THE RAINDANCE FIELD GUIDE
The Rain Dance prompt and reference workflow
See how portrait roles, a fixed pier reference and a three-beat prompt guide the 15-second video.
Inputs and model
The current standard recipe calls MiniMax H3 through Replicate with one or two portrait references and a fixed pier reference video. It requests 15 seconds, 768P and a 9:16 ratio. The free preview uses the same model at four seconds with a shorter opening-only instruction.
Prompt structure
- Assign identities: image one is the singer; image two, when supplied, is the pier partner. Preserve each face, hair and clothing separately.
- Establish atmosphere: golden-hour light, wooden pier, sea and gentle performance.
- Guide timing: opening performance at 0–5 seconds, partner approach at 5–10 seconds, beach sunset ending at 10–15 seconds.
- Constrain errors: no face merging, extra foreground people, implausible hands, captions or watermarks. Do not request a recognizable song or an artist’s voice.
These are model instructions, not proof that every result satisfies them. Check real samples and review the entire output. The reference influences motion and framing; the model may deviate. This product does not use a multi-keyframe interpolation workflow and does not currently accept custom prompts or reference uploads.
Why no 28-second option yet?
The current model configuration supports at most 15 seconds. A longer multi-shot workflow needs its own identity-consistency review, timing checks and cost verification before it can be offered. Try the current workflow.