Happy Horse 1.0 is Alibaba's newest video model — physics-aware, multi-shot, and unusually faithful to long structured prompts. Below: how to write for it, what makes it different, and prompt patterns you can copy.
Happy Horse 1.0 is the latest video generation model out of Alibaba's ATH AI Innovation Unit. The headline traits: it follows long, structured prompts more literally than most video models, it animates motion that obeys physics instead of drifting through it, and it currently sits at #1 on the Artificial Analysis Video Arena leaderboard for blind side-by-side preference.
On PixelDojo, Happy Horse runs in two modes — text-to-video and image-to-video — at 720p or 1080p, with clip lengths from 2 to 15 seconds. Pricing is 4 credits per second at 720p and 6 at 1080p, billed only on success.
This guide walks through what the model is good at, a prompt framework that takes advantage of those strengths, and ready-to-copy examples for the most common shot types.
Six things Happy Horse does meaningfully better than the previous generation. Lean into these in your prompts and you'll see the difference immediately.
Believable, not floaty
Trained on a motion engine that respects real-world physics. Collisions have weight, fabric folds, and limbs swing on natural arcs. Fewer of the warped, slow-motion-y movements that plague older video models.
Cuts inside one prompt
Happy Horse is unusually good at following long, structured prompts that describe more than one camera angle or beat. You can write a three-shot mini-sequence and the model will actually cut between them inside a single 10-15s clip.
Pleasant out-of-the-box look
Outputs land with sharp textures, balanced contrast, and pleasing color science before you ever touch a LUT. Skin tones, fabric weaves, and surface materials read naturally — fewer plastic-y faces and muddy blacks.
Your starting frame stays itself
When you animate a still, Happy Horse holds the subject's identity, outfit, and lighting through the entire clip. Faces don't morph mid-scene and props don't drift — well-suited for character-driven shots.
EN, ZH, JA
Native understanding of English, Chinese, and Japanese prompts. You can mix languages in a single prompt — useful if you have a precise term in one language that doesn't translate cleanly.
#1 on Video Arena
Currently sits at the top of the Artificial Analysis Video Arena leaderboard for blind side-by-side preference. Translation: humans pick its outputs over most competing models.
A reliable structure for Happy Horse prompts. Not all five blocks are required, but in this order they let the model build the shot the way a director would.
Who or what is in frame, and what they're doing right now. Be specific about clothing, age, posture, expression — these anchor the rest of the scene.
Where and when. Time of day matters more than you'd think — "golden hour", "blue hour", "overcast noon" each give the model a complete lighting direction.
Name the move explicitly. Push-in, pull-back, side-tracking, low-angle dolly, slow orbit. Static shots are fine too — just say "locked off" or "static camera".
What changes during the clip. For multi-shot prompts, number them: "Shot 1… Shot 2…". For single shots, name the micro-action (a blink, a turn, a step forward).
Closing tags steer the look: "hyper-realistic", "cinematic", "anime style", "film grain", "8k", "shallow depth of field". Happy Horse responds to film vocabulary.
Six prompts spanning the most common shot types. Each one is ready to copy and run as-is — or use the takeaway as a starting point for your own version.
Slow push-in, golden hour
Takeaway: Open wide, push in. Spelling out the camera move ("slowly pushes in to a medium shot") gives Happy Horse a clear arc to animate against — much stronger than a static shot of the same subject.
Multi-shot inside one prompt
Takeaway: Number your shots. Happy Horse handles 2-3 explicit camera changes well inside a 10s clip — give each beat its own camera angle and physical action so the cut points are unambiguous.
Single locked shot
Takeaway: For portrait work, name the micro-actions you want (blink, half-smile, leaf drift). Happy Horse renders those small motions with believable timing instead of either freezing the subject or over-animating them.
Stylized, hand-drawn feel
Takeaway: Lead with the style descriptor ("Studio anime style", "cel-shaded", "thin clean linework"). Happy Horse holds 2D aesthetics steadily through camera moves — you don't need to repeat the style every sentence.
15s commercial structure
Takeaway: For ads, structure as detail → action → hero shot. Premium adjectives like "clean, premium" steer color and pacing without naming brands.
Environment-first prompting
Takeaway: When the environment is the star, build it before the subject. Happy Horse will keep small distant figures coherent if the world around them has weight, light direction, and a named time of day.
Happy Horse is unusually consistent when animating a still. These habits get the most out of I2V mode.
The model can see your reference. Don't restate what's already in frame — focus the prompt entirely on what should happen next.
Sharp focus, well-lit, no compression artifacts. Happy Horse holds detail well, but it inherits the source image — garbage in, garbage out.
