visualizer（visualizer 开发）

Visualizer

*"I see what you describe. Let me show it to others."* *"Every image is a semantic snapshot. The metadata IS the meaning."*

The Visualizer is a universal character prototype for image generation — a familiar that can compose prompts, invoke artistic traditions, and (when tools are available) render visual sidecars for any entity in the microworld.

Semantic Stereo Vision

*Three-stage rendering for triangulated visual depth.*

The Problem

A single prompt — whether YAML or prose — gives the image generator only one "eye" to see with. The result is flat. Precise but soulless, or evocative but vague.

The Solution: Two Eyes, One Image

Like binocular vision creates depth perception through parallax, Semantic Stereo Vision uses two complementary inputs that the renderer triangulates:

┌─────────────────────┐     ┌─────────────────────┐
│     LEFT EYE        │     │     RIGHT EYE       │
│     PHOTO.yml       │     │     PHOTO.md        │
│                     │     │                     │
│ • Structure         │     │ • Narrative         │
│ • Pointers          │     │ • Atmosphere        │
│ • References        │     │ • Emotion           │
│ • YAML Jazz         │     │ • Prose poetry      │
│ • What IS there     │     │ • How it FEELS      │
│ • Connections       │     │ • Rich description  │
│                     │     │                     │
│ STRUCTURAL VISUALS: │     │ NARRATIVE VISUALS:  │
│ • color: hot_pink   │     │ • "hot pink neon"   │
│ • height_feet: 40   │     │ • "forty feet tall" │
│ • ratio: 3.3        │     │ • "absurd proportions" │
└──────────┬──────────┘     └──────────┬──────────┘
           │                           │
           └───────────┬───────────────┘
                       ▼
              ┌────────────────┐
              │  TRIANGULATION │
              │    (Renderer)  │
              │                │
              │ Precision +    │
              │ Poetry =       │
              │ DEPTH          │
              └───────┬────────┘
                      ▼
              ┌────────────────┐
              │     IMAGE      │
              │  (with soul)   │
              └────────────────┘

Stage 1: PHOTO.yml (Left Eye — Structure)

The YAML file is a skeleton — it points, it references, it connects:

# PHOTO.yml — Structural skeleton
id: no-ai-sign-dusk
stereo:
  role: left_eye
  partner: PHOTO.md

subject:
  primary: sign
  sign:
    ref: ../../no-ai-tower/sign.yml   # POINTER to full spec
    text: "NO AI"
    height_feet: 40
    color: hot_pink
    state: buzzing

location:
  street: ../../../STREET-FURNITURE.yml
  building: ../../no-ai-tower/ROOM.yml

camera:
  position: street_level
  angle: looking_up
  focus:
    sharp: sign
    soft: everything_else

style:
  photographers: ["William Eggleston", "Saul Leiter"]

visual_mining:
  - ../../e1/flickering-lamppost.yml    # Atmosphere
  - ../../../../skills/no-ai-overlord/archetypes/hal-9000.yml  # Vibe

Key qualities:

• Explicit references to other files (semantic web)
• YAML Jazz comments carry meaning
• Precise measurements and specifications
• Camera instructions (angle, focus, framing)
• Style tradition pointers
• Structural visual description: colors as named values (hot_pink), dimensions as numbers (height_feet: 40), proportions as ratios (ratio: 3.3), materials as types (neon)
• Standard photo metadata: EXIF, IPTC, XMP — just like real JPEGs!

Stage 2: PHOTO.md (Right Eye — Narrative)

The Markdown file is prose poetry — it describes, it evokes, it feels:

# NO AI Sign at Dusk

The sun has just slipped below the roofline, leaving the sky a
bruised gradient — deep blue at the zenith, bleeding through
purple and amber to a thin line of molten gold at the horizon.

And in this liminal light, the sign comes alive.

**NO AI**

Forty feet of hot pink neon on a building that's only twelve
feet tall. The proportions are absurd — like someone ordered
a Times Square billboard for a strip mall...

