The first time I tried to put two specific people in the same AI clip, I burned an afternoon. I described them in words — hair, jacket, height, the works — and got back two strangers who changed faces between takes. Prompts alone can describe a type of person. They cannot pin down this person.
Reference images fix that. And the reason I'm writing this now: Happy Horse 1.1 accepts up to 9 reference images in a single generation. That's enough to hand the model a character, a second character, a product, and a location, and ask it to compose a scene out of all of them instead of guessing.
This happy horse ai reference images guide covers what refs actually control, when one is better than nine, how to choose source shots that survive generation, and the mistakes that quietly wreck output. Happy Horse AI is the model Alibaba launched anonymously on the Artificial Analysis Video Arena in April 2026 and which currently sits at #1 there for both text-to-video and image-to-video — so the reference behavior below is running on a model that's genuinely good at holding onto what you give it.
What a Reference Image Actually Does
Short answer: a reference image is ground truth for appearance. Your prompt is instruction for behavior.
When you upload a reference, you're telling the model "this is what the subject looks like — don't invent it." Face structure, product shape, fabric, color, and lighting mood get anchored to the pixels you supplied. The prompt then spends its budget on the things an image can't express: the action, the camera move, the atmosphere, the dialogue.
That split is the single most useful mental model here. Most bad reference-based generations I've seen come from people writing prompts that re-describe what's already in the image ("a woman with long brown hair in a red coat") while saying almost nothing about what should happen. You're spending words on a question that's already answered.
There's a second, less obvious effect. Because Happy Horse generates video and audio jointly in a single forward pass, the visual anchor also nudges the audio — a quiet studio and a busy market produce different ambient beds from the same prompt. Say what you want to hear anyway, but expect the imagery to color it.
Rule of thumb: if a detail is visible in your reference, don't describe it in the prompt. If it isn't visible, the prompt is the only place it can come from.
The Up-to-9 Capability: How Many Refs Should You Use?
More references is not automatically better. Each one adds a constraint, and constraints can conflict. Here's how I actually allocate them.
| How many refs | Best for | Why it works |
|---|---|---|
| 1 | Animating a single photo, product hero shot, portrait in motion | Zero conflict — the model has one truth to hold |
| 2–3 | One character from multiple angles, or character + product | Fills in what a single flat shot can't show (profile, back, scale) |
| 4–6 | Multi character ai video, character + product + setting | Enough to compose a scene without pulling in ten directions |
| 7–9 | Complex scenes: several characters, props, wardrobe, environment | Maximum control; needs the most disciplined prompt and consistent sources |
The pattern to notice: references answer "what," never "how many of each." Nine images do not mean nine things in the shot. If you upload three photos of the same actor from three angles, that's one character described thoroughly — not triplets. You control the scene's composition in the prompt: say who is present, where they stand, and who speaks.
Start low. I run the first pass at one or two refs, see what the model got wrong, then add a reference that fixes that specific failure. Load all nine up front and you'll never know which image caused the problem.
How to Pick Good Reference Images
The quality ceiling of your clip is set before you ever write a prompt. A few things matter far more than resolution.
Clear, unambiguous subject. One dominant subject, cleanly separated from the background. Busy group shots teach the model the wrong thing — it may blend two faces or pick the wrong person as the anchor.
Consistent lighting and angle across a set. If you're feeding multiple refs of one character, keep the lighting family the same. A hard-flash selfie plus a golden-hour portrait plus a dim indoor shot gives the model three different-looking people, and it will average them into someone who matches none of your images.
Sharp, uncompressed sources. Screenshots of screenshots, heavily filtered photos, and small thumbnails all carry artifacts the model faithfully reproduces. It doesn't know your JPEG mush is an accident.
Neutral expression and posture for characters. A reference mid-laugh or mid-blink bakes that expression into the anchor and fights whatever performance you asked for.
Show scale when scale matters. If a product needs to look hand-sized, one reference showing it held is worth more than three studio cutouts on white. A well-lit phone photo beats a stylized render here almost every time.
The Browser Workflow
The whole thing runs in the browser. No install, no keys.
- Open the generator. Go to the Happy Horse AI video generator and choose the image / reference input mode rather than the plain text box.
- Upload your references. Start with one or two. Check the generator's current options for how many slots are exposed and which file types it accepts — that UI is the authority, not this article.
