H3 Max lip sync rejects audio under 5 seconds: what to do
Sume's MiniMax H3 Max lip sync takes 5 to 14.8 seconds of audio and returns invalid_request outside that window. Fabric is the fallback for shorter lines.
A 3 second line cannot go to minimax/h3-max/lip-sync on Sume: the API requires duration_seconds between 5 and 14.8 and refuses anything else with invalid_request. It never clamps, because the provider rejects audio under 5 seconds. For shorter lines use the default talk model, VEED Fabric 1.0 (veed/fabric-1.0), or join lines into one longer take.
This comes from the Sume H3 Max lip sync API notes on main (origin/main b28f3b0, read 2026-10-13); the route is part of the avatar_image_to_video job type.
| Audio length | H3 Max lip sync | Fabric 1.0 (default) |
|---|---|---|
| Under 5 s | Refused (invalid_request) | Use Fabric |
| 5 to 14.8 s | Allowed; billed ceil(seconds) | Allowed |
| Over 14.8 s | Refused; provider would silently clip | Use Fabric, or split at a pause |
Why not just pad the audio to 5 seconds?
The packet guidance says never to cut or synthesize the audio again to make it fit, because the waveform must stay identical to the approved voiceover. Silence padding is an edit of that file. If a line is short, the cleaner move is to give it to Fabric, or to write the neighbouring sentence into the same take so the single audio file naturally reaches 5 seconds.
What does a 5 second clip cost?
Sume reserves ceil(duration_seconds) times the provider per-second list ($0.05 at 480p, $0.08 at 768p, $0.16 at 1080p, from a fal page stamped 2026-09-22) times the 1.25 house margin. A 5 second clip at 768p reserves $0.50 (5 x $0.08 x 1.25). The same doc gives $0.94 for a Fabric reservation at 720p. The amount is reserved at admit, captured on completion and refunded on failure.
- Pick the model with
modelonavatar-image-to-video_create(MCP), or callPOST /v1/minimax/h3-max/lip-syncdirectly. - Use one lip sync model per run: the model sets the assemble
output.fpslock, and Fabric is 25 fps.
Sources
Related posts
More in Models
- Porting a Sume agent loop to Haiku 5.5: drop temperature, top_p, top_k
Anthropic says to omit temperature, top_p and top_k on Haiku 5.5, and some values return a 400. Strip them from your loop before the first Sume tool call.
- HelixBench: what spatial-acoustic consistency tests in AI video
HelixBench scores whether synthesized sound fields track camera movement. What the arXiv page says it measures, and a manual way to spot-check any AI clip.
- HelixWorld license: Apache 2.0 code, weights and paper differ
HelixWorld's code is Apache 2.0, its weights have a separate license on Hugging Face, and the arXiv page lists CC BY-NC-SA 4.0. Read all three.
- Image-to-video API with sound: five-second cost by Sume row
Omni Flash 1.1 (always on), Wan 3.0 (audio), MiniMax H3 and H3 Max (native stereo) take a first frame and render sound. A 5 s clip is $0.19 to $0.94 on Sume.
Written by Sume