Animated Short · AI Pipeline · Swiss Heritage
An internal research project at Vertical Labs — a ~2:13 animated short built to validate a new AI-native animation pipeline and serve as strategic client teaser for the Swiss chocolate industry. Twenty shots. Three acts. One tiny cocoa bean hero named Bérnois.
Discipline
Animated Short Film
Runtime
2 min 13s / 20 shots
Pipeline
Flora.ai × Seedance 2.0
Status
Completed
The Great Cocoa Avalanche is an internal research project at Vertical Labs — a ~2:13 animated short built to validate a new AI-native animation pipeline and serve as strategic client teaser for the Swiss chocolate industry.
High in the Alps, an alphorn's vibration dislodges a cocoa fave from a cliff. It cracks open, revealing Bérnois — a tiny cocoa bean with big dreams. He bounces down the valley meeting a cow, a chocolate-drinking St. Bernard, and a marmot marching band, until a yodeler's impossibly high note triggers a cocoa avalanche that buries the local chocolaterie. In the aftermath, a master chocolatier discovers Bérnois, takes him under his wing, and together they create a masterpiece truffle.
Every Swiss cliché is present and correct — alphorns, cowbells, yodelers, marmots in lederhosen — played with deep affection, never irony. The tone is Pixar-wholesome, maximum goofiness, zero cynicism.
Every shot is ≤10 seconds — built around Seedance 2.0's generation ceiling. Each is produced from locked start-frame and end-frame keyframes, connected by a natural-language motion prompt, assembled in Flora's node canvas with branching variants.
One of the core technical challenges of AI filmmaking is keeping characters consistent across 20 independent generation calls. Each shot is generated separately — maintaining visual identity across that many nodes requires deliberate character engineering upfront.
The cocoa bean protagonist. Defined as a top-level reference node — every shot inherits from this anchor.
The unwitting catalyst. His impossibly high yodel note triggers the entire avalanche. One scene, enormous narrative consequence.
The first character Bérnois encounters. A silent, massive, utterly unimpressed Swiss cow. The comedy is in the size contrast.
A St. Bernard with a chocolate-filled barrel collar. Stoic, enormous, oddly wise. Nods once at Bérnois. That's enough.
A trio of marmots in lederhosen playing pretzel-stick alphorns. Pure comic relief. They are having the best day of their lives at all times.
The inadvertent villain. His glass-shattering high note triggers the cocoa avalanche that sets the entire third act in motion. Zero awareness, maximum impact.
Flora's node system allows character references to be locked at the top of the graph. Every downstream shot node inherits them automatically — no copy-pasting prompts.
Despite reference locking, some shots produced character drift — slightly different proportions, altered eye shape, inconsistent colour temperature.
Character consistency in AI filmmaking is not a generation problem — it is an architecture problem. The quality of your reference nodes determines the consistency of your cast. Invest in the character sheet before you write a single shot prompt. The shots are easy. The character is the product.
Vertical Labs has standardized its AI film pipeline on Flora.ai for creative orchestration, Seedance 2.0 for video generation, and DaVinci Resolve for final assembly and edit. The Great Cocoa Avalanche is the first internal project to validate this three-tool workflow across a 20-shot, three-act narrative.
Node-based infinite canvas with FAUNA agent. Character references locked as top-level nodes. Every shot is a branching node with 3–4 variants compared side-by-side. Full lineage from prompt to final cut.
ByteDance's multimodal video model. Native audio co-generation, physics-aware motion, character consistency via @reference tagging, up to 15 seconds per shot at native 2K, 60fps.
Industry-standard editing and color grading. All 20 shots assembled, trimmed, and sequenced on the timeline. Final color grade, audio mix, and master export to 2K delivery.
Eight characters and environments generated once in Imagen 4 / Flux 2 Pro, bound to @tags at the top of the Flora canvas.
For each of the 20 shots, a start-frame and end-frame image are generated as locked keyframe nodes.
Keyframes feed Seedance 2.0 with the shot prompt. Native audio — dialogue, SFX, ambient — is generated in the same pass.
3–4 variants per shot. Best takes promoted to the assembly track. Across 20 shots, roughly 60–80 total video generations.
Twenty shots exported from Flora into DaVinci Resolve for final edit, color grade, audio mix, and delivery as a single ~2:13 2K master file.
Every element of this short was engineered for swappable reuse. Replace Bérnois with a brand mascot, re-run the node graph, and deliver a new film in the same visual language within days.
A single reference pass locks a mascot across a 20-shot film. Client mascots can be swapped in without re-architecting the pipeline. Every future short reuses the same node skeleton.
Alphorns, cowbells, marmots, St. Bernards — and a chocolatier-mentor apprenticeship arc. Every Swiss cliché is a love letter, never satire. Chocolatiers will feel celebrated.
Two minutes thirteen seconds of broadcast-quality animation in days rather than the months a traditional pipeline would demand. Campaign-velocity filmmaking for brands that used to commission one film per year.
The deliverable isn't just the film — it's the reusable Flora node graph. Future client work inherits the full 20-shot, three-act pipeline as a starting template, with character slots pre-wired.