HappyHorse AI Video Prompt Guide: Dialogue, Sound & Motion

In early April 2026 an unnamed video model showed up on the Artificial Analysis Video Arena and went straight to the top of both the text-to-video and image-to-video boards. Nobody knew who made it. A few days later, on April 10, Alibaba confirmed it was theirs. The model was called HappyHorse (you will also see it written “Happy Horse”), and it is built around one idea most video models still treat as an add-on: the picture and the sound are generated together.

That changes how you should prompt it. A silent, mood-heavy prompt that works fine elsewhere leaves half of HappyHorse unused. This guide covers what the model actually is, what the current version does, and how to write prompts that get clean motion, synced dialogue and believable sound, with 10 copy-ready examples at the end.

What HappyHorse Is (and Which Version to Use)

HappyHorse is Alibaba’s closed-weight video generation model, built by a team in its ATH business group. It is a separate product line from Alibaba’s Wan models. If you are running the open-source Wan 2.1 locally, our Wan prompt guide is the one you want.

The timeline so far:

HappyHorse 1.0 went into limited beta on April 28, 2026, through the official HappyHorse site, the Alibaba Cloud Model Studio API and Alibaba’s Qwen app. Alibaba listed text-to-video, image-to-video and subject-to-video generation, plus video-to-video editing, at up to 1080p.
HappyHorse 1.1 arrived in late June 2026 on the same channels. Alibaba’s release notes focus on four things: better consistency when you give it multiple reference images, stronger motion, better instruction following and tighter audio-visual sync.

Use 1.1 for anything with dialogue or sound. That is where most of the upgrade went.

One honest note on where it stands in September 2026: HappyHorse 1.1 is still the current version, but it no longer tops the charts. On Artificial Analysis’s text-to-video leaderboard with audio, newer models such as Gemini Omni Flash, Alibaba’s own Wan 3.0 (released in August) and MiniMax H3 now rank above it. That does not make the prompting habits below any less useful. Quoted dialogue, a named language and layered sound work the same way in every audio-native model.

The specs that matter for prompting (as listed by Alibaba and the platforms hosting the model):

Clips run 3 to 15 seconds, with 5 seconds as the usual default
720p or 1080p output
Video and audio generated in a single pass: dialogue, ambience, sound effects and music
Lip-sync in seven languages: English, Mandarin, Cantonese, Japanese, Korean, German and French
Reference-to-video with multiple reference images (hosts running 1.1 list up to nine, labeled character1 through character9)
Vertical, square and widescreen aspect ratios, including 9:16 for Shorts and Reels

How HappyHorse Reads a Prompt

The best way to think about a HappyHorse prompt is a short note from a director to a crew that includes the sound person. It needs to answer five questions:

Who is on screen? A concrete description, not “a person.”
What moves? The action and how it changes over the clip.
Where is the camera? A shot size and at most one camera move.
What do we hear? Dialogue, sound effects and background ambience.
What format? Length and aspect ratio, which you usually set in the interface.

The piece people skip is number 4. Platforms hosting the model note that prompts with no audio direction tend to come back with weak or missing sound. If you want a sound, name it.

The other habit to break is writing a photograph. “A moody woman in a rainy cafe, cinematic, beautiful lighting” describes a still image. HappyHorse needs something to happen over time: she wipes the fogged window with her sleeve, looks out, and her face drops. Motion first, mood second.

Dialogue: Quote It and Name the Language

HappyHorse treats quoted text as the line to be spoken and matches mouth shapes to it. Three rules make lip-sync noticeably cleaner:

Put the exact line in quotation marks. Paraphrased speech (“he tells her he is leaving”) gives the model nothing to sync to.

Name the language. The model supports seven lip-sync languages, but it should not have to guess which one you want. Write “says in Japanese” or “replies in German.”

Keep it short. One or two sentences fit a 5-second clip. Normal speech runs at roughly two to three words per second, so a 15-word line already fills most of a 5-second shot. If you cram in more, the model either rushes the delivery or cuts the line off.

For two speakers, give each a time window. Morphic’s HappyHorse 1.1 guide shows this pattern: “0-4s: character1 says in French…, 4-8s: character2 laughs and replies…” Timed turns stop both characters from talking over each other.

One known weak spot: lip-sync gets less reliable in crowded frames and fast-cut compositions. Two people at a table works. Six people at a party, all talking, does not.

Sound: Build It in Layers

Think like a sound designer and stack up to three layers:

Foreground: the dialogue or the one key sound effect (a door slam, a glass shattering)
Midground: sounds from the action itself (footsteps, keyboard clicks, oil sizzling)
Background: the room or location tone (cafe murmur, distant traffic, wind, rain on a roof)

You do not need all three every time. A product shot might only need the midground: “the soft click of the cap snapping shut.” A dialogue scene usually wants all three. Add a texture word to each sound (muffled, crisp, distant, echoing) and the model has a much clearer target.

If you do not want music, say “no music.” Otherwise the model may add a score you did not ask for.

Camera and Motion

HappyHorse was built with a cinematic look in mind: shallow depth of field, atmospheric light, fine texture. You get more of that when your camera direction is specific. “Cinematic camera” is not a direction. “Slow push-in from a medium shot to a close-up” is.

Stick to one camera move per clip. If you need a pan and then a push-in, generate two clips and cut them together. Our camera movement cheatsheet has the full vocabulary if you need it.

