The Bremen Town Musicians

The Project

What it's about

The same fairy tale, two depths — and consistently made child-friendly.

The goal is an animated video of "The Bremen Town Musicians" (Brothers Grimm) in two versions: a very simple one for ages 1–3 (~5–6 min, short sentences, lots of repetition, onomatopoeia) and an elaborated one for ages 4–6 (~10 min, real dialogue, character humour, slapstick).

In the original, the animals are to be killed — that is consistently defused: no threat, the animals are simply "old and no longer needed" and find a new purpose. The robbers aren't evil but clumsy and cowardly (slapstick instead of scares).

Recurring musical motif: each animal is one note — Hee-Haw · Woof · Meow · Cock-a-doodle-doo — together "the band". Usable as a leitmotif across the whole series.

Current status: proof of concept passed. The entire pipeline is proven and a first complete, voiced, text-faithful scene exists. Next up: produce the screenplay scene by scene.

The full cast

Size reference: three human-sized robbers, the upright donkey as the largest animal, then the dog, the rooster, the cat.

The Characters

The cast

Every character is created once as a still in a uniform picture-book watercolour style — the donkey is the style anchor, his image sets the look for all the others. All four animals are gently aged (greying muzzle, soft wrinkles), fitting the "too old" theme.

Donkey — Grauschimmel

Leader, optimist, has the Bremen idea. Carries a lute.

Voice: Crizz · deep, warm

Dog — Packan

loyal, clumsy, always hungry. Kettledrum as his anchor.

Voice: Adrian · strong, cheerful

Cat — Bartputzer

elegant, grumpy, sharp tongue.

Voice: still to be cast

Rooster — Rotkopf

loud, proud, the "early warner" — spots the light.

Voice: Paul · bright, dramatic

The robber gang

Captain + 2 henchmen. Clumsy and cowardly, never scary.

Narrator: Bettina · warm

Robbers' house (night)

Setting background: a warmly lit window in the dark forest.

Robbers' parlour (interior)

Stage for the window peek: a table full of food, candlelight.

Size chart

Defines the relative standing heights of the four musicians.

Group reference

The four in natural poses — defines how the characters relate.

Per character: a pose library

Identity references + action references

The key to character stability (more below): every character gets several views. Identity = front · face close-up · profile · rear. Action = a matching pose per motion (e.g. walking, with the lute on its strap instead of in the hands). Just like an animation studio keeps a model sheet per action.

Identity Front (anchor)

Identity Face close-up

Identity Profile

Identity Rear

Action Walking, profile — lute on the strap

Action Walking, ¾ view

Action Dog — four-legged, drum as anchor

Action Cat — four-legged, tail up

Action Rooster — strutting on two legs

Sheet 4 poses in 1 image (2×2)

Sheet All 4 animals in 1 image

The Pipeline

Four building blocks, three small scripts

None of this needs a web UI or a human clicking somewhere — everything runs headless as a Node script that calls the APIs and assembles with ffmpeg.

{
  "id": "fenster", "duration": 6, "variants": 2, "pick": 2,
  "refs": [
    "assets/characters/esel.jpg",
    "assets/characters/refs/esel_profile.jpg",
    "assets/characters/raeuber.jpg",
    "assets/characters/stube.jpg",
    "assets/characters/haus.jpg"
  ],
  "prompt": "Night exterior, locked static camera... the old grey donkey
    peers inside the lit window... THROUGH THE WINDOW three funny scruffy
    robbers sit around a big wooden table laden with food, feasting...",
  "narration": [
    { "speaker": "hund", "text": "Was siehst du denn, Grauschimmel?" },
    { "speaker": "esel", "text": "Einen Tisch voll mit herrlichem Essen —
       und drei Räuber, die sich's so richtig schmecken lassen!" }
  ]
}

Models compared

Which video model for which shot?

Every model runs through the same Atlas Cloud API. The key insight: no single model wins — each shot type has its model.

The same donkey, five approaches

Click to play — the clips start automatically as soon as they scroll into view.

Challenges & solutions

What was hard — and how we solved it

The honest part. Almost every step forward came out of a visible failure. Each card shows Problem → Cause → Solution, many with the before/after video evidence.

4 Extra arms while walking

ProblemWhen walking, the donkey grows extra arms that don't belong.

CauseAll identity refs show him upright with "arms" (playing the lute). The model mixes that arm pose with walking legs → anatomy errors.

SolutionProvide action references per motion: a walking profile with the lute on its strap (not in the hands) + a sharpened negative against "extra arm, holding the lute". Every character needs a pose library per action.

The finished scene

"Approaching the robbers' house" — 37 seconds, end to end

The real milestone: not "does the pipeline work?" but "we can produce scenes on an assembly line" — style-true, character-stable, with German character voices, faithful to the Grimm source.

Strategic approaches

Which tools — and how much AI?

A fundamental question ran in parallel: shouldn't this really be built in a proper 3D engine? The honest trade-off — and the chosen direction.