Crop the source to the orientation you want (16:9, 9:16, 1:1) before uploading. The model adapts to the source ratio rather than a separate setting.
"Slow push-in", "slight handheld sway", "the wind picks up". Even gentle named motions read more naturally than "make it move".
If you need a specific face, build it as a still in your favorite image model first, then animate with Happy Horse — identity drift is dramatically lower than T2V from scratch.
5 seconds is the sweet spot for "alive portrait" style clips. Longer durations let drift accumulate; multi-shot prompts shine more in T2V.
Happy Horse Reference-to-Video takes 1–9 reference images plus a prompt. Each image is mapped to a token — character1, character2… — and the model preserves each subject's identity through motion. Use it for multi-character scenes, product placement with a specific reference object, or locking a face across multiple shots.
Three references → one shot
Takeaway: Number every reference. Order matches your upload grid: first image is character1, second is character2, etc. The model places each one in the role you've named.
One reference → cinematic restage
Takeaway: With a single reference, you can drop a known face or character into any environment. Identity stays steady; outfit, lighting, and setting follow the prompt.
Sharp, well-lit photos with the subject filling most of the frame. Avoid heavy crops, extreme angles, or busy backgrounds — the model picks up everything that's in the reference.
Don't combine multiple characters into one image. Upload them as separate references and label them character1 / character2 in the prompt.
If character1 has red hair, you don't need to write "red-haired" in the prompt — the model already knows. Save your prompt budget for action and camera direction.
Reference doesn't have to be a face. Clothing, vehicles, weapons, signature objects all work. Tag them with character tokens just like people.
Happy Horse Video Edit transforms an existing clip with a text instruction. Two modes happen organically depending on what you write: style transfer (a complete restyle of the whole video) and local replacement (swap an outfit, prop, or object using a reference image). Same look-and-feel as the base model — physics-aware, multi-shot-aware — but applied to footage you already have.
No reference, prompt-only
Takeaway: For style transfer, name the medium (watercolor, charcoal, oil, anime, claymation), the palette, and one or two texture cues ("visible brushstrokes", "grainy paper"). Tell the model to preserve motion and composition.
Named-reference style
Takeaway: Naming a studio or director shorthand ("Studio Ghibli", "Wes Anderson", "Roger Deakins") locks in a familiar visual language without you having to enumerate every detail. Combine with "maintain the original camera move" to preserve the shot's intent.
"Watercolor", "oil painting", "charcoal sketch", "Studio Ghibli", "comic book", "clay animation". One specific term beats five vague adjectives.
For "swap the sweater", "change the car to this one", or "give the character this hairstyle", upload the target object as a reference image and write the swap as a single instruction.
Default Auto lets the model decide. Pick Original when your input clip has dialogue, music, or signature audio you need to keep — the visuals change but the audio track is preserved verbatim.
MP4 or MOV, 3-60 seconds, max 100MB. Output is up to 15 seconds; longer source clips are clipped to the first 15. For best results, use a clip with a clear primary subject.
Write "Shot 1: …Shot 2: …Shot 3: …". Happy Horse treats each numbered block as its own camera angle and will cut between them inside one clip.
Be explicit. "Slow dolly in", "side-tracking shot", "low-angle push", "locked-off static". Vague motion = inconsistent results.
"Golden hour", "blue hour", "overcast noon", "night, neon-lit". These trigger complete lighting setups instead of you having to name colors.
Closing the prompt with "hyper-realistic, cinematic, 8k, shallow depth of field" pushes the model toward its highest fidelity output.
For single-shot prompts, give the camera one job. Mixing too many simultaneous moves (zoom + orbit + pan) usually produces messy motion.
Specific Japanese terms (e.g. 木漏れ日 for dappled sunlight) can sharpen mood that English struggles to name precisely. Happy Horse reads them natively.
Iterate at 720p × 5s (cheapest path), then re-roll the winning prompt at 1080p × 10-15s for the final.
Open Advanced Settings and pin a seed. You can then tweak the prompt and isolate what each change does without random variation getting in the way.
Resolution
720p · 1080p
Aspect derived from source / chosen ratio
Duration
2 – 15s
5s for portraits, 10-15s for multi-shot
Cost
4 cr/s · 6 cr/s
720p · 1080p — refunded on failure
Languages
EN · ZH · JA
Mix freely in one prompt
Happy Horse 1.0 is Alibaba's latest video generation model. On PixelDojo it ships in four flavors: text-to-video, image-to-video, reference-to-video (1-9 character/prop images), and video edit (style transfer + local replacement on existing clips). All run at 720p or 1080p with clip lengths up to 15 seconds. The base model currently sits at #1 on the Artificial Analysis Video Arena leaderboard for blind preference.