Skip to content
CONCEPTS
4 min readЧитать на русском

Happy Horse 1.0 — Prompt Rules

You've used Seedance 2 enough that JSON shot prompts come naturally. You try the same instincts on Happy Horse 1.0 — same brief, comparable structure, same vocabulary — and the output comes back generic. Bland subject. Vague camera. Adequate lighting that could be anything. The model didn't ignore the prompt; it just didn't follow rules you didn't know it had.

What follows are those rules. Happy Horse rewards a different prompt shape than Seedance — load-bearing slot order, an explicit duration-plus-beats prefix, exactly one camera move, language purity. Learn them once and the gap closes.

Happy Horse 1.0 is a 15B-parameter open-source text-to-video / image-to-video model with native synchronized audio and multilingual lip-sync. Ranked #1 in blind T2V / I2V voting at time of writing; positioned as a Seedance 2 competitor. The rules below are the prompt-engineering shape that consistently survives generation.

Capabilities & limits

| Dimension | Value | |---|---| | Modes | T2V, I2V | | Audio | Joint video + synchronized audio generation | | Lip-sync | Multilingual | | Max duration | 8s per clip (hard cap) | | Max resolution | 1080p | | Default aspect | 16:9 (others supported) | | Quality tiers (host-exposed) | "Pro", style preset "Balanced" |

Canonical 8-slot structure

Order is load-bearing — Happy Horse has front-loaded attention; the first ~40 tokens carry disproportionate weight.

1. Subject          ← primary entity, front-loaded
2. Action           ← what the subject does
3. Scene            ← where, props, environment
4. Camera movement  ← exactly ONE primary move
5. Lighting         ← source, time of day, quality
6. Visual style     ← cinematic / cel-shaded / studio / documentary
7. Mood             ← emotional palette
8. Platform format  ← aspect, pacing, output context

Skip a slot only if you have nothing concrete; never pad with generic adjectives. Output as one flowing paragraph (not headed sections) but preserve the order strictly.

Hard rules

  1. Front-loaded attention. Primary subject in the first words. Motion descriptors second. Stylistic modifiers last. Never bury the subject behind an adjective stack.
  2. Specificity over implication. Happy Horse follows literally and does not infer missing camera or lighting — silence yields generic output. This inverts the "deliberate omission" patterns that work with image models like Nano Banana 2 (which fill gaps via the image channel). For HH: omission = bug.
  3. Beat structure mandatory. Open every prompt with a duration flag and a per-beat plan. Example: 8s duration. First 3s: close-up on hands. Final 5s: camera pulls back to reveal full scene.
  4. One primary camera move per clip. Combining dolly + orbit + rack focus causes drift, simplification, or model fallback. Pick one.
  5. Language purity. English or Chinese — never hybrid. Mixed-language prompts measurably degrade. Pick the language at the top and hold it.
  6. N ≥ 2–3 sample discipline. Output variance is high even with identical input. Always generate 2–3 takes and pick.

Vocabulary palettes that survive generation

Drawn from the official template gallery (cinematic / scene / product). These specific phrasings reliably translate.

  • Camera moves: slow crane move, smooth side tracking, macro transitions, rotating pedestal, controlled backlight pull.
  • Lighting: golden-hour, lantern light, studio reflections, moody dust, clean cel shading.
  • Pacing: elegant pacing, premium ad pacing, smooth transitions.
  • Tone: premium cinema tone, emotional evening palette, controlled atmosphere.

Anti-patterns

  • Multi-camera choreography stacked in one clip (dolly + orbit + rack focus).
  • Language mixing (English/Chinese hybrid prompts).
  • Missing duration / beat structure → erratic pacing.
  • Generic adjective stacks without an explicit camera + light anchor.
  • Burying the subject behind style modifiers ("Cinematic, moody, premium 8s shot of a horse running…" — subject is too late).

Comparison vs. Seedance 2 (where rules diverge)

| Axis | Seedance 2 | Happy Horse 1.0 | |---|---|---| | Output contract | Compact JSON (refs, g, s[]), ≤3500 chars | Single flowing paragraph, slot-ordered | | Camera/visual split | Camera in c, visual in p (separate fields) | One paragraph, single camera move embedded | | Reference handling | refs[].s CSV scene-mapping, grid overlay tricks | I2V mode: image is the reference, no JSON map | | Omission | Heavy use of deliberate omission | Inverted — must specify camera and light explicitly | | Beat structure | Implicit per-scene | Explicit duration + per-beat plan in prompt body | | Language | English mandatory regardless of input | Pure English or pure Chinese | | Camera moves per scene | Multi-stage motion encouraged | Exactly one |

Worked example

Brief: "Luxury beverage launch, glass bottle, energy theme, premium ad."

Prompt:

8s duration. First 3s: glossy black bottle stands centered on a rotating pedestal. Final 5s: a horse-shaped energy trail spirals up around the bottle. Studio interior with deep gradient backdrop. Slow crane move arcing in from low angle. Studio reflections, controlled backlight, moody dust catching the rim light. Premium cinema tone, clean cel-shaded energy trail, glossy product surfaces. Confident, kinetic, high-end. 16:9, premium ad pacing.

Self-checks:

  • ✅ Subject ("glossy black bottle") in first 10 tokens
  • ✅ Duration + beat structure at the open
  • ✅ Exactly one camera move (slow crane move)
  • ✅ Pure English
  • ✅ Camera AND light explicitly specified
  • ✅ Mood + format slot present

Sources