Comicmaster

Workflow

Step 1: Story Plan (Agent generates this)

When asked to create a comic, generate a story_plan.json:

{
  "title": "The Last Robot",
  "style": "western",
  "preset": "dreamshaperXL",
  "reading_direction": "ltr",
  "characters": [
    {
      "id": "char_01",
      "name": "Kai",
      "visual_description": "young man, 25, short black messy hair, brown eyes, grey hoodie"
    }
  ],
  "panels": [
    {
      "id": "panel_01",
      "sequence": 1,
      "scene": "ruined cityscape at dawn",
      "action": "Kai stands on a rooftop looking over ruins",
      "characters_present": ["char_01"],
      "camera_angle": "low_angle",
      "shot_type": "wide",
      "mood": "melancholic",
      "narrative_weight": "high",
      "character_emotions": "contemplative",
      "character_poses": "standing at edge, hands in pockets",
      "lighting": "dramatic",
      "background_detail": "crumbling skyscrapers, orange dawn sky",
      "dialogue": [
        {"character": "char_01", "text": "Three years since the shutdown...", "type": "speech", "position_hint": "top_right"}
      ],
      "narration": "The world had changed."
    }
  ],
  "pages": []
}

Tip: Leave pages: [] to use dynamic auto-layout based on narrative_weight.

Step 2: Review

After generation, review the output pages (in output/comicmaster/{project}/pages/). Individual panels are in panels/, bubbled versions in panels_bubbled/.

Step 3: Post-Processing (optional)

# Upscale for print (768px → 1536px)
python3 scripts/upscale.py output/panels/ --scale 2.0

# Apply color grading for unified look
python3 scripts/color_grading.py output/panels/ --grade noir

Story Plan Schema

Required Fields

Optional Panel Fields

Dialogue Entry

{"character": "char_01", "text": "Hello!", "type": "speech", "position_hint": "top_left"}

Types: speech, thought, shout, whisper, narration, caption, sfx, explosion, electric, connected, scream Position hints: top_left, top_center, top_right, middle_left, middle_right, bottom_left, bottom_center, bottom_right

SFX entries support "rotation": -15 for angled text.

Bubble Types (11 total)

Dynamic Layouts (recommended)

Set narrative_weight per panel and leave pages: [] — sizes calculated automatically:

Fixed Templates (alternative)

Character Consistency (IPAdapter)

Automatically generates character reference sheets and uses IPAdapter for consistency:

1. Reference generation: 1024×1024 portrait per character
2. Upload: Ref uploaded to ComfyUI automatically
3. Per-panel conditioning: Weight adapts to shot type (0.4 wide → 0.8 close-up)
4. Auto preset switch: PLUS for body shots, PLUS FACE for close-ups
5. Multi-character: Chained IPAdapterAdvanced nodes with decreasing weights

Batch Optimization

For 6+ panels with IPAdapter, the pipeline automatically:

• Groups panels by IPAdapter requirements (none → single char → multi char)
• Sorts within groups by character combination
• Minimizes model switches (tested: 9→5 switches on 10 panels)
• Shows ETA and progress every 10 panels

Color Grading

Apply unified color palette across all panels:

python3 scripts/color_grading.py panels/ --grade noir
python3 scripts/color_grading.py --list  # show presets

Panel Upscaling

# Auto-detect best method (ComfyUI model → PIL fallback)
python3 scripts/upscale.py panel.png --scale 2.0 --method auto

# Batch upscale all panels
python3 scripts/upscale.py panels/ --scale 2.0

Presets (from ComfyUI skill)

Illustrious XL / NoobAI-XL Notes

These models use Danbooru tag-based prompts instead of natural language descriptions. When you set "preset": "illustriousXL" or "preset": "noobaiXL" in the story plan, the pipeline automatically:

1. Switches to tag-based prompting — character descriptions are converted to Danbooru tags (e.g., 1boy, short black hair, brown eyes, grey hoodie instead of prose)
2. Adds quality tags — masterpiece, best quality, newest, absurdres, highres prefix
3. Uses optimized negative prompts — model-specific negatives tuned for each architecture
4. Sets Clip Skip 2 — required for Illustrious models
5. Handles v-prediction — NoobAI-XL v-pred adds ModelSamplingDiscrete node automatically

When to use which:

See references/illustrious-xl-guide.md for detailed model comparison, download links, and prompting guide.

Export Formats

Files

Workflows (JSON for ComfyUI)

Loadable in ComfyUI UI → Load:

Style Guides

See references/style-guides.md for prompt templates per style (Western, Manga, Cartoon, Noir, Realistic).

Camera Angle Guide

Quality Tracking

Run after each comic generation to measure panel quality:

# Score all panels in a directory
python3 scripts/quality_tracker.py output/panels/ --output scores.json --verbose

# Score single image
python3 scripts/quality_tracker.py panel.png

# With composition analysis (center bias, thirds, flow, harmony, palette)
python3 scripts/quality_tracker.py output/panels/ --composition --verbose

# With panel-sequence analysis (flow continuity, pacing, shot variety)
python3 scripts/quality_tracker.py output/panels/ --composition --sequence --report

# Detailed text report
python3 scripts/quality_tracker.py output/panels/ --composition --sequence --report

Technical Metrics

Composition Metrics (--composition)

Colorimetry Metrics (--composition)

Sequence Metrics (--sequence)

Scoring System

Each panel gets:

• Technical Score (0-100): Weighted composite of sharpness, contrast, saturation, entropy, edges, exposure
• Composition Score (0-100, with --composition): Weighted composite of center bias, thirds, balance, harmony, palette
• Overall Score (0-100): 55% technical + 45% composition (or 100% technical without --composition)

Score labels: good (≥65), ok (≥40), poor (<40)

Outputs quality_scores.json with per-panel and aggregate scores. ~50ms/panel basic, ~120ms/panel with composition. Zero external dependencies (PIL + numpy only).

Performance Reference