Reference Images: Label Them

For character consistency across clips, use reference-to-video and refer to each upload by its label in the prompt: “character1 hands the box to character2.” Unlabeled references are one of the most common causes of the wrong face on the wrong body. For a broader workflow on keeping one character stable across a whole series, see our consistent characters guide.

10 Copy-Ready HappyHorse Prompts

Each of these is written for a single generation. Swap the details, keep the structure.

1. Street interview (vertical, English dialogue)

Vertical 9:16 handheld phone video on a busy downtown sidewalk at midday. A young man in a gray hoodie holds a microphone toward a woman in her sixties carrying grocery bags. She laughs, shifts the bags to one hand and says in English: "Honestly? I just want the bus to show up on time." Medium close-up, slight handheld sway. Audio: her voice clear and close, traffic and chatter in the background, a bus brake hissing in the distance. No music.

2. Two-person cafe scene with timed turns

character1 and character2 sit across a small round cafe table by a rain-streaked window, warm lamp light. 0-4s: character1 slides a folded letter across the table and says in English, "Read it before you say anything." 4-8s: character2 unfolds it, eyes widening, and whispers, "You're serious." Static medium two-shot, shallow depth of field. Audio: soft cafe murmur, rain tapping the glass, the rustle of paper.

3. Product ad with one clean sound

A matte black wireless earbud case sits on a white stone surface in soft morning light. A hand enters frame, flips the lid open with a thumb, and lifts one earbud out. Slow push-in to a close-up of the open case. Audio: a crisp magnetic click as the lid opens, faint room tone, no music, no voice.

4. Mandarin food vendor

Night market in Taipei, steam rising from a wok under a bare bulb. A middle-aged street cook tosses noodles in a flaming wok, then looks straight at the camera and says in Mandarin: "最后一份了，要不要？" Medium shot, slight low angle, handheld. Audio: loud oil sizzle and metal spatula scraping, crowd chatter, a scooter passing behind.

5. Horror doorway beat

A dark hallway in an old apartment, only a thin strip of light under a closed door at the end. A woman in a white T-shirt walks slowly toward the door, stops, and presses her ear to it. From the other side, a child's voice whispers in English: "Don't open it." She freezes. Slow push-in from behind her shoulder. Audio: floorboards creaking under her steps, a low electrical hum, the whisper close and dry. No music.

6. Pet comedy (no dialogue, reaction sounds)

A fluffy orange cat sits on a kitchen counter next to a full glass of water. It stares at the camera, then slowly pushes the glass with one paw until it tips off the edge. One continuous take, static eye-level shot, bright kitchen daylight. Audio: the glass sliding on the counter, a splash and shatter on the tile floor, a person off-screen gasping "No no no!" in English.

7. Japanese morning routine vlog

Vertical 9:16 selfie-style video in a small Tokyo apartment at sunrise. A young woman in a cream sweater holds a cup of coffee near the window, smiles at the camera and says in Japanese: "おはよう、今日は雨みたい。" She turns the camera toward the window where rain streaks the glass. Handheld, soft natural light. Audio: gentle rain, a kettle clicking off in the background, her voice soft and close.

8. German product review

A man in his thirties sits at a wooden desk with a camera tripod and ring light behind him. He holds up a small mechanical keyboard, taps a key twice and says in German: "Hört ihr das? Das ist der Unterschied." Medium close-up, static shot, warm practical lighting. Audio: two sharp keyboard clicks, quiet room tone, his voice clear and slightly amused.

9. Nature documentary moment

Dawn over a misty wetland. A gray heron stands motionless in shallow water, then strikes down fast and lifts its head with a small silver fish. Telephoto shot, static, shallow depth of field with the reeds soft in the foreground. Audio: water dripping from the heron's beak, frogs and distant birdsong, no narration, no music.

10. Reference-to-video brand character

character1 stands behind the counter of a small bakery holding a tray of fresh croissants. She sets the tray down, brushes flour off her hands, looks at the camera and says in French: "Tout juste sortis du four." Slow push-in from medium shot to medium close-up. Warm window light, golden tones. Audio: the tray landing on the wooden counter, a doorbell chime as a customer enters, soft street noise outside.

What HappyHorse Still Struggles With

On-screen text. Signs, labels and titles often come out garbled. Add text in your editor instead of prompting for it.
Crowded dialogue. Lip-sync works best with one or two speakers in frame.
Long stories. Each generation tops out at 15 seconds. Plan multi-clip sequences shot by shot, the way our storytelling guide lays out.
Silent prompts. If you want sound, you have to ask for it.

Where Your Ideas Come From Matters More Than the Model

Every HappyHorse prompt above follows the same pattern: a concrete subject, a visible action, one camera move, and named sounds. The hard part is not the structure. It is coming up with ten different scenes worth generating when you have a posting schedule to keep.

That is the job LzyPrompt does. You type a plain idea like “cat knocks glass off counter,” and it gives you 10 prompt variations, each with the camera style, setting, lighting, action beats and sound cues already written, plus a caption ready to post. The prompts are plain text, so you paste them into HappyHorse, Veo, Kling or whichever model you use. If you want a structured starting point, our Veo 3 prompt generator and Kling AI prompt generator pages show how the same idea becomes a full prompt, and the pet video prompts and street interview prompts pages have more scene ideas to adapt.

Try LzyPrompt free for 7 days and turn one idea into ten HappyHorse-ready prompts in a few seconds.