Quick answer: H3 accepts up to three reference audio clips (MP3 or WAV, 15 MB each) on image-to-video lanes from 480P to 2K, and never on 2K text-to-video. Audio conditions the result, but dedicated lip-sync mode is unavailable and exact synchronisation is not guaranteed at any tier. Image-to-video still needs at least one reference image.
Almost every video model takes images. Far fewer take audio as an input, and the ones that do tend to document it in one line. H3 accepts up to three reference audio clips alongside up to nine reference images, which is enough to change how you write a shot — but only on some lanes, and only up to a point.
The limits, plainly
| Constraint | Value |
|---|---|
| Audio references per job | up to 3 |
| Accepted formats (in the web generator) | MP3, WAV |
| Per-file size | 15 MB |
| Lanes that accept audio references | image-to-video at 480P, 768P, 960P, 1080P, 1440P and 2K |
| Lanes that do not | text-to-video at 2K (no references at all) |
That last row surprises people: the most expensive tier's text-to-video scene takes no reference audio. If your job is "voice file in, speaking shot out", the 2K text lane is the wrong door — use image-to-video at 2K, or any of the cheaper image lanes.
One more rule that rejects quietly: image-to-video needs a reference image. Audio alone is not enough for that scene — attach at least one image, or the job fails validation before it is submitted. Supported text-to-video scenes can accept audio without an image, except at 2K; video-to-video needs its source clip.
What a reference clip actually conditions
Three things change when you attach audio, and it is worth separating them because people conflate them:
- Performance timing. The clip gives the model a rhythm to work against — pauses, breaths, the length of a phrase. Shots conditioned this way tend to have fewer arbitrary cuts and more sustained beats.
- Source character. Voice timbre and room tone pull the ambience and the character's presence toward the clip, useful when the audio came from the same production.
- Lip movement is a limit to keep explicit. Dedicated lip-sync mode is unavailable, including at 480P and 768P; requests for that mode are rejected. Reference audio still conditions the result, but exact synchronisation is not guaranteed. Our own planner documents it in one sentence: "Optional reference audio conditions the result; exact lip sync is not guaranteed."
That sentence is the honest boundary. If a client asks for frame-accurate lip sync at 1080P, the right answer is to shoot or to post-sync, not to promise a generation lane will land it.
The mistake that eats dialogue: short audio, long clip
A five-second clip that has to carry a seven-second line will cut the line. Two ways to avoid it:
- Match the lengths before you submit. If the line is 6.2 seconds, request seven seconds of output, not five.
- Pad the audio, not the video. A deterministic audio-pad step takes your clip, an optional start offset (0–59 s) and a target duration (2–60 s) and returns a padded track. Feed the padded file as the reference, generate the full length, and you keep the whole line instead of a truncated one.
Padding is a file operation, not a generation: it does not consume generation credits, and it leaves the original asset untouched. It is also the difference between "the take is unusable, regenerate" and "the take is usable".
When is audio conditioning not worth it?
- You have no audio yet. The model can generate its own ambience; a reference clip with nothing in it just makes the shot duller.
- You need a music video to hit a beat grid. H3 quantises output to 24 fps frames, so a requested five seconds lands at about 5.17 — close, but not a beat-synchronised cut. Generate long and cut to the timeline.
- The audio is a final mix with dialogue, music and effects. Reference clips work best when they carry one thing. A mastered mix conditions everything at once and the result is mush.
FAQ
Does H3 do lip sync? Dedicated lip-sync mode is unavailable at every tier. Reference audio conditions the result, but exact sync is not guaranteed.
Can I send audio without an image? Image-to-video requires at least one image. Supported text-to-video scenes can take audio without an image, except at 2K; video-to-video requires a source clip.
Why did the spoken line get cut? The clip was shorter than the shot. Pad the audio to the target duration rather than stretching the video.
Where this runs
The reference-audio field appears in the H3 generator's supported modes (up to three clips, MP3/WAV, 15 MB each) — https://callirra.com/minimax-h3 documents the tiers and the reference limits per lane. The audio-pad step lives in Constellation, documented alongside the other finishing tools at https://callirra.com/constellation.
I work on Callirra, so the support matrix above is our enforced table rather than a guess. If you find a lane where sync is better than described, that is worth publishing — it would change the advice in this post.