Multi-reference generation lets you supply multiple source images — one per character — so the AI model can maintain each person's distinct appearance throughout a video. The model maps each reference to a named token (e.g., [person1], [person2]) in the prompt, binding facial features, body proportions, and clothing to that token. This is fundamentally different from single-reference models that can only preserve one identity, forcing multi-person scenes to hallucinate secondary characters.
HappyHorse accepts 1 to 9 reference images — the highest count in the current AI video model landscape. Each reference binds to a separate character token, so a group scene with 9 distinct people is possible in a single generation.
Reference images are assigned to tokens like [person1] through [person9]. Use these tokens in your prompt to position and direct each character independently: "[person1] hands a coffee cup to [person3] while [person2] waves from the background."
Because all references are loaded in the same generation pass, characters can interact naturally — handshakes, conversations, passing objects. Single-reference models require compositing separate clips to achieve this.
Facial features, skin tone, hairstyle, and clothing remain stable from frame 1 through the end of the clip. No mid-clip identity drift, even with camera angle changes or partial occlusions.
References can be headshots, full-body photos, or stylized illustrations. HappyHorse extracts identity features regardless of the source image format, though front-facing photos with neutral expressions produce the most accurate results.
Go to PonPon Video and select HappyHorse from the model dropdown.
Click the reference image upload area and add 1 to 9 images. Each image should clearly show one person's face — front-facing, well-lit, minimal occlusion. Label or note the order (person1, person2, etc.).
Reference each uploaded image using its token: [person1], [person2], etc. Describe the scene with specific actions for each: *"[person1] sits at a desk typing while [person2] stands behind pointing at the screen. [person3] enters through the door carrying a folder."*
Choose your clip length and aspect ratio. For multi-character scenes, 16:9 widescreen gives more room for character positioning. Longer durations (8–10s) allow more complex interactions.
Click Generate and check that each character matches their reference throughout the clip. If one character drifts, try a clearer reference photo with better lighting or a more frontal angle.
Whether you're a solo creator, an agency, or a brand — every model adapts to how you work.
[person1] and [person2] sit across from each other at a coffee shop table. [person1] gestures while speaking, [person2] nods and smiles. Warm afternoon light through the window. Shallow depth of field. 16:9, 8 seconds.
Model: HappyHorse · References: 2 images · Duration: 8s · Aspect: 16:9
[person1], [person2], [person3], and [person4] stand in a row in a modern office lobby. Each waves at the camera in sequence from left to right. Clean white background, professional attire. 16:9, 10 seconds.
Model: HappyHorse · References: 4 images · Duration: 10s · Aspect: 16:9
[person1] sits at the head of a dining table, [person2] and [person3] on either side, [person4] at the far end. [person1] raises a glass for a toast, others follow. Warm candlelight, rustic wooden table. 16:9, 10 seconds.
Model: HappyHorse · References: 4 images · Duration: 10s · Aspect: 16:9
[person1] in a lab coat hands a product box to [person2] in business casual. [person2] inspects the box and nods approvingly. Clean studio background, soft key light. 16:9, 6 seconds.
Model: HappyHorse · References: 2 images · Duration: 6s · Aspect: 16:9
Short films, web series, and explainer videos with a recurring cast. Upload your character references once and generate consistent scenes across episodes — no continuity errors between shots.
Turn a corporate team photo into an animated introduction video. Upload individual headshots for each team member and prompt a scene where they interact — waving, shaking hands, or presenting together.
Generate personalized family videos for holidays or celebrations. Upload photos of family members and create scenes like a family dinner, a birthday party, or a group walk in a park — each person recognizable.
Fashion and lifestyle brands can maintain the same model identity across multiple product videos. Upload the model's reference and generate them wearing different outfits in different settings without rebooking a shoot.
| HappyHorse (1-9 references) | Other Models | |
|---|---|---|
| Max reference images | 9 images per generation — each bound to a separate character token | Kling 3.0: 1 reference. Veo 3.1: up to 3 assets. Seedance 2.0: 1–2 references |
| Multi-character interaction | All characters rendered in one pass — natural interactions between referenced people | Single-reference models require generating characters separately and compositing |
| Identity binding method | Named tokens ([person1]–[person9]) in prompts — explicit control per character | Most models use a single implicit reference — no way to direct multiple identities |
| Group scene quality | Each person maintains their reference identity — no face blending between characters | Models with 1 reference often blend secondary characters' features with the primary |
| Use case fit | Best for multi-person narratives, team videos, family content | Better suited for single-subject content: portraits, solo product demos, monologues |
Identity extraction works best with well-lit, front-facing headshots or waist-up photos. Side profiles, sunglasses, or heavy shadows reduce matching accuracy. One person per reference image.
More references increase generation complexity. Start with 2–3 characters to validate your prompt structure, then add more. Beyond 5 characters in a single scene, positioning becomes harder to control precisely.
With multiple characters, vague spatial descriptions lead to crowded or ambiguous compositions. Specify positions: "[person1] on the left, [person2] in the center, [person3] on the right."
While HappyHorse supports up to 9 references, scenes with 7–9 characters leave less visual space per person. Identity accuracy remains high, but individual character detail decreases as the frame gets more crowded.
Join thousands of creators, agencies, and brands who use PonPon every day.
Start with free daily credits. No credit card required.