Hailuo’s H3 model (branded in-app as Hailuo 3.0) changes what marketers should expect from short AI video. The model now accepts text, images, video and audio together and can return 4–15 second clips with integrated stereo sound and a regenerate path to 2K. That means old habits—stacking stylistic tags, long adjective lists, or treating the prompt like a wish list—no longer produce reliable results. If you need believable product detail, continuity across cuts, or a match between picture and sound, treat the prompt like a production brief.

1. Choose the right input mode

Hailuo AI offers three inputs; pick the one that preserves what must be accurate on screen.

  • Text only — Fastest and suitable for loose, atmospheric shots that won’t be inspected closely: streetscapes, surface textures or background room tone.
  • Start frame (image-to-video) — Use this when the product, label or packaging must be faithful. The model begins from a real photo, so proportions and printed details are far more likely to stay true than if you describe them in text.
  • Reference mode — When continuity matters (a presenter’s face or voice, a repeatable camera move, or a product shown from multiple angles), upload reference files. H3 accepts multiple photos, short video clips and audio to anchor identity and motion.

Generate in the intended aspect ratio (vertical for Reels, widescreen for display) rather than cropping later; framing that looks right at generation can break when you reframe after the fact.

2. Write the prompt as a timed shot list

H3 treats prompts like production instructions. Replace long modifier lists with a short scene line and a numbered, time-stamped shot list. Each shot should state framing, camera move and a single action. Put sound notes in the same block.

Minimal structure to copy:

  1. Scene: place, time of day, lighting, who’s present.
  2. [Shot 1] Framing, camera action, one clear action.
  3. [Shot 2] At 00:04.000, cut to — framing, camera and action.
  4. Soundscape: room tone and key Foley hits.
  5. Music: specify or explicitly state none.

Example (10s bakery ad):

A small, sunlit bakery counter in the morning. A woman in her 40s in a flour-dusted apron stands behind it.

[Shot 1] Medium shot, locked camera. She slides a tray of croissants onto the counter and looks up with a small smile.

[Shot 2] At 00:04.000, close-up of one croissant as her hands break it open. Steam rises.

[Shot 3] At 00:07.000, cut back to medium. She says, “Baked at six. Gone by ten.”

Soundscape: quiet café room tone, metal tray clatter, soft street noise.

Music: none.

3. Plan sound as part of the shot

H3 generates stereo audio in the same pass as video. Specify room tone, single-frame Foley hits (tray clatter, door close), and whether music runs. Unspecified audio is guessed by the model and can break perceived realism faster than small visual imperfections.

4. Use reference files for identity and continuity

Supply photos, short clips and audio when a face, voice or product needs to remain recognisable across cuts. Reference mode is the safest option for UGC-style presenter spots or multi-angle product views. A practical rule: the more your message depends on a detail, the more that detail should come from a supplied file rather than a written tag.

5. Iterate: draft at lower resolution, then regenerate for 2K

Use a quick 768p draft to lock timing, pacing and camera moves. Once the edit and timings are correct, use H3’s regenerate step to produce a 2K version. For start-frame workflows, begin with a sharp, evenly lit photo already cropped to the intended aspect ratio to avoid surprises after upscaling.

Keep shots few and precise. The official guidance and third‑party builders generally limit prompts to a handful of timed shots—this focuses the model and reduces motion artifacts caused by too many rapid cuts.

6. Practical checklist and prompt hygiene

  • Decide mode: text / start frame / reference before you start.
  • Set the aspect ratio to match your delivery platform.
  • Write a one-line scene description, then 1–4 timed shots with framing and one action each.
  • Include a concise soundscape and state music preferences explicitly.
  • Supply reference files for any persistent subject or product.
  • Draft at 768p; regenerate to 2K only when edit and timing are locked.
  • Avoid long lists of stylistic tags; prefer precise shot instructions.

MiniMax publishes an official H3 prompt guide and third‑party tools such as ComfyUI have complementary documentation; those lists are useful for exact labels if you’re integrating via API. The core rule remains the same: scene first, then timed shots and sound.

What to test next: run A/Bs that change only one variable—start frame vs text-only, or with vs without reference images—to measure how identity and continuity improve when you supply real files. Prioritise sound and timing over extra visual adjectives. With H3, a concise, well-ordered brief produces more believable results than a longer, vaguer prompt.