MiniMax H3 Max Lip Sync on fal: Talking Photos From Audio Up to 15 Minutes (October 2026)
A photo plus a recording becomes a talking video, priced from $0.05 per second at the 768P base and accepting audio up to 15 minutes. What it takes, what the price leaves out, and how it compares with OmniHuman 2.0 and Kling Avatar.
Most talking-photo models stop at a few minutes of audio. MiniMax H3 Max Lip Sync, which appeared on fal on 18 September 2026, accepts a recording of up to 15 minutes and returns a video exactly as long as the audio. At a listed base rate of $0.05 per second that is the cheapest route we know of to a long talking portrait, and the only one that reaches podcast length in a single call. It is also new, untested by us on real faces, and priced in a way that hides the cost of higher resolutions.
What it takes
- A photo of the speaker. The aspect ratio can be anything from 0.4 to 2.5 (tall portrait to wide), and the video follows the photo's shape, which matters if you want a 9:16 result: start from a 9:16 image.
- An audio file of at least five seconds and at most fifteen minutes. The output length equals the audio length.
- Resolution: 480P, 768P, 1080P or 2K. The $0.05 base applies to 768P; fal does not publish the multiplier for 1080P or 2K.
- Transcription: an
enable_transcriptionoption, useful when the recording is speech and you want the timing aligned to words.
You supply the voice. The model does not generate speech, so it is a tool for people who already have audio: a recorded voice-over, a podcast clip, a cloned voice you have the rights to, a translation produced elsewhere.
What it costs
| Audio length | At the 768P base rate |
|---|---|
| 15 seconds | $0.75 |
| 1 minute | $3.00 |
| 5 minutes | $15.00 |
| 15 minutes (the maximum) | $45.00 |
Those are base rates. Treat anything above 768P as an unquoted cost until you run one short clip at the resolution you want and read the bill.
How it compares with the other audio-driven options
| Model | Price per second | Longest audio | Notes |
|---|---|---|---|
| MiniMax H3 Max Lip Sync | $0.05 (768P base) | 15 minutes | New; not verified on real faces by us |
| OmniHuman 2.0 | $0.105 | 300 seconds | Body and gesture follow the audio |
| Kling AI Avatar v2 Pro | $0.115 (Standard $0.0562) | about 5 minutes | In production use in our onboarding video |
| Sync-3 lip-sync | $0.133 | n/a (works on a video) | Fixes the mouth on footage you already have |
The pattern: H3 Max Lip Sync is the budget and the length champion, OmniHuman is the one built to move the whole upper body with the delivery, Kling Avatar is the one we have shipped, and sync-3 solves a different problem (editing the mouth in an existing video, not animating a still).
Where it fits
Long talking portraits. A ten-minute lesson, a podcast episode illustrated by a host portrait, a narrated explainer. At fifteen minutes the cost is real but a fraction of the alternatives.
Draft versions of shorter pieces. Before spending $1.58 on a 15-second OmniHuman clip, a 75-cent H3 pass tells you whether the photo and the audio work together at all.
Not for Reels in one call. A Reel is under a minute and benefits from the more expressive models. The long-audio ceiling is a feature for a different format.
What we have not tested
We have not run H3 Max Lip Sync on a real person's photo, so we cannot say how stable the face is over a long clip, how well the lip movement tracks fast speech or a non-English voice, or whether the head stays put or wanders. MiniMax's moderation rejects named public figures and fal's recommender gives no identity guarantee for any model in this category. The two risks to watch on long outputs are drift (the face softens as the clip goes on; cutting the audio into two-minute pieces and joining them is the standard remedy) and cost surprise at higher resolutions.
A sensible test costs about a dollar: a 15-second clip of your own voice on a clean frontal photo at 768P. If the lips and the face hold up, run a two-minute version before you commit to a long one.
A word on rights
Whatever audio you use, you need the rights to it, and the photo needs to be of someone who agreed. Cloning a specific real person's voice and putting words in their mouth is the version of this format platforms take down fastest; the same holds for AI baby podcast clips, where a generic character and your own voice is the version that stays up.
Common questions
Does it generate the voice?
No. You supply the audio and it animates the face to match. If you only have a script, you need a voice first, either your own recording or a text-to-speech track you have rights to.
Can I make a vertical video?
The video follows your photo's shape, with an aspect ratio between 0.4 and 2.5. Start from a 9:16 portrait and the result is vertical.
Why not use it for every talking clip?
It is new and we have not tested it on real faces. For short Reels we prefer models with a record, and use this one for very long audio where its price and 15-minute ceiling matter.
What happens to identity over ten minutes?
Faces soften over long clips on every model we know. Cut the audio into pieces of one to two minutes, generate each, and join them.
How to do this without code: A talking photo from a recording is the Animate photo tool with audio, and H3 Max Lip Sync is a candidate for long audio. 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: Kling LipSync Audio-to-Video on fal: Limits, Price Unknowns and Why Reels Need a Scene-by-Scene Approach (October 2026) · MiniMax H3 Max Image-to-Video on fal: Price, Limits and Examples (October 2026)
For developers: the fal endpoint
| Model | fal endpoint |
|---|---|
| MiniMax H3 Max Lip Sync (image-to-video) | minimax/h3-max/lip-sync/image-to-video |
Inputs: an image (aspect ratio 0.4–2.5), an audio file (5 s to 15 min), resolution (480P / 768P / 1080P / 2K), enable_transcription. The listed $0.05 per second is a 768P base rate read on 18 October 2026.