with GPT Image 2 or Seedream, type the script and synthesize it with one of 60 built-in voices, then feed both straight into OmniHuman 1.5. Every clip on this page was made exactly that way, end to end, without recording a second of audio or uploading a single photo of a real person.
The text mode is not an afterthought: 60 voices across English, Chinese, Japanese, Korean, French, German, Spanish, Italian, Russian and Portuguese, each with a preview you can play before spending anything. Type up to 600 characters, pick a voice, and the platform synthesizes the speech and drives the avatar with it in one run. The Mandarin tea-shop clip above is this exact path — no microphone involved.
OmniHuman 1.5 bills by the second of audio — 130 credits per second, rounded up — and the workspace shows the exact cost of your clip before you press generate. The same plan covers GPT Image 2, Nano Banana Pro, Seedream, Seedance, Kling and Veo, so the sensible workflow is cheap: draft the portrait for a handful of credits, lock the script, and only then spend on avatar seconds. Trim the silence off your audio — you are paying for every second of it.
Our seven-clip test set was generated on August 19, 2026 — a product pitch, a news brief, a gym pep talk, a Mandarin tea-shop welcome, a storybook character, a deliberate three-quarter-profile failure, and a wide stage shot. Every accepted job rendered successfully on its first take, in 3m 20s to 4m 42s for around ten seconds of audio. We also publish what went wrong: on a busy Tuesday evening the upstream rejected 13 of our 20 submissions with capacity errors before accepting them (rejected runs are refunded automatically), one accepted job queued for 14 minutes, and the profile clip on this page shows exactly what happens when you ignore the frontal-portrait advice.
All 7 clips below were generated here on August 19, 2026 — each one is the first render its inputs produced, no retakes and no cherry-picking. Portraits are GPT Image 2 outputs; every voice was typed as text and synthesized with the built-in library. Accepted jobs returned in 3m 20s to 4m 42s for roughly 10 seconds of audio (one outlier queued 14 minutes); on a busy evening the upstream rejected several submissions with capacity errors before accepting — rejected runs are refunded. Sound on.
Input: GPT Image 2 portrait of a creator holding a portable speaker + the built-in Trustworthy Man voice. Script: "This little gadget honestly surprised me. One charge lasts the whole week, it pairs instantly, and the sound is way bigger than the size suggests. I've been using it every single day."
Input: GPT Image 2 portrait of an anchor at a news desk + the built-in News Anchor voice. Script: "Good evening, and welcome to the briefing. Tonight we look at how small studios are using digital presenters to publish in ten languages at once — without booking a single camera crew."
Input: GPT Image 2 waist-up photo of a coach mid-gesture + the built-in Energetic Guy voice. Script: "Alright team, listen up! Thirty seconds, that's all I'm asking for. Three rounds, full effort, no excuses. When that timer starts, you give me everything you've got. Let's go!"
Input: GPT Image 2 portrait of a tea-shop owner + the built-in Mandarin Sweet Girl voice. Script: "大家好，欢迎来到我们的小店。这一款云雾绿茶是今年春天刚采的，用八十度的水慢慢泡，入口特别清甜。喜欢的话，记得点个关注哦。"
Input: GPT Image 2 storybook illustration of a character called Pip + the built-in Sweet Girl voice. Script: "Hi there! I'm Pip, and I live inside a storybook. Every page you turn is a brand new adventure — so, shall we see what's on the next one?"
Stress test — input deliberately breaks the guidance: a three-quarter profile turned well away from the camera, + the built-in Wise Lady voice. Script: "People ask why I still write my letters by hand. Because slowness is not a flaw. Some things are only worth saying at the speed of ink."
Input: GPT Image 2 wide 16:9 shot — a small spotlit figure on a dark stage, testing how the model handles a face that is far from the camera. + the built-in Movie Narrator voice. Script: "Every story begins in the dark, long before the first spotlight finds it."
Each of these maps to a clip demonstrated above — not a wish list.
Type the narration and get a presenter on screen — no camera, lighting or reshoots. The 35-second cap per clip maps neatly onto one concept per clip, which is how most course editors cut anyway.
The speaker-pitch case above is the whole workflow: generate a creator-style portrait holding your product, type the pitch, pick a voice. No talent fees, and no likeness release to chase — the face never existed.
The same face can deliver the same message in ten languages: keep the portrait, swap the voice. Lip-sync follows whatever audio you provide — the Mandarin case above is the proof.
A persona you generate is a persona you own. Design the face once, give it a consistent library voice, and it can front every episode — the illustrated storybook character above shows it does not even have to be photoreal.
You already have the hard part: the audio. Trim a highlight to 35 seconds, add a portrait, and you have a visual clip for social without opening an editor.
Speech video versus scene video — what each model needs from you, and what it hands back in this workspace.
| OmniHuman 1.5 | Seedance 2.5 | Veo 3.1 Pro | Kling 3.0 Turbo | |
|---|---|---|---|---|
| What you provide | 1 portrait photo + your audio, or typed text and a library voice | A prompt, plus optional image, video and audio references | A prompt, plus an optional reference image | A prompt or a start image |
| Who speaks the words | You do — the avatar lip-syncs to your exact soundtrack | The model generates voices from the prompt | The model generates voices from the prompt | Audio is generated with the clip |
| Clip length here | Length of your audio, up to 35 s | 4–30 s | 4–8 s | 3–15 s |
| Credits | 130 per audio second | 110/s at 480p · 230/s at 720p | 320–1,920 per clip (4–8 s · 720p–4K · ± audio) | 55/s at 720p · 70/s at 1080p |
| Measured here | ~4 min for ~10 s of audio (7 clips, August 2026) | Not measured on this task | Not measured on this task | Not measured on this task |
Credits, duration and input options are read from this workspace's own model registry on August 19, 2026 — they describe what you get here, not the vendors' published specifications. Timings are our own measurements on the clips shown above; we have not run the same scripts through the other three models, so those cells say so rather than guessing.
