OmniHuman V2 on fal: 5-Minute Talking Videos

Introduction

OmniHuman V2 from ByteDance is now live on fal. Feed it one image of a person and an audio track, and it returns a talking video with lip movements, expressions, and body motion matched to the speech. It is a closed, API-only image-to-video model, and you can run it today in the fal playground or through the bytedance/omnihuman/v2 endpoint. The two numbers that jump out: audio clips of up to 300 seconds, and a price of $0.105 per second.

What shipped

fal

View Post on X

at 22:01Z Oct 15 (about 06:01 MYT Oct 16), and the model page lists ByteDance as the lab with a commercial license.

Input: one image of a human figure plus a wav or mp3 audio file.
Optional prompt: steer the scene, movements, and camera. Prompts are supported in Chinese, English, Japanese, Korean, Spanish (Mexican), and Indonesian.
Audio length: 4 to 300 seconds, up to 15 MB.
Image rules: jpeg, png, webp, bmp, tiff, gif, heic, or heif; up to 30 MB; aspect ratio (width ÷ height) between 0.4 and 2.5; each side between 300 and 6,000 pixels.
Output: a video file plus the billed duration.
Price: $0.105 per second of output.

What changed from v1.5

OmniHuman v1.5 is still listed on fal, which makes the comparison easy:

Cheaper: v1.5 costs $0.16 per second, so V2 is roughly 34% cheaper. A one-minute clip drops from about $9.60 to about $6.30, and a full five-minute monologue on V2 comes to about $31.50 (our maths from fal's listed prices).
Much longer: v1.5 caps audio at 30 seconds in 1080p mode and 60 seconds in 720p mode. V2 accepts up to five minutes in one call.
Simpler controls: V2 drops v1.5's resolution and turbo_mode switches. The schema doesn't state an output resolution, so check your first render before you plan a 4K edit around it.
Multilingual prompting: V2 lists six prompt languages explicitly.

Try it

Here is a minimal API call. Swap in your own hosted image and audio URLs:

curl --request POST \
  --url https://fal.run/bytedance/omnihuman/v2 \
  --header "Authorization: Key $FAL_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "image_url": "https://example.com/presenter.png",
    "audio_url": "https://example.com/voiceover.mp3",
    "prompt": "Warm, confident presenter, gentle hand gestures, slow push-in camera"
  }'

And a paste-ready performance prompt for the optional prompt field:

A warm, confident presenter speaks directly to camera in a softly lit studio, natural blinks, small expressive eyebrow movements, relaxed open hand gestures on key words, slight head tilts at the end of sentences, slow steady push-in camera, shallow depth of field

Pro Tip

Trim silence from the start and end of your audio before uploading. You pay per second of output, and a five-minute track that opens with ten seconds of dead air is ten seconds of a frozen face you are paying for.

Why it matters

Talking-head video has been stuck in short bursts: generate 30 seconds, stitch, hope the seams don't show. A single five-minute pass at about $6.30 per minute changes the maths for course videos, product explainers, dubbed shorts, and virtual hosts. For creators, the prompt field is the fun part, because you can direct performance and camera instead of accepting a static nodding head.

Caveats

API only: no open weights, no ComfyUI node yet. Run it on fal.
Resolution unstated: the V2 schema has no resolution setting, so test before you commit.
Likeness and consent: this turns any portrait into a speaking person. Use images and voices you have the rights to, and label synthetic video where your platform requires it.
Early days: the fal listing is the launch artifact we're citing; we'll update if ByteDance publishes deeper specs or benchmarks.

Original Source