龙舟小波 · Bo's Dragon Boat Race
端午赛龙舟, told as a 6-beat cartoon — and starring one chibi dragon who looks exactly the same in every shot. The whole reel is locked to a single character picture. The part most AI video gets wrong, done right with primitives that exist today.
The story a complete arc in 6 beats
Dawn mist on the river → the crew boards and Bo lights the drum → the race explodes off the line → a rival pulls ahead → Bo breathes a rainbow gust and the boat lunges to the line → fireworks, zongzi, victory. It reads in one watch, it's funny, and it's unmistakably the same dragon throughout.
Caption-ready: 祝大家端午安康 🐲🚣 #端午 #龙舟 #DragonBoatFestival
The one image that locks it 小波 — Bo's character sheet
Everything starts here. A single character sheet of 小波 (Bo) — the carved dragon-head prow come to life: jade-green scales, a gold water-drop belly spot, two little curved horns, a red neck ribbon, big mascot eyes. This image is the only identity input. Every one of the 6 beats is generated by carrying Bo's identity forward from this sheet (subject-preservation), not by re-drawing him each time — which is exactly why his horns, belly-spot and ribbon stay put.
This is the same idea as VibePaper's "draw a character sheet, then connect it into every shot" — done here with place_subject, which fans one subject across N scenes with the preservation prompt applied automatically. gpt-image drew the sheet; nano-banana placed Bo into each beat.
Want the lock declared once on a canvas and reused across a whole series? That's the Wire Graph version of this same reel — see the comparison below.
The 6 beats Bo, consistent, scene to scene
How it was made primitives that exist today
No LoRA, no fine-tuning — pure reference conditioning. The end-card text ("端午安康") is a clean overlay rendered with a CJK font, never baked by the image model (AI text rendering is the one thing to never trust). Total spend ≈ $2.70: one sheet + seven keyframes (one re-render) + eight i2v clips (two retries) + a music bed.
Honest verdict what held, what needed a nudge
✓ What held
- Bo held 6/6. Same horns, belly-spot, ribbon and proportions across every beat — the one-image lock worked.
- The arc reads in a single watch; it's legible and fun.
- The festival end-card text is crisp (overlay, not generated).
⚠ What needed a nudge
- 2 of 6 i2v clips failed on the first pass ("fetch failed" — fal-side latency, not the orchestrator). One retry each fixed them.
- Beat 4's first keyframe had a stray artifact + clutter; re-rendered once (~$0.04) before animating.
- Fast-water motion can thin fine details — premium
seedance-i2vheld them better than the fast tier would.
The same reel, two ways manual lock vs Wire Graph
This reel was made the manual reference-passing way — Bo's sheet hand-fed into each shot. The forthcoming Wire Graph produces the identical reel by wiring Bo's node into all 6 beats once (VibePaper-style), which is what pays off for multi-character series.
| Manual (this reel) | Wire Graph | |
|---|---|---|
| Where the lock lives | per step (re-passed each beat) | once, on the graph |
| Multi-character | fiddly (track who's in which shot) | native (draw another line) |
| Re-run one beat | re-pass the sheet | surgical — re-pull that node's wires only |
| Surfaces | playbook the agent runs | one schema on CLI · MCP · webapp |
| Available | today | after Wire Graph Phase 1 |