Lip-sync a 20-second line: H3 Max takes 14.8 s of audio, so split it

The limit and the workaround

Sume's POST /v1/minimax/h3-max/lip-sync takes audio of 5 to 14.8 seconds, so a 20-second line must be cut into two takes of at least 5 seconds each, and each take run as its own job. Cut at a sentence boundary so nobody sees a word split in half, then join the two videos.

If the line is long and the speaker is a still photo, the other route is POST /v1/veed/fabric-1.0, which takes a still and audio of up to 10 MB with no 14.8-second ceiling.

Pick the route

The right route depends on what you start with: a video of a person, or a still.

Cutting a 20-second line in two

Write the script in two parts before you generate the voice, so each part is a whole thought of roughly 7 to 12 seconds. Generate each part as its own tts-1.0 job. This is better than cutting finished audio, because the voice finishes a sentence naturally. If you already have one 20-second file, use POST /v1/timeline-1.0/audio with operation: "split" to slice it into ranges at $0.01 flat, and choose a silent point between sentences.

Send each audio part with the same video to its own lip-sync job. Make sure each part is within 5 to 14.8 seconds, because a short tail under 5 seconds is refused. Then join the two clips in order with Sume's timeline. Expect a slight change of expression at the join; hide it with a cut-away or a caption. If the two takes must look continuous, start the second take from a frame close to where the first ended, and keep lighting, framing and resolution identical so the seam is hard to spot. Cost scales with the number of takes, so a 40-second line would need three or four, and at that length Fabric is simpler.

A rule from the docs

Sume's docs say that a face that talks is Fabric with an accepted still plus TTS, and that video models do not lip-sync to TTS. So do not expect a text-to-video model to sync a voice file you provide; use a lip-sync route for that.

• Check the duration_seconds of each audio part before you submit.
• Pay attention to price: H3 Max is billed at list times 1.25, so test one take at 480p first, and compare it with a Fabric render of the same line.
• Keep the same voice and settings for both parts.
• Never resubmit a paid job that looks slow; poll its status.