- Write a behavior-only prompt. Action, camera, atmosphere, audio. Name each subject the way you'd name them on a set ("the woman in the denim jacket steps forward and speaks") so the model can map your instruction onto the right reference.
- Keep the first pass short. Happy Horse targets 1080p in short clips, generally in the five-to-ten-second range. Iterating at the short end is cheaper than re-rolling a long clip.
- Generate, then diagnose. Look for one specific failure — wrong face, wrong garment, wrong room — and fix it with one change: either a sharper reference or a clearer prompt line. Not both at once, or you learn nothing.
- Lock what works. Once a reference set gives you a character you like, reuse that exact set across every clip in the project. Consistency across shots comes from reusing references, not from re-describing.
Multi-Character and Scene-Accurate Generation
This is where the up-to-9 capability actually pays for itself.
For multi character ai video, give each character their own clean reference and then make the prompt do the blocking: who's on the left, who enters, who speaks first, who reacts. Vague prompts with two character refs are the classic failure — the model has two valid anchors and no instruction about which belongs where, so it may swap them between frames or merge features.
For scene-accurate work, split responsibility. One or two refs for the subject, one for the product or prop, one for the environment. Then write the prompt as if you're describing a shot to a camera operator who has already seen all the photos. Say what changes, not what things look like.
A genuine technical note: because generation is a single joint pass rather than a pipeline of stages, everything — motion, fidelity to your refs, dialogue, Foley — is decided together. A weak reference doesn't just hurt the visuals; it can drag down the whole take. Fix the input and re-roll, don't patch.
To go deeper on writing the behavior half, the prompt guide covers structure in detail, and the image-to-video walkthrough covers the single-photo case end to end.
Common Mistakes
Conflicting references. Two photos of the same character with different hair length, different wardrobe, or different eras. The model doesn't know which one is current — it blends. Pick one look per generation.
Low-quality sources. Blurry, tiny, heavily compressed, or watermarked images. Artifacts carry through. So do watermarks.
Filling all nine slots because they exist. Every extra reference is another constraint the model has to satisfy. If a reference isn't fixing a specific problem, it's adding noise.
Prompting the picture instead of the performance. Re-describing appearance wastes the prompt and can actively contradict the reference.
Cropped-out context. A tight face crop tells the model nothing about body, clothing, or proportion, so it invents all three. If the body matters, show the body.
Expecting a reference to be a storyboard. References anchor appearance. They don't dictate framing, sequence, or timing — that's prompt work. Same goes for fine detail like small text or logos: expect degradation, and composite anything that must be exact.
FAQ
How many reference images does Happy Horse AI support?
Happy Horse 1.1 supports up to 9 reference images in a generation. Check the generator's current options for exactly how the slots are presented and what file requirements apply.
Do I need to use all 9?
No, and usually you shouldn't. One clean reference beats nine mediocre ones. Add references only to solve a specific problem you saw in a previous take.
Can I put two different people in one video?
Yes — that's the main use for a multi-reference setup. Give each person their own clear reference and use the prompt to specify positions, actions, and who speaks. Ambiguous prompts are what cause identity swapping, not the model's ref limit.
Will my character look identical across several clips?
Reuse the exact same reference set for every clip and you'll get strong consistency. Change references between shots and you'll get drift. Treat your reference set like a casting decision — make it once.
Does the reference affect the audio too?
Indirectly. Video and audio are generated jointly in one pass, so the scene implied by your references influences ambient sound. Still state the audio you want explicitly in the prompt if you care about it.
The Bottom Line
Reference images are the difference between "an AI video of a person" and "an AI video of your person." The up-to-9 support in Happy Horse 1.1 means you can compose real scenes — several characters, a product, a location — instead of hoping a paragraph of adjectives lands.
The workflow that works: start with one clean reference, write a prompt that's all behavior and no description, diagnose one failure at a time, and only add references that fix something specific. Reuse the winning set across every shot in the project.
Load your best photo into the Happy Horse AI video generator and run a five-second take before you build a nine-reference setup. You'll learn more from that one clip than from any checklist. When you're ready to scale up, Happy Horse 1.1 is where the multi-reference work happens — and how to use Happy Horse AI covers the rest of the interface.
