HappyHorse 1.1 Reference-to-Video: How Many Photos You Really Need and What It Costs (October 2026)
HappyHorse 1.1 builds a new 3 to 15 second shot from up to nine photos of one person, at $0.14 per second at 720p. How the character1 to character9 prompts work, what the 1080p default does to the price, and why more photos help.
HappyHorse 1.1 reference-to-video is the model behind the idea of "put me in a scene I describe": you provide up to nine photos of a person and a prompt, and it generates a new shot with that person in it. It is Alibaba's model, it lists on fal at $0.14 per second at 720p, and it is the one we recommend first for the job. The questions that come up most are how many photos it really needs, what the defaults do to the bill, and what it cannot do. This guide answers them from the schema and from our experience; for a general explanation of the family, see what HappyHorse is.
What it costs, and the default that doubles it
The $0.14 figure is the 720p rate. The endpoint's default resolution is 1080p, which bills about twice that, so leaving the setting alone turns a $0.70 five-second clip into roughly $1.40.
| Length | 720p ($0.14) | 1080p (about $0.28) |
|---|---|---|
| 5 seconds (the default) | $0.70 | about $1.40 |
| 10 seconds | $1.40 | about $2.80 |
| 15 seconds (the maximum) | $2.10 | about $4.20 |
Read from fal on 18 October 2026. Set both resolution and aspect_ratio on every call: the aspect ratio defaults to 16:9, which gives a vertical creator a landscape clip they cannot post. The endpoint supports nine ratios, including 9:16 and 4:5.
How the references work
You do not simply upload photos. You cite them in the prompt as character1 through character9, in the order you supplied them in image_urls. A prompt is therefore a sentence about named characters: "character1 stands at a kitchen counter, laughs and turns to the window". If you supply several photos of one person and cite only character1, the model uses the others as context rather than as separate people.
That mechanism explains the advice that surprises people: one frontal selfie produces a mediocre result, and several angles produce a good one. The model is matching a named subject across views. With a single front-on photo it must invent the profile, and the invention is what viewers feel as "not quite them".
How many photos
- One: works for a locked-off, front-facing shot and not much else. Expect drift when the head turns.
- Three or four: a frontal, both three-quarters and a profile, in similar lighting. This is where most of the gain is.
- Up to nine: useful when outfit and body shape matter, or for two people. Add full-body frames in the clothes you want.
In our own comparison on 24 September, going from one photo to four raised our face-similarity score, but not by enough to pass our bar on a difficult source clip, which is why we are repeating it on an easier one. The lesson is the order of operations: references help, and a clear source and a calm prompt help more. Four good references beat nine mixed ones, so do not pad.
Reference image rules
JPEG, PNG or WEBP; shortest side at least 400 pixels (720p or better is recommended); up to 20 MB each. Whole head in frame, no sunglasses, no heavy shadow across one side. Treat each photo as evidence for the model: the more consistent the lighting between them, the less it has to guess.
What it does and does not do
It builds a new shot. The framing, the setting and the action come from your prompt. It does not copy a source clip's choreography and does not keep a source clip's background or cuts, so it is the wrong tool for "remake this Reel exactly" (that is a video-edit job: HappyHorse video-edit, Kling O3 edit or Recast).
It generates native audio with multilingual lip-sync, so a short speaking line is possible in the same call. It accepts real faces, which is not true of Seedance. Identity still drifts with duration and motion, so five seconds of calm action beats twelve of vigorous.
Where it compares
At the same $0.14, Kling O3 reference-to-video takes only four references but treats an object as an element too; Wan 3.0 Prime takes ten photos and video references but is billed in extra ways. HappyHorse's edge remains the nine references and the vertical formats, and our choice of it as the default for scene work is documented in the O3 comparison.
Common questions
Why does my clip come out landscape?
The aspect ratio defaults to 16:9. Set it to 9:16 or 4:5 on every call.
Why was the bill double what I expected?
The default resolution is 1080p, which bills about twice the 720p rate. Set resolution deliberately.
Do I need nine photos?
No. Three or four good angles in similar lighting capture most of the gain. Four good references beat nine mixed ones.
Can it copy a dance?
No. It builds a new shot and does not read a source video's motion. Use Motion Control for choreography or a video-edit model to keep a source clip.
How to do this without code: HappyHorse reference-to-video is the model behind the Me in scene tool, with the resolution and aspect ratio set for you. In AutoHustle this is the Me in scene tool: a few photos of you and one sentence describing the scene (about 2 credits per second, quoted before you press Generate). See How to put yourself in any scene and Put yourself in a movie scene.
Related reading: How Many Photos Do You Need for an AI Video of Yourself? · Seedance 2.5 on fal: 30-Second Takes, 50 References, and Why It Still Cannot Use Your Face (October 2026)
For developers: the fal endpoint
| Model | fal endpoint |
|---|---|
| HappyHorse 1.1 reference-to-video | alibaba/happy-horse/v1.1/reference-to-video |
Inputs: image_urls (1–9), prompt (citing character1…character9), duration (3–15, default 5), resolution (720p/1080p, default 1080p), aspect_ratio (nine values, default 16:9), seed. Price read on 18 October 2026: $0.14 per second at 720p; 1080p about double.