HappyHorse vs Kling O3 for Reference-to-Video: Same Price, Different Strengths
Both cost $0.14 a second and both hold a character across a generated shot from your photos. Where the aspect ratios, the reference handling and the audio differ.
Reference-to-video is the family that puts you into a shot you describe, built from photos rather than from a source clip. Two models do it well at the same price, so the choice is purely about fit.
Prices checked against the live endpoints on 3 August 2026 — both are $0.14 per second, so a ten-second clip is $1.40 either way.
HappyHorse 1.1 Reference to Video
Up to nine reference images, addressed in the prompt by position — character1 through character9, matching the order you supply them. Duration 3–15 seconds, default 5. Resolution 720p or 1080p, billed the same. Reference images need a shortest side of at least 400px and cap at 20 MB.
The standout is aspect-ratio coverage: nine ratios including 9:16, 4:5, 1:1, 21:9 and 9:21. For vertical social work that full set is unusual, and it is the main reason this is our default for Reels-shaped output.
It also generates native audio with multilingual lip-sync in the same call.
The trap: the default aspect ratio is 16:9. A vertical creator who runs at defaults pays for a landscape clip they cannot post. Set it every time.
Kling O3 Pro Reference to Video
Kling's reference endpoint, same $0.14 per second, up to four references. Its strength is cinematic language — the O3 line preserves motion and camera style well, and it sits inside the Kling family, so a workflow that also uses Motion Control or Kling 3.0 Pro Image to Video stays in one place operationally.
How to choose
| If you need | Pick |
|---|---|
| True 9:16 or 4:5 output | HappyHorse |
| Voice and lip-sync in the same call | HappyHorse |
| More than four reference photos | HappyHorse |
| Cinematic camera feel | Kling O3 |
| Consistency with a Kling-based pipeline | Kling O3 |
At the same price, capability is the only sensible tiebreaker.
The reference photos matter more than the model
This is the part people skip. Reference-driven models are matching a named subject across angles, so a single frontal selfie leaves the model inventing your profile — and the invention is what viewers read as "that is not quite them".
Supply frontal, both three-quarters and a profile, in similar lighting, whole head in frame, no sunglasses. If wardrobe or build matters to the shot, include full-body frames in the clothes you want. Four good references beat nine bad ones.
The more expensive alternative
MiniMax H3 Reference to Video at $0.26 per second — nearly double — adds 2K output and lets you cite reference clips for motion and audio alongside images. Released 31 July 2026, so capability is documented and a track record is not. Worth trying when the two $0.14 options both fail on a specific shot; not worth defaulting to.
What none of them do
Copy the choreography of an existing clip. Reference models build a new shot; they do not ingest motion. For that it is Kling 3.0 Pro Motion Control at $0.168 per second, with character_orientation set to video — on image the face-binding element silently does nothing.