Freefall // Bloom — CapCut Seedance 2.5 Video Challenge

Project

Freefall // Bloom" — an entry for the CapCut Seedance 2.5 Video Challenge. The idea: a heroine falls through a shifting sequence of elements — starlit void, rain, snow, bloom — and instead of fighting the fall, she lives through each stage as a dance. Every element is its own emotional state, and the soft landing at the end isn't "the fall stopped" — it's "a new road began." Camera: a single continuous shot (one-take), a vertical tracking shot following the fall combined with a slow orbit around the heroine — no cuts

Task

A 30-second video, an honest one-take, holding character consistency throughout — per the contest rules, built through CapCut Video Studio with Seedance 2.5, on 6 reference images (identity lock, lighting, atmosphere, per-phase environment, palette, pendant detail).

References

A full set of 6 reference images was originally prepared for generation — each one carrying a different part of the look:

• Identity/outfit lock — full body, silhouette, gown, ribbon, pendant. The primary identity reference, used across every phase.
• Warm lighting — close-up, golden dusk light — reference for the Deceleration and Landing phases.
• Cool atmosphere / ribbon motion — reference for the opening Void phase.
• Rain environment — reference for the Rain phase.
• Bloom palette — reference for the Bloom phase.
• Pendant close-up detail — reference for the pendant's glow and the introspective Snow-phase moments.

What I did

Along the way I ran into a profile issue on CapCut: despite having received the contest credits, every generation attempt threw a "credit estimation service temporarily unavailable" error — I couldn't generate a single video through the contest platform itself. Rather than stall, I tested alternative paths

Tools

1. Recraft — start/end-frame interpolation only, no real frame-by-frame scene control. Produced a working but simplified draft (version 1, trial).
2. Comfy Cloud (Seedance 2.5 directly)** — a more flexible path, using the same prompt and references originally prepared for the contest.

That's where a technical limitation surfaced: the reference-to-video node (ByteDance Reference-to-Video) in the standard template only supports connecting 2 reference images, not all 6 as originally planned. The final video ended up generated with effectively 2 of 6 references — full text prompt and phase-by-phase timing intact, but without the complete visual reference set.

Prompt

Result

The finished clip: 30 seconds, 720p, one continuous take with no cuts, generated on Seedance 2.5 via Comfy Cloud. Despite the 2-of-6 reference limitation, all six emotional phases read clearly through palette and movement.