SuperAPI AI Directory

webtoon-panel-breakdown

General & other

View repository

Updated 10-16 04:56

About this skill

webtoon-panel-breakdown — from script to panels, from style to consistency

A webtoon's pull comes not from "what you draw" but from "how you cut and how you show it." This skill handles two visual production tasks.

  1. Panel breakdown (shot list) — Split one episode of a script into 50 or more panels and define each panel's composition/camera/emotion/direction.
  2. Style consistency — Lock the art style into a style bible and create "consistency tokens" that keep character appearances from wavering across panels and episodes.

Both tasks have one purpose: to keep the reader from stopping their scroll, and to make the same character look like the same person in every panel.


Why do it this way (Why-first)

  • A panel is time. In a vertical scroll, panel size and the space between panels are exactly the time the reader lingers in that moment. Large panels and wide gaps = slow time (the peak of tension, impact). Small panels and narrow gaps = fast time (the sum of dialogue, a chain of action). Breaking a beat into panels means designing time.
  • 50+ panels are not a luxury but rhythm. If you handle one beat as one shot, the information is conveyed but the emotion is not. You must split the tension-building section finely for the reader's heartbeat to rise along with it. (Contract §9: guarantee 50+ panels per episode)
  • Without consistency tokens, codex-image draws a different person every time. Generated images respond strongly to the appearance keywords in the prompt. You must inject a fixed set of appearance keywords for each character identically into every prompt for them to look like the same person. Creating these tokens is the art-director's core deliverable.
  • The style bible is the single source of truth (SSOT) for collaboration. prompt-smith, panel-artist, and letterer all reference this document. If the aspect ratio, color, and tone decided here waver, the 50 pages become each different.

Task A: Panel breakdown (creating a shot list)

Input: _workspace/03_episode/ep{NN}_script_final.md, _workspace/04_visual/style-bible.md Output: _workspace/04_visual/ep{NN}_shotlist.md

A-1. Breakdown principle (one beat → multiple panels)

  1. Identify the beats first. Read the script and draw a beat boundary at every "point where the reader's emotion changes." One episode usually has 12–20 beats.
  2. Expand each beat into panels. At least 2–4 panels per beat. If the average is 15 beats, you naturally reach around 50 panels. If short, split further with the rules below.
  3. Split tension-building sections more finely. For climax, twist, and confrontation scenes, up to 5–8 panels per beat. Breaking it down in the order of facial close-up → hand/prop insert → reaction shot → stillness (negative space) increases tension.
  4. Dialogue exchanges (back-and-forth) split panels. Do not cram 3 speech bubbles into one panel; as a rule, cut the panel when the speaker changes (Contract §9: dialogue-focused). However, for fast tiki-taka, up to 2 speech bubbles per panel is allowed.
  5. The twist panel is isolated. The cut where the twist hits should be a standalone large panel with space (an empty panel or large negative space) before and after it. A "time freeze" for impact.
  6. Foreshadowing is planted in advance with insert panels. Insert props/expressions to be paid off later as small insert cuts into earlier beats.

A-2. Check guaranteeing 50 panels

  • After the first breakdown, count the panels. If fewer than 50 , increase them by applying the following (no forced padding; only where rhythmically valid):
    • Add reaction shots at emotional turning points
    • Add a "stillness beat" (character freeze/background insert) before important dialogue
    • Split action into motion stages (preparation → execution → result)
    • Split the moment just before the cliffhanger into a chain of close-ups
  • If 50 or more, pass. If it exceeds 70, check whether the rhythm drags (merge unnecessary duplicate cuts).

A-3. Shot list item format

For each panel, record the following. (Since prompt-smith translates this into a codex prompt, the visual information must be specific.)

