Upload one clear, front-facing photo with a visible face. JPG, PNG, or WebP up to 10 MB; Prizmad resizes the long side to 2048 px. Use at least 300 px per side and an aspect ratio from 0.4 to 2.5.
Upload the voice track as MP3 or WAV, 4–300 seconds and up to 15 MB. Token cost follows the verified audio duration.
Optional direction for the scene, delivery, movement, and camera. Prompts work best in English, Chinese, Japanese, Korean, Spanish (Mexican), or Indonesian.
Cost is calculated after audio upload
Portrait Image is required
Your generated video will appear here
OmniHuman V2 turns one portrait photo and a voice track into a lip-synced talking video with natural expressions, head movement, and optional performance direction. On Prizmad it runs on your existing token balance at 21 tokens per second of speech.
OmniHuman V2 is an audio-driven talking-head model: give it a single clear portrait and a speech recording, and it animates the person to match the voice. Lips, facial expressions, blinks, and head movement follow the performance in the audio instead of a generic loop, so the result reads like a real presenter rather than a slideshow with a moving mouth.
The optional Performance Prompt adds direction on top of the audio: scene, delivery, gestures, and camera. Use it to keep a calm spokesperson read, add natural hand gestures on key words, or request a slow push-in for a premium beauty shot. Prompts work in English, Chinese, Japanese, Korean, Mexican Spanish, and Indonesian.
Accept one portrait as JPG, PNG, or WebP up to 10 MB; Prizmad resizes the long side to 2048 px. Use at least 300 px per side with an aspect ratio from 0.4 to 2.5. The voice track can be MP3 or WAV from 4 to 300 seconds (up to 15 MB). Longer takes are supported in a single pass — no stitching 30-second fragments together.
Turn a product photo or creator portrait into a presenter that speaks the hook and offer — the core talking-head format for paid social.
Build a consistent presenter from one approved photo and produce new explainers, testimonials, and landing-page videos without another shoot.
Keep the same face and swap the voice track per market. One portrait can carry multiple languages and offers while the visual identity stays consistent.
A close-up portrait with natural expression and soft studio light reads like a premium beauty creator rather than an obvious avatar.
Record a clean voiceover once and turn it into a full-length presenter video without filming yourself or editing mouth movements by hand.
A single front-facing portrait is enough. The model drives lips, eyes, expression, and subtle head motion from the audio.
Optional direction for delivery, gestures, scene, and camera — from a calm spokesperson to energetic UGC energy.
Audio from 4 to 300 seconds in one generation, so a full product explainer or localized ad read stays seamless.
21 tokens per second of verified speech. A 5-second hook is 105 tokens; a 60-second read is 1,260 tokens.
Performance prompts are supported in six languages, making it practical to keep one face and swap the market-specific voice track.
Finished MP4 videos land in your asset library with full commercial rights and can be dropped straight into the video wizard or downloaded.
Upload a clear, front-facing portrait with one visible face. Use even lighting and keep the head and shoulders in frame.
Upload the voice track as MP3 or WAV — between 4 and 300 seconds and up to 15 MB. Trim silence at the start and end so you do not pay for dead air.
Optionally add a Performance Prompt describing delivery, gestures, the scene, and camera movement.
Check the token estimate, which follows the verified audio duration, then click Generate.
When the render finishes, preview the MP4 and download it or use it in a project.
Billing follows the verified audio duration, rounded up to whole seconds: 4 seconds = 84 tokens, 5 seconds = 105, 30 seconds = 630, and the 300-second maximum = 6,300 tokens. The workspace shows the current cost before you generate.
OmniHuman V2 uses the same token balance as image generation, video, avatars, voiceover, captions, and music. There is no separate model account or API key.
If the render fails, the run is refunded through the standard AI Studio flow. You are only charged for a completed video.
One portrait photo and one speech audio file. The portrait can be JPG, PNG, or WebP up to 10 MB; Prizmad resizes the long side to 2048 px. Use at least 300 px per side with an aspect ratio from 0.4 to 2.5. Audio must be MP3 or WAV, 4–300 seconds, and up to 15 MB.
21 tokens per second of verified audio. A 5-second clip is 105 tokens, a 30-second clip is 630 tokens, and a full 5-minute take is 6,300 tokens. The cost updates automatically once the audio duration is verified.
It adds optional direction for the scene, delivery, gestures, movement, and camera — for example a warm confident presenter with natural blinks, relaxed hand gestures, and a slow push-in. Leave it empty to let the audio drive the performance.
From 4 to 300 seconds in a single generation, matching the audio length.
The current model does not expose seed or resolution controls. Output framing follows the source portrait, and the render is produced at the model's native quality.
Yes. Every video generated on Prizmad comes with full commercial rights, so you can run it on TikTok, Meta, and YouTube or embed it on landing pages.
Published 2026-10-19 · Last updated 2026-10-19