I2V Storyboard Multi-Cut: a Multi-Shot Sequence from a Single I2V Input
Happy Horse 1.0 — and most other image-to-video models — accept exactly one input image, used as the first frame of the clip. You want a four-shot music-video sequence: close-up, wide, intercut detail, top-down. Two options exist. Generate four separate clips and stitch them — style drifts between cuts and identity won't lock. Or accept a single continuous shot — and lose the rhythm the brief asked for. Neither hits what the brief wanted.
The trick is to fold the four shots into the one allowed input — as a 2×2 storyboard panel — and then prompt the model to briefly show it and snap-cut into each panel in turn.
This is the I2V storyboard multi-cut: a prompt-engineering trick for image-to-video models that accept only one input image. Feed a multi-panel storyboard as that input, instruct the model to show it for ~1 second as an opening reference grid, then snap-cut into each panel one by one. You get a multi-cut montage from a single I2V call.
Why naive attempts fail
The two naive approaches both lose something:
- A 4-shot sequence from a single input image → can only get one continuous shot.
- Stylistically locked across cuts → can't pass four reference frames separately.
- Intercut close-ups, wides, top-downs → impossible without re-running the model four times and stitching, which loses style and identity coherence.
The trick
- Compose a single input image as a 2×2 (or N×M) storyboard panel — each panel is one of the shots you want, drawn in the final visual style.
- In the prompt, tell the model the storyboard fills the frame for ~1 s as the opening reference.
- Then issue explicit snap cuts into each panel in order, with each cut's content re-described in cinematic terms (subject, framing, action) — not as "panel 1", but as the actual shot.
- Lock visual-style language across all cuts so the four shots read as one piece.
Why it works:
- The model honors the first frame as-given, so the storyboard appears at t=0 — that's free.
- Without an explicit "leave the first frame at 1 s" cue, the model lingers on the input for the whole clip. The cue gives permission to move on.
- Re-describing each shot in cinematic terms forces the model to re-render in motion instead of statically panning across the panel grid.
- Shared style descriptors across cuts make the model treat the four shots as one stylistic universe, even though each is a separate beat.
Pattern (drop-in)
8s duration. First 1s: <hand-drawn / cel-shaded / photoreal> N-panel storyboard
fills the frame as an opening reference grid, <style descriptors>. At 1s, snap
cut to Scene 1 (Xs): <scene 1 rewritten as a real shot — subject, action,
framing>. Snap cut to Scene 2 (Xs): <scene 2 ...>. Snap cut to Scene 3 (Xs):
<...>. Snap cut to Scene 4 (Xs): <...>. Locked-off framing on every cut, no
in-shot camera movement — only the cuts move. <Lighting>. <Visual-style line
shared across all cuts>. <Mood>. <Audio>. <Aspect>, <pacing>.
Constraints to respect
- Total ≤ 8 s. With a 1 s storyboard hold + N cuts, each cut gets
(8 − 1)/Nseconds. For 4 cuts that's ~1.75 s — short but viable for music-video pacing. - Locked-off framing per cut is the cleanest way to stay inside Happy Horse's "one camera move per clip" rule — the cuts themselves are the only motion. Adding in-shot dollies or orbits inside each cut breaks the rule and causes drift.
- Style language must be shared across cuts. If style descriptors only appear at the end of the prompt, the early cuts may render in a different look. Repeat the style anchor inside each cut description, or wrap them with one final cross-cutting style line.
- Order must match panel reading order in the input (TL → TR → BL → BR for 2×2). Out-of-order cuts confuse the model and produce hybrids.
Failure modes (sample N ≥ 3, pick best)
- Storyboard sticks. The model animates panel 1 for the whole 8 s, ignoring the cut cue. Mitigation: emphasize "At 1s, snap cut" and keep the storyboard hold at 1 s.
- Static pan across the grid. The model treats the panel layout as the actual scene and pans across it. Mitigation: rewrite each scene cinematically so the model re-renders, not pans.
- Style drift between cuts. One cut goes photoreal while others stay illustrated. Mitigation: repeat the style anchor inside each cut, plus a final cross-cutting style line.
- Cuts blur into morphs. The model interpolates instead of cutting. Mitigation: use the literal phrase "snap cut" — softer phrasings ("transition to", "then we see") read as morphs to HH.
Worked example — abstract ink-bloom study
Input: a single 2×2 photoreal-macro storyboard of an ink-in-water study —
- TL: extreme macro of a single dark ink drop kissing a still water surface
- TR: side profile of a dark ink plume blooming downward through clear water
- BL: top-down of ink spreading in slow concentric ripples
- BR: wide static of the glass tank against a soft backlit gradient
Prompt:
8s duration. First 1s: 4-panel photoreal macro storyboard fills the frame as
an opening reference grid, high-key minimalist style, soft diffused light.
At 1s, snap cut to Scene 1 (1.75s): extreme macro of a single dark ink drop
the moment it kisses a still water surface, surface tension dimpling, shallow
DOF. Snap cut to Scene 2 (1.75s): side profile of a dark ink plume blooming
downward through clear water, slow billowing tendrils. Snap cut to Scene 3
(1.75s): overhead top-down of ink spreading in slow concentric ripples on a
white backdrop. Snap cut to Scene 4 (1.75s): wide static of a glass tank
filled with the swirling ink against a soft backlit gradient. Locked-off
framing on every cut, no in-shot camera movement — only the cuts move.
Soft diffused side light, high-key minimalist palette of off-white, ink black
and pale blue, photoreal macro look with shallow DOF, consistent across all
cuts. Meditative, hypnotic, contemplative. Subtle water ambience, faint
low-volume synth pad underneath, no dialogue. 16:9, contemplative pacing.
The pattern is style-agnostic — swap the storyboard and rewrite the four scene descriptions; everything else stays.
When NOT to use this trick
- The shots need genuinely different lighting / time-of-day. The model holds one lighting state across the clip; multi-time-of-day cuts will fail.
- The shots involve different characters who must look identical. Identity drift across cuts is high; if exact character lock matters, generate per-shot and stitch externally.
- You need more than 4 cuts. 4 × 1.75 s is already tight; 5+ cuts collapse into < 1.4 s each and the model degrades to flashes.
- The output is for a context that prohibits cuts (e.g. continuous-shot brand films).
Pairs with
- happy-horse-prompt-rules — base ruleset; this trick layers on top.
- happy-horse-prompt-engineer — system prompt that operationalizes the ruleset; this pattern is compatible with its 8-slot output.
- seedance-2-cinematic-video-prompt-engineer — Seedance 2's
s[]array makes the trick unnecessary there (multi-scene is native), but the storyboard-as-style-lock idea still applies. - chaos-frame-feed-architecture — adjacent technique: a panel grid as a deliberate composition, not just an input hack.
- typed-reference-composition — related discipline: multi-input prompts with explicit per-reference role declaration.