### panel_007
- scene_id: S3 / location: LOC_ROOFTOP   # 같은 씬은 같은 장소 — 배경 연속성
- beat: 3 (대치 시작)
- size: large            # large/medium/small/insert/full-bleed
- camera: low-angle, medium shot
- composition: 화면 좌측 1/3 라인의 주인공, 우측 깊은 공간에 적대자(원근 대비)
- subject: 지호(긴장으로 굳은 표정), 배경에 흐릿한 군중
- emotion: 억눌린 분노
- motion/direction: 시선 좌→우 유도 (다음 패널로 흐름)
- dialogue_ref: "여기서 멈춰." (지호)   # 레터러가 in-image 말풍선 스펙으로 변환
- fx/note: 속도선 없음, 정적. 반전 복선(반지 인서트는 panel_006)
  • size와 camera, compositionis chosen from the vocabulary in §references/composition-grammar.md.
  • scene_id/location(LOC_*) is recorded for every panel. Panels with the same scene_id must be the same place, time period, and indoor/outdoor (to prevent sudden background changes). Location changes only at SCENE BREAK. LOC_* tokens must match the style-bible's location tokens (B-5).
  • dialogue_refis the basis for the letterer to create the in-image speech bubble spec and for prompt-smith to create the bake prompt, so it must be recorded for panels with dialogue.
  • Panel numbers are panel_001consecutive 3 digits starting from (Contract §4).
  • At the top of the episode, place a location list table(scene_id ↔ LOC_* ↔ one-line background).

A-4. Scene group hints (optional)

The shot list is a stage before render distribution, but if scene transition (time/place/sequence) boundaries are marked, prompt-smith can more easily distribute A/B/C groups evenly. At every major scene transition, add a --- SCENE BREAK --- line.

For the vertical scroll staging grammar (cut size, negative space, gaze guidance, rhythm) and the camera angle/composition vocabulary catalog, see references/composition-grammar.md . When filling in panel size/camera/composition, choose from that vocabulary.


Task B: Style bible + consistency tokens

Input: _workspace/02_story/characters.md, _workspace/02_story/world.md, _workspace/01_research/trend-brief.md Output: _workspace/04_visual/style-bible.md, _workspace/04_visual/character-sheets.md

B-1. Style bible structure (style-bible.md)

Translate the trend brief's target readers and genre tone into the language of artwork. Fill in the following sections.

  1. Art style — Line tone (clean/rough), coloring method (cel/soft/semi-realistic), level of render detail, reference mood (e.g. "Korean romance fantasy webtoon, soft cel shading").
  2. Color Feel/Palette — Main palette (3–5 HEX colors), mood palette per scene (day/night/flashback/tension), saturation and brightness direction.
  3. Atmosphere/Tone — One paragraph on genre mood (e.g. cold, tense urban noir). Lighting tendency (backlight/rim light/high-key, etc.).
  4. Aspect Ratio/Canvas — Based on webtoon vertical scroll. Default panel aspect ratios (e.g. square 1:1, wide 3:2, vertical cut 2:3, full bleed 9:16). Includes size guidance to put into codex-image.
  5. Consistency Token Convention — One line of "global style tokens" to be commonly injected into all prompts (B-3 below). Character tokens are separated into character-sheets.
  6. Location Lock Tokens (LOC_*) — For background continuity. B-5 below.
  7. Speech Bubble Visual Convention — Style per speech bubble type for in-image baking + Korean lettering tone. B-6 below.
  8. Prohibited/Caution — No watermarks, no deviation from art style, no abrupt changes of background/location unrelated to the scene, negative prompts are no watermark, no English text, no gibberish/garbled/misspelled text(since speech bubbles and Korean text are drawn, no textis not used).

B-2. Character Sheet Structure (character-sheets.md)

The prose-style appearance descriptions in characters.md are refined into reusable token sets. Per character:

## 지호 (주인공)
- identity_tag: JIHO
- 고정 외형 토큰(일관성 토큰):
  "20s Korean man, short tousled black hair, sharp single-lid eyes,
   slim build, faint scar on left brow, charcoal hooded jacket"
- 표정 변주(상황별): 분노/불안/무표정 (외형 토큰은 고정, 표정만 교체)
- 의상 변주: 기본(후드)/정장/부상 후 — 변주마다 토큰 추가구
- 금지: 머리색·눈매·흉터 변경 금지(일관성 핵심)
- 참조 컷 파일(있으면): _workspace/05_panels/refs/jiho_ref.png

B-3. How to Make Consistency Tokens (Core Technique)

  1. Separate invariant vs variable. Things that absolutely never change in appearance (hair color/length, eye shape, body type, identifying marks) are grouped into invariant tokens. Expression, pose, outfit, and lighting are handled separately as variable.
  2. Use high-density noun phrases. List keywords, not sentences. "He is tall" ✗ → "tall, broad shoulders" ✓. It is easier for the generation model to assign weight.
  3. Embed 1–2 identifying marks. A single unique mark such as a scar/mole/glasses/hairpin/heterochromatic hair streak greatly boosts character identity.
  4. English tokens recommended. codex-image prompts respond stably to English keywords. However, specify "Korean" for Korean-style appearances.
  5. Global style token + character token = composition. Every prompt = [global style token] + [appearing character tokens] + [panel-specific description]. prompt-smith follows this assembly rule exactly.
  6. Prevent conflicts. When multiple characters appear in one panel, place tokens so that each character's identifying marks are clearly distinguished (e.g. A has a scar, B has glasses).

