Prompt Architecture

Status: Candidate production standard

Prompts must be readable, diffable, and reusable. Store the structured record; compile it into model-specific text only at generation time.

Inheritance

Universe → Show → Season → Cast → Location → Shot → Model adapter

Later layers may add detail but must not silently contradict an earlier lock.

Human-readable shot record

identity:
  show: Little Lanterns
  season: 01 Alphabet Adventures
  episode: S01E__
  panel: 06

style:
  medium: cinematic anime-inspired preschool animation
  rendering: clean shapes, soft dimensional light, controlled texture
  palette: warm lantern gold against teal-blue shadows

cast:
  visible: [Mia]
  mia:
    hair: chestnut-brown chin-length bob with one small teal clip
    eyes: warm brown
    face: light freckles
    wardrobe: teal zip jacket, cream T-shirt, denim pants
    signature: tiny round lantern pendant

location:
  name: Lantern Club room
  anchors: [arched window, low map table, hanging lanterns]
  time: blue hour

shot:
  framing: close-up
  lens_intent: intimate portrait compression
  camera_height: eye level
  subject_scale: face fills most of frame
  composition: Mia alone; clean eyeline space camera-right
  action: she notices the letter glowing
  emotion: quiet wonder
  continuity: teal clip visible; background softly recognizable

output:
  aspect_ratio: 3:4
  dimensions: 864x1152

Compilation rules

Include only visible character features in the final shot prompt.
Put shot framing and subject scale near the beginning and reinforce them in composition.
State exactly who is visible; “Mia alone” is stronger than omitting Henry.
Use stable location anchors, but vary camera position and focal depth.
Preserve style language verbatim across a sequence.
Keep negative constraints in the model adapter, not scattered through story fields.
Record model, workflow version, seed, dimensions, and output link.

Why distinctive features matter

Perfect pixel identity is unrealistic across independent generations. Strong, orthogonal visual anchors—silhouette, hair shape, palette, signature object, and movement—make small variance tolerable while preserving recognition.

Composition rules learned from Anima tests

Require a single continuous camera image, one frame, one moment.
Require the environment to fill every pixel from corner to corner.
Reject solid-color borders, blank margins, vignettes, circular frames, isolated dioramas, poster layouts, split screens, comic panels, storyboards, and duplicated scenes.
Prefer one simple physical action. Multiple sequential actions increase the chance that Anima will create two panels or a before-and-after layout.
State subject scale and camera geometry, but do not state a height relationship between Mia and Henry.
Treat small accessories as secondary continuity details until reference-guided generation proves them stable.

Exact candidate character text

The following description blocks are copied verbatim from the positive prompts embedded in the selected v007 ComfyUI PNG metadata. Preserve the wording when testing prompt-only continuity; change only the surrounding shot, action, and environment text.

Mia

Mia is an ordinary young girl with a softly rounded face, straight chestnut-brown chin-length bob with a simple side part and one small teal hair clip, warm brown eyes, a few light freckles, teal zip-up jacket over a cream T-shirt, denim pants, cream sneakers, and a tiny round lantern pendant at the collarbone. Balanced natural child proportions and an everyday kid-next-door appearance.

Source: Mia-ReadingNook-v007_00001_.png

Henry

Henry is an ordinary young boy with a lively natural silhouette, compact tousled medium-brown hair with one small cowlick, warm brown eyes, mustard zip jacket over a navy T-shirt, denim pants, simple sneakers, a short burnt-orange neckerchief, and a small blue cross-body satchel. Balanced natural child proportions and an everyday kid-next-door appearance.

Sources: Henry-Soccer-v007_00001_.png and Henry-Pond-v007_00001_.png

Current style target

The current target is the rendering family represented by Mia-ReadingNook-v007_00001_.png:

Warm cinematic storybook animation matching a polished contemporary children's picture book: soft amber backlight and gentle window glow, rich but muted teal, cream, rust, navy, and honey-brown palette, rounded natural child faces, medium-large expressive brown eyes, fine dark-brown linework, softly shaded 2.5D forms, subtle fabric, hair, paper, and wood texture, detailed lived-in environment, gentle atmospheric depth, balanced full-bleed composition.

Keeping that text unchanged produced a coherent style family but not an exact lock. Exact matching will require a validated style-reference workflow, reference conditioning, or a tested style LoRA. See Character, Style, and I2V Tests — 2026-07-27.

Prebuilt LoRA consistency candidate

The strongest prebuilt style-consistency result so far used:

model: anima-base-v1.0.safetensors
lora: Anima/noobai-style-anima-lora-final.safetensors
lora_strength: 0.75
trigger: noobai-style
dimensions: 1024x1024
steps: 30
cfg: 4
sampler: er_sde
scheduler: simple

NoobAI kept a coherent clean anime/storybook look across ten varied frames. It did not lock character identity: Mia's bangs and side part drifted, while Henry was more stable. Treat it as a candidate rendering adapter, not a replacement for character references or an approved final style.

Wan mouth-motion finding

Wan 2.2 14B Turbo produced appealing automatic movement from the NoobAI first frames with empty positive and negative prompts, but repeatedly animated the children's lips as if they were speaking.

This exact short positive prompt was tested on six matched sources at CFG 1:

Silent natural motion. Preserve the exact mouth shape; no talking or lip-sync.

Lip motion persisted. Closed-mouth source frames, alternative concise wording, slightly higher CFG, and the full non-Turbo workflow remain controlled tests. Do not claim that speech-like mouth motion has been suppressed until one of those tests succeeds.

Prompt-Architecture