manga-generation - SKILL.md Agent Skill

name: Manga Generation description: Generate multi-panel manga from character reference and story beats triggers: - User wants to create a manga - User provides story beats or narrative - Character has been created and user wants a story model: gemini-3-pro-image-preview (Nano Banana Pro)

Manga Generation Skill

Generate sequential manga panels with visual continuity using Nano Banana Pro.

Capabilities

Sequential Generation: Panels generated in order, each using previous as reference
Visual Continuity: Same colors, lighting, environment maintained across panels
Camera Variety: Parses camera directions from story beats (close-up, wide, etc.)
Streaming Support: Progressive display as each panel completes (~25s each)
Multiple Styles: manga, webtoon, chibi, ghibli

Input Requirements

Parameter	Type	Required	Description
`character_image_path`	Path	Yes	Reference image for character consistency
`character_name`	str	Yes	Character name for prompts
`story_beats`	list[str]	Yes	2-6 visual descriptions (one per panel)
`dialogues`	list[str]	No	Optional dialogue for each panel
`style`	str	No	Visual style (default: manga)

Story Beat Format

Include camera direction in story beats for best results:

"Close-up: Momo's eyes widen in surprise"
"Wide shot: The room is a mess of scattered toys"
"Low angle: Momo leaps heroically toward the treat"
"Medium shot: Momo lands gracefully, treat in mouth"

Output

Streaming (recommended for UI)

async for event in generator.generate_manga_streaming(...):
    if event.type == 'panel':
        show_panel(event.data['image_url'])

Non-streaming

result = await generator.generate_manga(...)
for panel in result.panels:
    print(panel.image_url)

Usage Example

from skills.generate_manga.generate_manga import MangaGenerator

generator = MangaGenerator(client=gemini_client)

# Streaming (for progressive UI)
async for event in generator.generate_manga_streaming(
    character_image_path=Path("outputs/characters/momo_full_body.png"),
    character_name="Momo",
    story_beats=[
        "Wide shot: Momo spots a butterfly in the garden",
        "Medium shot: Momo crouches, ready to pounce",
        "Close-up: Momo's intense focus, whiskers twitching",
        "Low angle: Momo leaps into the air, paws extended"
    ],
    dialogues=["Ooh!", "Must... catch...", "", "GOTCHA!"],
    style="manga"
):
    if event.type == 'panel':
        # Display panel immediately
        display(event.data['image_url'])

Technical Notes

Aspect Ratio: 9:16 vertical (portrait orientation)
Generation Time: ~25-30 seconds per panel
Image Chaining: Previous panel passed as reference for continuity
Output Location: assets/outputs/manga/{manga_id}_panel_{n}.png

Character Consistency (Dual Anchor)

Before generating panels, Gemini Pro (gemini-3-pro-preview) analyzes each character reference sheet and produces a 2-3 sentence appearance description. This runs once per character (~2s), not per panel.

Each panel prompt includes two anchors per character:

Visual reference: Character sheet image passed as [IMAGE N]
Text description: APPEARANCE: VibeCoder is an older man with long grey hair, full grey beard, black hoodie, gold headphones...

This prevents the image model from hallucinating different characters on complex story beats.

Visual Continuity

For panels 2+, the prompt includes:

Previous panel as visual reference ([IMAGE N - PREVIOUS PANEL])
Explicit instructions to match colors, lighting, environment
"NEVER replace a character with a different person or animal"

This ensures consistent backgrounds, lighting, color palette, and character identity across all panels.