Compare OmniHuman 1.5 with other video models — same workspace, one click to switch.
| Model | omnihuman-1.5, by ByteDance (research paper published August 2025) |
|---|---|
| Mode | Talking avatar: one photo + one audio track → speaking video, in the avatar workspace on this page |
| Inputs | Exactly 1 portrait image, plus audio up to 35 seconds (mp3/wav upload) — or typed text up to 600 characters, synthesized to speech |
| Voice library | 60 built-in voices across 10 languages: English, Chinese, Japanese, Korean, French, German, Spanish, Italian, Russian, Portuguese |
| Optional guidance | A short performance note for actions, emotion and camera — useful, never required |
| Output | One continuous shot that keeps your photo's framing; clip length equals audio length. Measured output: 1248×1664 from a 2:3 photo, 1920×1088 from 16:9, 25 fps, H.264 with mono AAC audio |
| Credits | 130 per second of audio, rounded up — 130 minimum, 4,550 at the full 35 seconds; the exact cost is shown before you generate |
| Measured speed | 3m 20s–4m 42s per accepted job for ~10 s of audio in our 7-clip test; one outlier queued 14 minutes (August 2026, a busy Tuesday evening) |
Upload a clear, front-facing, waist-up portrait — or generate one with GPT Image 2 or Seedream in this same workspace, which is how every face on this page was made. One person per photo; keep the face large and unobstructed.
Upload an audio clip (mp3 or wav, up to 35 seconds), or switch to text mode: type up to 600 characters and pick one of 60 voices in 10 languages, each with a preview you can play first.
A short note like "calm, small nods, subtle smile" steers gesture and emotion. Leave it empty and the model reads the rhythm and meaning of the audio on its own.
The exact credit cost appears before you run — it scales with audio length. Generation is asynchronous; the clip lands in your history with every input attached, so a working setup is reproducible next week.
One account, the full image and video generation workflow.
Where OmniHuman sits among Seedance, Veo, Kling and Hailuo — and when a talking head beats b-roll.
How to make the input portrait — a clean, front-facing face is half of a good avatar clip.
If you need motion rather than speech, these cost a fraction per second.
Stylized characters can talk too — see the illustrated storybook case above for what to expect.
Run OmniHuman 1.5 on the free tier's daily credits, or upgrade for premium models, faster generation, and 4K output.
Perfect for getting started
Billed annually · $240/year
Billed annually · $380/year
ByteDance's audio-driven avatar model. It takes one portrait photo and one audio track and returns a video of that person speaking the audio — lips, expressions, head motion and hand gestures all follow the sound. The research paper behind it was published in August 2025.
Yes — that is the difference from prompt-driven video models. You supply the soundtrack, either as an mp3/wav upload or by typing the script and picking a built-in voice, and the avatar lip-syncs to it word for word. Prompt-driven models generate their own voices and rarely say precisely what you wrote.
No — it bills at 130 credits per second of audio, rounded up, so a 10-second clip is 1,300 credits. The exact cost is shown before you generate, and failed runs are refunded automatically. Keep scripts tight and trim silence: every second of audio is a second you pay for.
Front-facing, waist-up, one person, with the face large, evenly lit and unobstructed — reviewers and our own tests agree. Our deliberate three-quarter-profile test above shows what happens otherwise: the lips barely articulate, the model slowly rotates the head further away, and objects appear on the desk that were never in the photo. Generated portraits work exactly as well as camera photos.
Yes — our storybook test above is a flat 2D illustration, and OmniHuman 1.5 animated it convincingly while keeping the art style: eyes, brows and mouth move like a hand-animated character. One quirk we observed: when the mouth opens wide, the model draws surprisingly realistic teeth for a flat illustration. ByteDance's paper also claims non-human subjects; we have only tested this illustrated character.
Up to 35 seconds of audio per clip on this platform — longer scripts need to be split and stitched. In our August 2026 test, accepted jobs returned in 3m 20s to 4m 42s for roughly 10 seconds of audio. One caveat from the same evening: under heavy load the upstream rejected several submissions with capacity errors before accepting the job. Rejected runs cost nothing, but plan for retries at peak hours rather than a deadline five minutes away.
Non-frontal faces, above all: our three-quarter-profile test shows near-static lips, a head that drifts further from the camera, and desk objects that were never in the photo. Small identity details can wander mid-clip — an earring changed shape in that test, and lip colour reads slightly stronger than the input in two others. Multi-person dialogue is not available here, output resolution is modest (1248×1664 in our portrait tests — no 4K), and it only makes speech video: for action, camera movement or scene changes, use Seedance 2.5 or Kling 3.0 Turbo.
Videos you generate can be used in personal and commercial projects under our terms of service. The face is the part to take seriously: animating a real person's photo without their consent can violate likeness and publicity rights, and several jurisdictions now require disclosure of synthetic presenters. Generated faces — like every face on this page — avoid the consent problem entirely.
OmniHuman 1.5 avatar model integrated with a 60-voice TTS library (6 voice archetypes × 10 languages) and ready-made avatar gallery.
