🎯 MiniMax H3 Omni Reference: The 12-File Budget

The reference system is the part of H3 that most repays reading the docs, because the headline numbers are a trap. Nine images, three videos, three audio clips — that adds to fifteen, and fifteen is not allowed. The real ceiling is twelve.

By the numbers

Reference type	Per-type cap	Duration rules
Images	≤ 9	256–5760 px per side, aspect 0.4–2.5
Video clips	≤ 3	2–15 s each, ≤15 s total
Audio clips	≤ 3	2–15 s each, ≤15 s total
All types combined	≤ 12 files	—
Formats	JPG/JPEG/PNG/WEBP/HEIC/HEIF · H.264 or H.265 + AAC/MP3 · WAV/MP3
Size caps	Image ≤30MB · Video ≤50MB · Audio ≤15MB · Body ≤64MB

The 12-file total is the constraint that shapes everything. You can have nine images or a deep multi-modal stack, not both.

Three budgets that actually work

Identity-heavy (9 + 2 + 1 = 12). Nine images carrying face, wardrobe and set from multiple angles; two short video clips for motion; one audio clip for voice. This is the configuration for episodic work where a character has to survive across many shots.

Motion-led (6 + 3 + 3 = 12). Six identity images, the full three video clips for camera and movement transfer, three audio clips. Use when you are matching an existing camera language more than locking a face.

Voice-first (8 + 1 + 3 = 12). Eight images, one video, three audio clips. For talking-head and dialogue coverage where the voice has to stay recognisable and the camera is simple.

Whichever you pick, spend image slots on different jobs rather than near-duplicates — the same discipline our consistent characters guide applies across models. Front, three-quarter and profile of the face. One clean wardrobe shot. One set or location plate. A signature prop if the character has one. Nine slots go fast.

Give every reference exactly one job — in the prompt

The reference files are only half the mechanism. The prompt has to assign them:

References: image 1 = her face and build. image 2 = the jacket.
image 3 = the gantry location. Keep the face from image 1 exactly and the
jacket from image 2 exactly. Place her in the location from image 3.

Two references that disagree about a hairstyle do not produce either hairstyle — they blend into a third wrong one. That is the single most common cause of "the references aren't working."

Reference hygiene matters more than reference count: even lighting, unobstructed faces, consistent wardrobe across the image set, clean noise-free audio. Six good references beat nine mixed-quality ones.

Reference video does motion, not appearance

A reference video transfers camera path and pacing. It will also transfer appearance unless you tell it not to:

References: video 1 = camera movement reference. Follow its camera path and
pacing exactly. Do not copy the appearance from video 1, only its motion.
Image 1 = the subject. Image 2 = the environment.

This is the closest thing H3 has to a motion-control feature, and it is worth knowing that reference video seconds bill as input on top of your output seconds. A 10-second motion reference on a 5-second render bills 15. Trim reference clips to the shortest span that carries the move — the 2-second floor exists for a reason.

Reference audio carries a voice, but never travels alone

Three audio clips, 2–15 seconds each, 15 seconds total. Each one must be accompanied by at least one image or video reference; an audio-only set is rejected outright.

References: audio 1 = the character's voice. Image 1 = the character.
He speaks in the voice from audio 1: "We're not going back."

Because H3 generates audio in the same pass as the picture — the capability that separates it from Kling 3.0 — a voice reference means the dialogue comes back in that timbre over matching mouth movement. That is the piece that removes an entire ElevenLabs-plus-lipsync stage from a dialogue pipeline.

Watch the total: three 6-second clips is 18 seconds and over the limit. Two 7-second clips is fine.

Frames or references — never both

Image-to-video and reference-to-video are mutually exclusive. A request containing both a first_frame and a reference_image is rejected.

In practice that means choosing between a locked opening composition and a locked identity. Most character work picks references and describes the opening framing in the prompt instead. If the exact first frame genuinely matters more than identity consistency, use the frame mode and accept more drift.

Pre-flight your assets

Two limits fail server-side and cost you a round trip, so check locally: images must be 256–5760 px on each side with an aspect ratio between 0.4 and 2.5, and every reference clip must be 2–15 seconds. A tall 9:16 crop at 0.56 passes; a 10:1 banner at 10.0 does not.

Our own client validates geometry with Pillow and clip durations with ffprobe before submitting, which turns a paid rejection into an instant local error. The Python setup guide has the request-side detail.

How we assessed this

Every limit here is read off MiniMax's own video-generation documentation and the v2 create endpoint reference, cross-checked against the morphic model spec page. The 12-file total is stated explicitly by MiniMax as "mixed input is capped at 12 files in total."

Worth flagging because it circulates: some pages cite the per-type maximums (9 + 3 + 3) and conclude fifteen files are allowed. They are quoting the per-type caps correctly and missing the cross-type ceiling. Both constraints apply.

This is a documentation-led workflow guide, not a render review. We have not published drift rates or reference-count benchmarks, because measuring those honestly needs volume on a model that is days old.

What didn't make the cut

Per-slot weighting. There is no reference-strength parameter on the v2 API — no percentage slider, no priority field. Any guide describing one is describing a different product. Your only weighting mechanism is the prompt language assigning jobs to files.

For the same problem on the other frontier model, where the reference budget is four times larger, see Seedance 2.5 character consistency.