Provide a video and a driving audio track. The model adjusts the visible
speaker's mouth movements to the supplied audio while aiming to preserve the
person, scene and motion. The supplied audio replaces the original soundtrack;
source audio is optional. This endpoint does not translate speech, synthesize
voices or accept a user prompt.
Request fields
| Field | Type | Default | Purpose |
|---|
video_url | string, required | — | Source video, 5–900 seconds. Display aspect ratio must be between 0.4 and 2.5. |
audio_url | string, required | — | Driving audio, 0.1–900 seconds. |
duration_mode | string | cut_off | How to handle different video and audio lengths; see below. |
resolution | 480P or 768P | 768P | Output resolution tier. |
seed | integer or null | Random | Generation seed; the resolved value is returned. |
enable_safety_checker | boolean | true | Uses the standard fal safety-check settings. |
Output video is normalized to 24 fps and uses the supported aspect ratio nearest
the source. Prompt expansion runs automatically. Only the documented request
fields are accepted; strength, adapter and attention settings are fixed.
Duration mode
The selector is named Duration mode in the frontend and duration_mode in
JSON. Its values describe timeline handling:
| Value | Behavior | Output duration |
|---|
cut_off | Trim to the shorter input. | Shorter of video and audio. |
loop | Repeat the video forward when the audio is longer; trim the video when the audio is shorter. | Audio duration. |
bounce | Alternate forward and reverse playback when the audio is longer; trim the video when the audio is shorter. | Audio duration. |
silence | Keep the video length and pad shorter audio with silence. Reject audio longer than the video. | Video duration. |
For a 5-second video and 8-second audio, cut_off produces 5 seconds,
loop and bounce produce 8 seconds, and silence returns a validation error.
For a 5-second video and 3-second audio, cut_off, loop and bounce
produce 3 seconds; silence produces 5 seconds with 2 seconds of silence.
Python example
Install fal-client and set the FAL_KEY environment variable to your API key:
import fal_client
result = fal_client.subscribe(
"fal-ai/minimax-h3-max-lipsync-v2v",
arguments={
"video_url": "https://example.com/speaker.mp4",
"audio_url": "https://example.com/new-soundtrack.wav",
"duration_mode": "cut_off",
"resolution": "768P",
"seed": 42,
"enable_safety_checker": True,
},
)
print(result["video"]["url"])
print(result["duration"])
Replace the example URLs with accessible media URLs. subscribe uses fal's queue
and waits for the result. To manage polling yourself, use fal_client.submit
with the same endpoint ID and arguments.
Response
| Field | Purpose |
|---|
video | Output file with url, content_type, file_name and file_size. |
seed | Resolved generation seed. |
duration | Final output duration in seconds. |
timings | Optional timing breakdown; inference time is summed over generated windows. |
Pricing
| Resolution | Per billed output second | 10-second output |
|---|
| 480p | $0.10 | $1.00 |
| 768p | $0.20 | $2.00 |
Errors
- Invalid media, unsupported request fields, invalid duration modes and audio
longer than the video in
silence mode produce validation errors (422).
- Policy violations stop the request through the standard safety checks.
minimax/h3-max/lip-sync/video-to-video