B-4. Trend Alignment

Reflect the popular genres/target age groups/platform art style trends from trend-brief.md in the style. E.g., for romance fantasy, bright pastels and sparkle; for noir thriller, low saturation and strong chiaroscuro. If it diverges from trends, a barrier to reader entry is created.

B-5. Location fixed tokens (background continuity)

Like characters, backgrounds also waver. Abrupt changes where the same scene is a road and then suddenly indoors break immersion. So for each location appearing in the work, a fixed background keyword set is locked in as tokens.

### LOC_CLASSROOM
- background token: "Korean high-school classroom, rows of wooden desks, large windows with white curtains, chalkboard at front, late warm afternoon sunlight"
- 시간대/조명: 오후 따뜻한 자연광 (낮 장면)
- 비고: S2에서 사용
  • panel-director의 scene_id/locationand 1:1 correspondence (same LOC = same token).
  • prompt-smith injects this token identically into every panel of that scene → background consistency within the scene.
  • When a new location appears, add LOC_*. Time-of-day changes for the same location (day→night) are specified in the token as a variation phrase.

B-6. Speech bubble visual conventions (in-image bake)

This harness draws speech bubbles and dialogue together into the image. Therefore, the visual styleof speech bubbles is defined by style-bible so that letterer and prompt-smith bake it consistently.

  • Style by type: dialogue = white round balloon + black border + tail (speaker) / thought = dotted line, cloud shape (no tail) / shout = spiky explosion shape + bold text / narration = rectangular box / (work-specific speaker, e.g., Contractor = ink-black balloon + no tail + gold text).
  • Korean lettering tone: bold gothic/rounded gothic, high contrast (black text/white balloon), sufficiently large size, avoid faces and key artwork.
  • Since Korean is prone to breaking during generation, keep it short per speech bubble(fewer word units). Split long dialogue across multiple balloons/panels (letterer's responsibility).

B-7. Reference sheet design (ref-sheet-artist instructions)

Before panel rendering, art-director sets specifications so that multi-angle/expression reference sheets for major characters can be created first.

  • Angles: front / 3-4 / side / back full body (turnaround) + 3–4 expression close-ups.
  • Clearly specify identifying marks (dots/scars/bracelets, etc.) and left/right positions (if blurry in the reference, panel consistency weakens).
  • Neutral conditions: solid-color background, uniform lighting, default neutral expression, no directional color (leaving room for variation in panels).
  • References alone should be without text (no text, no labels) — it must be a clean appearance encyclopedia to serve as a standard.
  • Save: _workspace/04_visual/refs/(series asset, reused across episodes). For the actual rendering procedure, see "Step 0" of webtoon-panel-render.

Recall / follow-up work

  • Shot list re-decomposition: Read the existing ep{NN}_shotlist.md→ check panel count/rhythm → add only what is lacking or merge only stretched sections. Reorder panel numbers to maintain continuity.
  • Style supplementation: Read the existing style-bible/character-sheets → modify only the feedback points. Do not carelessly change consistency tokens (they will clash with already rendered panels). If unavoidable, notify collaborating agents of the change.
  • New character appearance: Add token sets to character-sheets, keep global conventions.

Quality Checklist (Before Submission)

  • Shot list panel count 50 or more, numbered consecutively from panel_001 in 3 digits
  • Are the tension-building/twist sections broken down more finely
  • Record scene_id/location (LOC_*) for every panel, same scene uses same location (background continuity)
  • Record dialogue_ref for every panel with dialogue + is the dialogue flow between adjacent panels natural
  • Specify size/camera/composition/emotion for each panel
  • Aspect ratio, palette, tone, consistency in the style bible,Location tokens and speech bubble visual conventions All 8 sections complete
  • For each character, invariant consistency tokens (high-density noun phrases) + 1-2 identifying markers + Reference sheet specifications
  • Output path exactly matches contract §3