H3 Max Lip Sync: Make One Photo Speak Your Recording

H3 Max Lip Sync makes one still face speak one supplied recording. You give the dedicated endpoint an image and an audio file; it returns a video with the original soundtrack attached and mouth movement synchronized to it. The recording is not merely a style reference, and the model is not being asked to invent a new line.

That distinction is the whole reason to use this route. If the exact voiceover has already been approved, a prompt-generated voice is the wrong tool. If you want a new scene with newly generated dialogue, Lip Sync is the wrong tool.

AI Live Generator does not currently expose this specialist route or accept an audio upload. The site's generator uses standard H3 Max Turbo text-to-video and image-to-video with newly generated sound; this article is a preparation and integration guide for the separate Lip Sync endpoint.

This is not the ordinary image-to-video form

The current minimax/h3-max/lip-sync/image-to-video contract has two required inputs: image_url and audio_url. It also exposes resolution, transcription guidance, seed, and the safety checker. It has no prompt field.

Use that constraint to choose the workflow before preparing files:

Trying to smuggle scene direction into the file name will not help. Lip Sync is a performance transfer with a deliberately small control surface: face, recording, output size, transcription behavior, and seed.

Pick a portrait that can carry the performance

The endpoint accepts photographs, illustrations, renders, and paintings. The live schema allows image aspect ratios from 0.4 to 2.5, so vertical, square, and wide frames all fit the contract. Technical acceptance does not make every composition equally readable.

For the first pass, use one adult face in a front or slight three-quarter view. Leave the mouth unobstructed. Avoid a hand, microphone, cup, scarf, heavy shadow, or crop crossing the lips. Give the face enough pixels to read without making it an extreme close-up. If two people are equally prominent, the request no longer communicates clearly which mouth owns the recording.

Expression matters too. A still with tightly closed teeth, a wide frozen grin, or a dramatic profile asks the animation to travel farther before it can form ordinary speech. A relaxed mouth and clear jawline give the motion a quieter starting point. Keep glasses and distinctive features if they matter to identity, but do not expect resolution alone to repair a weak or obstructed source.

Before uploading, write one sentence about what must remain recognizable: the person, an illustrated character, a costume, or the lighting. That gives you a specific review target after the mouth begins to move.

Treat the recording as the timeline

The output uses the supplied soundtrack, so edit the recording before the request. Remove the false start you do not want, leave intentional pauses where the face should stop speaking, and listen for room noise, clipping, or music that masks consonants. A clean recording makes the synchronization problem easier to judge; it does not need to sound artificially dry.

One speaker is the safest first test. If the file contains an interviewer and a guest, laughter from off camera, or a chorus, a single still face has no reliable way to explain who owns each syllable. Split the material or choose a shot whose visual logic matches the sound.

There is an important documentation conflict as of October 20, 2026:

• the live API schema says audio must be at least five seconds, supports up to 15 minutes, and that the output matches the supplied audio duration;
• fal's dedicated H3 Max Lip Sync landing page still says 5–15 seconds, with longer audio clipped to the first 15 seconds;
• the current model page prints a 1.2× billing multiplier above 15 seconds and even gives a 30-second cost example.

Those pages cannot all describe the same production limit. Until fal reconciles them, do not promise a long-form render based on one paragraph. A five-to-fifteen-second file is the conservative shared range. For anything longer, check the endpoint contract at request time; then confirm the returned duration and listen to the end of the output before building a long workflow around it.

Turn transcription on for words, off for the waveform

enable_transcription changes how the route interprets the recording. With transcription guidance, the system can use the recognized words to help place mouth shapes. That is the sensible starting point for clean spoken dialogue.

Waveform-only synchronization is useful when a transcript would be misleading: singing, vocables, heavily processed speech, an invented language, or an effects-heavy vocal. Turning transcription off does not magically improve every song. It simply removes a text interpretation that may be wrong.

Use the switch as a diagnosis, not as a quality label:

1. For ordinary speech, submit the recording with transcription enabled.
2. If the words are recognized but the mouth timing is poor, inspect the face, pauses, and audio clarity before changing modes.
3. If transcription itself is likely to misunderstand the performance, make one waveform-only comparison while keeping the image, audio, resolution, and seed unchanged.

Changing five inputs at once gives you two different videos and no explanation.

Start at 768P unless you are answering a different question

The live schema lists 480P, 768P, 1080P, and 2K, with 768P as the default. The endpoint chooses the supported aspect ratio nearest the source image.

Resolution changes detail and cost; it does not change the words in the recording. Use 480P for a cheap layout or timing check when small output is acceptable. Use 768P for the first serious review. Move to 1080P or 2K only after the face, framing, and recording already work and the final placement benefits from more detail.

fal's H3 Max Lip Sync page currently describes per-second billing by resolution. Recheck the model page before budgeting because provider pricing and promotions can change. The useful cost formula is simple:

audio seconds × current rate for the selected resolution = generation cost

Do not estimate a 2K talking portrait from the general H3 Max headline price; this specialist route publishes its own resolution tiers.

A small request is easier to audit

A minimal speech request can be expressed without inventing a prompt:

{
  "image_url": "https://example.com/authorized-portrait.jpg",
  "audio_url": "https://example.com/approved-line.wav",
  "resolution": "768P",
  "enable_transcription": true
}

Keep the API key on the server. The official client supports queued submission, status checks, and result retrieval; a browser should call your protected server route rather than receive the provider credential. Store the returned request ID with the source asset IDs so a retry does not become an untraceable duplicate job.

The response includes the output video, seed, and duration. Save the seed with the result, but remember what it can and cannot do: it helps reproduce a request with the same inputs; it does not turn a different portrait or edited recording into the same performance.

Review the result in three passes

First, listen without watching. Confirm that the returned soundtrack is the recording you supplied and that no required beginning or ending was lost. This is also where the length disagreement in the public documentation becomes visible.

Second, watch muted. Look for face drift, a changing jawline, teeth that appear between words, and mouth motion continuing through a deliberate pause. Compare the eyes, hairline, glasses, and clothing with the source—not only the lips.

Third, watch and listen together at normal speed. Check the first consonant, the longest phrase, and the last closed-mouth beat. A slow-motion loop can locate a defect, but the acceptance decision should match how the clip will actually be watched.

If the output uses the right words but the mouth is wrong, change one visual or synchronization variable. If the recording itself contains the wrong line, no mouth setting can repair the script. The existing lip-sync troubleshooting guide gives a stricter wrong-words-versus-bad-timing review, while the custom-audio guide explains when target audio or reference audio is the better contract.

Get permission for both halves of the performance

A talking portrait combines a likeness and a voice. Use an image and recording you own or have permission to process, and make the synthetic treatment clear where viewers could mistake it for a real statement. Permission to post a photograph is not automatically permission to make the person say new words; access to a recording is not automatically a license to reuse the performance commercially.

The strongest first result is rarely the most elaborate one. One clear face, one short authorized recording, one resolution, and one review checklist tell you whether this route solves the actual job. Once that works, variations become a production decision instead of a guessing game.