OmniHuman 2.0 on fal: Photo Plus Audio to Talking Video at $0.105 per Second (October 2026)
ByteDance's OmniHuman 2.0 turns one photo and 4 to 300 seconds of audio into a performing, gesturing video, at $0.105 per second. What it accepts, what it costs per clip, its record of outages on fal, and how it relates to 1.5.
OmniHuman is ByteDance's audio-driven human video model: you give it a portrait and a recording, and the person in the portrait performs the recording, with face, mouth, head and upper-body movement all following the audio. Version 2.0 appeared on fal in mid-October 2026 (the vendor's API documentation was updated on 13 to 15 October), at $0.105 per second, against $0.16 for the 1.5 release it replaces. fal now marks 1.5 as superseded. For talking photos driven by a recording this is the new default candidate, with a record you should know about before you build on it.
What it costs
| Audio length | OmniHuman 2.0 ($0.105) | OmniHuman 1.5 ($0.16) |
|---|---|---|
| 10 seconds | $1.05 | $1.60 |
| 15 seconds | $1.58 | $2.40 |
| 30 seconds | $3.15 | $4.80 |
| 1 minute | $6.30 | $9.60 |
| 5 minutes (the maximum) | $31.50 | $48.00 |
Cost depends only on the length of the audio; there is no resolution setting to multiply it. Price read from fal on 18 October 2026, within days of the release. The saving against 1.5 is about a third.
What it accepts
- Audio: 4 seconds to 300 seconds, WAV or MP3, up to 15 MB. The video is exactly as long as the audio.
- Image: JPEG, PNG, WEBP or HEIC up to 30 MB, with an aspect ratio between 0.4 and 2.5, so a 9:16 portrait works and sets the shape of the result.
- Prompt: optional guidance in Chinese, English, Japanese, Korean, Spanish (Mexico) or Indonesian.
There is no parameter for resolution, so what you get is what the model produces. That is simpler than most video endpoints and also means you cannot trade quality for price.
Why people choose it
It is built for delivery. When the recording carries the content (a joke, a pitch, a lesson), a model that reads the waveform and puts the pauses, emphasis and pace on the face and shoulders is doing something a script-to-video model cannot. HappyHorse generates its own voice; OmniHuman takes yours. If you already recorded the read, you are not paying a model to regenerate audio you own. HappyHorse vs OmniHuman explains the split in detail.
Where it fits
Audio-only Reels turned into a person. A podcast clip, a voice memo, a commentary track over a trend. The source has no usable face, so the job is "give this audio a speaker", which is the case the avatar models exist for. See turning an audio-only Reel into a talking avatar.
Talking photo with a real recording. A short message, a product explainer, a greeting in your own voice.
Cost-conscious rewrites of the 1.5 workflow. If you built on 1.5 you get the same job for about a third less (the two versions compared).
The record, honestly
OmniHuman on fal has been unavailable twice in recent months, on 16 July and on 4 September 2026, and our own test on the latter date failed on availability rather than on policy. Version 1.5 accepted real people's photos on fal when it ran. We have not yet run 2.0 on our own assets; our planned test is a 10-second clip from a generated character at about $1.05. Before you depend on it, have a fallback: Kling AI Avatar v2 Pro at $0.115, or MiniMax H3 Max Lip Sync for very long audio, both of which we have used or are evaluating.
Limits to design around
Like every model in this class, identity softens over time, so long recordings are better cut into short pieces and joined. The photo matters more than the prompt: frontal, evenly lit, whole head in frame. And rights are yours to handle: the recording, the voice and the photographed person all need to be cleared. Cloning a specific real person's voice and putting words in their mouth is the use most likely to be removed by a platform.
Prompting and framing tips
Use a photo where the speaker is roughly shoulder-up to chest-up, because the model animates the upper body and gives the best results when it can see it. Keep the prompt short and calm: describe the setting and the manner of speaking, not the emotions you want. Start with the first ten to fifteen seconds of the audio to check the result before paying for the whole file, since cost scales with audio length. If you are recording the audio yourself, leave a half second of silence at the start and end, and avoid music beds under the voice; clean speech gives the cleanest lip movement.
Common questions
How long can the audio be?
From 4 seconds up to 300 seconds, as WAV or MP3 up to 15 MB. The video is exactly as long as the audio.
Is there a resolution setting?
No. The endpoint has no resolution parameter, so price depends only on audio length.
Is it reliable?
OmniHuman on fal was unavailable on 16 July and 4 September 2026, so keep a fallback such as Kling AI Avatar or H3 Max Lip Sync.
Does it work with real faces?
Version 1.5 accepted them on fal. We have not run 2.0 on our own assets yet, so test with your own photo first.
How to do this without code: A photo plus a recording is the Animate photo tool with audio, and OmniHuman 2.0 is the model we are testing for it. In AutoHustle this is the Animate photo tool with a voice recording: audio in, a talking photo out (about 3 credits per second with the audio-driven model, quoted before you press Generate). See How to make a talking photo and Animate a photo.
Related reading: How to Animate a Photo Into a Talking Video With Lip Sync (2026) · Talking Photo Apps Compared 2026: Four Models, $0.06 to $0.14 a Second
For developers: the fal endpoint
| Model | fal endpoint |
|---|---|
| OmniHuman 2.0 | bytedance/omnihuman/v2 |
| OmniHuman 1.5 (superseded) | fal-ai/bytedance/omnihuman/v1.5 |
Inputs: image (JPEG/PNG/WEBP/HEIC, ≤30 MB, aspect 0.4–2.5), audio (WAV/MP3, 4–300 s, ≤15 MB), optional prompt. No resolution parameter. $0.105 per second of audio, read on 18 October 2026.