MiniMax H3 vs Gemini Omni 1.1 vs Wan 3.0: three audio-video models, three prompt dialects

Three video models arrived on LUVI within a week of each other in September, and they share the headline — sound generated together with the picture, dialogue included. They do not share a prompt language. We wrote a prompting guide for each from the vendors' own documentation; this is the comparison that fell out of that work.

The numbers

• Base price on LUVI. H3 from 2,080 credits for 8 s (H3 Max 1,216). Omni 1.1 Flash from 2,200 credits for 10 s. Wan 3.0 from 2,000 credits for 5 s (Prime 2,800).
• Native resolutions. H3: 480P, 768P, 2K. Omni: 360p, 720p; 1080p and 4K are Google's upscaler. Wan: 480p, 720p, 1080p.
• Clip length. H3: 4–15 s. Omni: 3–10 s, extendable to 40 s. Wan: up to 30 s.
• Modes on LUVI. H3: text, image, reference. Omni: text, image, reference, edit, extend. Wan: text, image, reference.
• Dialogue languages. H3: 11 (Turkish not among them). Omni: English evaluated, others may work. Wan: Chinese and English.
• Reference media. H3: up to 9 images, 3 clips, 3 audio, 12 files total. Omni: images and up to 3 short clips, no audio. Wan: up to 10 images, 5 clips, 5 audio, 20 files.

MiniMax H3: why is the prompt a document?

H3 does not want a caption. It wants three named fields — integrated_multimodal_description, overall_soundscape, non_diegetic_music — and it wants all three every time. Leave an audio field out and the model does not fall silent; it invents. Camera is prose built from twelve exact verbs and four modifiers ("The camera pushes in with small amplitude at slow speed toward…"), cuts are timestamped to the millisecond ("[Shot 2] At 00:03.500, the camera cuts to…"), dialogue sits in a tag with its language: <d>[English] I get off at the next station.</d>. A proper noun — a celebrity, a film, a brand — blocks the whole request. Bracketed camera commands from the older Hailuo line are dead syntax here.

Best at: scripted multi-shot with dialogue, and reference packs that lock a face and a voice at once.

Gemini Omni 1.1: why must a caption ask for a single take?

Omni is the opposite temperament: a 45-word caption in plain prose is the ideal prompt, and the model fills in the world. The trap is that Omni cuts by default — ask for a lakeside shot and you may get a three-shot montage of one. Every single-take prompt needs a continuity clause in the first sentence ("In a single unbroken scene…"). Sound goes in its own sentence behind "Sound design:", music has to be asked for, and dialogue uses a colon with no quotation marks — quotes burn the words into the frame as on-screen text. There is no negative field; exclusions are terminal fragments ("No dialogue. No text overlay on screen."). Edits are one clause plus "Keep everything else the same."

Best at: quick, cheap, natural-looking clips; conversational edits; extending a clip you already have.

Wan 3.0: what is the additive formula?

Alibaba publishes the prompt as a formula and the model reads it in that order: Entity + Scene + Motion, then Aesthetic control + Stylization, then Sound (voice = lines + emotion + tone + speed + timbre + accent; effect = source + action + ambience; music = score + style), then multi-shot as "Shot 1 [0–3 s] … Shot 2 [4–6 s] Hard cut transition, fixed camera…". Uploaded media is addressed by name and used as a noun: "The cat in Image 1 plays in the room in Image 2." Its own do-not list is short and worth memorising: no real people by name, no rapid scene changes inside one clip, no shot built on exact text or word-exact lip-sync.

Best at: long clips (30 s), mixed reference packs with audio, and Chinese-language dialogue.

The rule that breaks each one

• H3: omit an audio field. You get invented sound, not silence.
• Omni: forget the continuity clause. You get cuts.
• Wan: name a real person. The request fails.

Which one should I pick?

A dialogue scene with a locked cast: H3. A fast draft or an edit of something that exists: Omni. A long clip or a reference pack that includes a voice: Wan 3.0. All three guides are on the model pages — MiniMax H3, Gemini Omni, Wan — and the Workspace applies them when the guide toggle is on. For how the credit prices above are built, see how credits work.