VIBEPAPER JOURNAL / AI CREATION

MiniMax H3 in VibePaper: Assigning Roles to Mixed References, Planning Sound With the Picture, and Where It Still Breaks

VibePaper · Updated

MiniMax H3 takes text, images, video and audio as one creative context, then generates a synchronized picture-and-sound scene. The practical difference from a text-to-video endpoint is that every file you attach is a decision: it has to answer one question about the shot, or it adds noise. This article covers how H3 reads mixed references, how to assign roles, how sound is planned alongside the picture, what MiniMax publishes about its 768p and 2K passes, and what happened when we tried to generate in Paper Gen. Where a claim comes from MiniMax, we say so.

  • MiniMax H3
  • AI video generation
  • mixed references
  • audio-visual generation
  • VibePaper
A still frame from a MiniMax H3 video
A MiniMax H3 frame. The steps below follow the VibePaper Paper Gen interface.

At a glance

MiniMax H3 as published by MiniMax, with the Paper Gen settings that shape a job
CapabilityPublished figure or behaviorWhat it means in Paper Gen
Clip length4–15 seconds per generationPaper Gen defaults to 5 seconds. Set it before you submit, because it defines the job.
Resolution768p base generation; up to 2K via the regeneration passPaper Gen defaults to 768p. Iterate at 768p and reserve higher resolution for a result you have approved.
Frame and audio24 FPS video, native 32 kHz stereo audio, one passWrite dialogue and ambience into the prompt beside the visuals they belong to.
Mixed referencesUp to 9 images, 3 videos, 3 audio clips; 12 files totalAttach them in Reference mode, one job per file.
Frame controlFirst frame, last frame, or bothUse First / last frame when the opening or closing image is already approved.
Aspect ratioLandscape, square and portraitPaper Gen defaults to 16:9. Pick the frame shape before you submit.

One context, several sources

Most video models read a prompt and return a clip. H3 reads a set of material and tries to build one scene from it. MiniMax describes the input as a single creative context, so subjects, motion, camera language, voices, ambience and rhythm can each come from a different file rather than all coming from the sentence you type.

That is H3’s main difference, and it changes the job. Before you attach anything, decide what each file is for. A reference that does not answer a question about the shot adds noise, and it can pull the result away from the one role you wanted it to carry.

Assigning a role to each reference

These roles follow from what each input type can carry. They are how we plan a job before opening Paper Gen; they are not a list of switches the interface exposes.

Matching each reference type to the part of the shot it should control
Reference typeThe question it answersExample role in a shot
ImageWho or what is in the frame, and what it looks likeThe character’s face and wardrobe, or the room’s materials
VideoHow something moves, and how the camera movesA camera push-in, a hand movement, the pace of a turn
AudioWhat the scene sounds like and who speaksA voice, a room tone, a rhythm for a cut
First / last frameHow the shot begins and resolvesThe composition at the first and final frame of the clip

Planning sound with the picture

H3 generates picture and sound together. MiniMax says the visual and audio latent spaces are separate, but generation runs inside the same system, and the audio is produced as stereo from left and right channels that are processed independently and recombined. In practice this means the sound belongs to the shot, not to a layer added afterwards.

So write the sound into the same beat as the picture it belongs to. “The wheel hums as the hands slow down” gives the model one event to synchronize, where “add ambient sound” gives it none.

One constraint to know before you start: an audio reference cannot be the only input. Audio has to sit alongside an image or video reference, and each video or audio file must run 2–15 seconds, with a 15-second cap on the total for each type.

Iterate at 768p, then regenerate at 2K

MiniMax describes two passes. H3-Base produces the 768p audio-visual sequence. H3-Regenerate-2K takes that result and the original context back into H3 and regenerates at higher resolution. It is not a plain upscale: because the model can revisit the source instructions and references, MiniMax says it recovers detail that a pixel-only enlargement cannot.

The practical loop follows from that. Use 768p to test the composition, the motion and the sound. Once a take works, regenerate that one take at higher resolution. Do not expect the 2K pass to fix a shot that was wrong at 768p.

Two checkpoints behind the model

MiniMax publishes H3-Base as two checkpoints. H3-Base-FL2VA covers text-to-video with optional first-frame, last-frame or first-and-last-frame control. H3-Base-Ref2VA handles mixed references. Both are released for deployment and fine-tuning.

Scope matters. The hosted Context-IR orchestration layer, H3-Regenerate-2K and the native sparse-attention implementation are not part of that open release, so the full official 2K pipeline still depends on MiniMax’s services. Paper Gen gives you the hosted workflow, not the raw checkpoints.

How to make a MiniMax H3 clip in VibePaper

  1. 01Open Video mode and pick MiniMax H3In Paper Gen, switch from Image to Video, open the model selector, go to the Minimax group and choose MiniMax H3. Hailuo 2.3 Fast sits beside it in the same group.
  2. 02Set ratio, resolution and durationH3 defaults to 16:9, 768p and 5 seconds. Confirm all three before you write the prompt, since they define the job.
  3. 03Assign a role to each fileDecide which reference controls the character, the motion, the sound or the framing. Drop any file that does not answer a question about the shot.
  4. 04Write the prompt as a shot with soundSubject, one action, one camera move, the light, then the ambience and any dialogue, each tied to the moment it happens.
  5. 05Generate, then read the result or the errorA job that starts appears in Your creations. A failed job shows its reason under the settings row, so you can copy the exact message for support.
The path in Paper Gen, from choosing H3 to reading what came back.

Where each control lives

The controls that shape an H3 job sit in two places: the input mode switch above the prompt box, and the settings row beneath it.

What the interface looks like

Model first. In the picker, MiniMax has its own group with H3 at the top, beside Hailuo 2.3 Fast. Ratios, durations and reference rules follow the model, so choose it before the other settings.

Limits and failures we ran into

A published feature list says what a model is designed to do. The limits decide whether it fits your work, so here is what we hit and what we have not yet verified.

  • Generation failed on the day we tested. An H3 job at 16:9, 768p and 5 seconds returned “Could not start generation. Please try again. Nexra API returned HTTP 404”. Retrying after a short wait returned the same error. We have reported it, and we are not showing an H3 result we did not get.
  • The price was not visible on the Generate button for H3 when we checked, so we could not confirm the cost before submitting. Read the Generate control before each job rather than assuming a price from another model.
  • Reference lengths are strict. Video and audio references must each be 2–15 seconds, and an audio reference cannot be the only input. Check the lengths before you upload.
  • Rights still apply. Every uploaded likeness, voice, logo or copyrighted element needs clearance, whatever the model.

A prompt checklist for your first attempt

Medium close-up: a potter’s wet hands shape a clay bowl on a spinning wheel in a sunlit studio. Slow push-in, shallow depth of field. Soft window light, dust drifting through the beam. Sound: the wheel hums as the hands slow, water drips into a metal bucket. No dialogue, no text, no logos.

Questions we get about MiniMax H3

How long can one MiniMax H3 clip be?

MiniMax publishes 4 to 15 seconds per generation. Paper Gen defaults to 5 seconds. For a longer sequence, plan several clips with a shot list for each.

Does H3 generate sound as well as picture?

Yes. MiniMax publishes native 32 kHz stereo audio generated with the picture in one pass. Write the sound into the same beat as the picture it belongs to.

Can I use H3 without references?

Yes. Text-to-video works, and first-frame or last-frame control is optional. References earn their place when a specific character, motion, voice or rhythm has to carry across the clip.

Why did my H3 job fail?

Paper Gen shows the reason under the settings row. On our test day the reason was “Nexra API returned HTTP 404”, a service error rather than a prompt problem. Copy the exact message and retry later, or send it to support.

References and scope

Technical concepts follow the official and primary sources listed below. Workflow suggestions and examples are editorial analysis, not guarantees of results.

  1. MiniMax-H3 model card, MiniMax on Hugging FaceSource for the H3-Base-FL2VA and H3-Base-Ref2VA checkpoints, the 2–15 second video and audio reference lengths, and the parts kept out of the open release: Context-IR, H3-Regenerate-2K and the sparse-attention implementation.