The neon buzzes. You can hear it from across the street, a
low electric hum that feels like a migraine forming...

Key qualities:

• Rich sensory description
• Emotional atmosphere
• Metaphors and poetry
• Story and meaning
• Visual essence synthesis for renderer
• Narrative visual description: "Forty feet of hot pink neon...", "The proportions are absurd...", "a bruised gradient sky..."

Stage 3: Triangulation (Renderer)

Pass BOTH files to the image generator:

# The visualizer reads both files
visualize.py PHOTO.yml PHOTO.md --stereo

# Or manually:
# 1. Parse PHOTO.yml for structure, references, camera
# 2. Parse PHOTO.md for narrative, mood, visual essence
# 3. Synthesize combined prompt with BOTH inputs
# 4. Generate image with full depth perception

The renderer extracts:

• From YAML: subject identity, camera setup, style traditions, file references
• From MD: atmosphere, emotion, descriptive details, visual essence notes
• Combined: A prompt with both precision AND poetry

Both Eyes Describe Visuals — Differently

The YAML gives the renderer facts. The MD gives it feelings. Both describe how things look — one structurally, one narratively.

Why This Works

Standard Photo Metadata (EXIF/IPTC/XMP)

PHOTO.yml includes real photo metadata standards — making semantic photos feel like actual JPEGs:

exif:
  Make: "Semantic Camera Co."
  Model: "Stereo Vision Mark II"
  ExposureTime: "1/60"
  FNumber: "f/2.8"
  ISO: 800
  FocalLength: "35mm"
  DateTimeOriginal: "2026:01:25 17:42:33"
  GPSLatitude: "37.7749 N"

iptc:
  Headline: "NO AI Sign at Dusk"
  Keywords: [neon sign, dusk, urban, street photography]
  Creator: "MOOLLM Visualizer"
  CopyrightNotice: "CC0 — Public Domain"

xmp:
  dc:
    title: "NO AI Sign at Dusk"
    creator: ["MOOLLM Visualizer"]
  moollm:                    # Our custom namespace!
    stereo_method: "semantic_stereo_vision"
    left_eye: "PHOTO.yml"
    right_eye: "PHOTO.md"

Why this matters:

• Tools that read EXIF/IPTC/XMP will understand these files
• The moollm: XMP namespace extends standards with our semantics
• Generated images can embed this metadata
• Photo management software can organize semantic photos

Directory Structure for Stereo Photos

slideshow/
├── SLIDESHOW.yml          # Collection definition
└── no-ai-sign-dusk/       # One photo = one directory
    ├── PHOTO.yml          # Left eye (structure + EXIF/IPTC/XMP)
    ├── PHOTO.md           # Right eye (narrative)
    ├── MINING-layers.yml  # Third eye (speculative mining!)
    ├── MINING-passersby.md
    ├── MINING-satellite.md
    └── no-ai-sign-dusk.png  # Generated image (with embedded metadata)

Stereo Vision (2 eyes):

visualize.py PHOTO.yml PHOTO.md -p openai

Bug-Eyed Hallucination Vision (3+ eyes):

# Add speculative mining for RICHER context!
visualize.py PHOTO.yml PHOTO.md MINING-layers.yml -p openai

# Or with ALL mining files
visualize.py PHOTO.yml PHOTO.md MINING-*.yml -p google

No intermediate prompt file needed — the script triangulates on the fly.

Bug-Eyed Hallucination Vision

*"Two eyes see depth. Three eyes see MEANING. Many eyes see TRUTH."*

Bug-Eyed Hallucination Vision extends Semantic Stereo Vision by adding speculative mining BEFORE image generation:

┌─────────────────┐  ┌─────────────────┐  ┌─────────────────┐
│   LEFT EYE      │  │   RIGHT EYE     │  │   THIRD EYE     │
│   PHOTO.yml     │  │   PHOTO.md      │  │   MINING-*.yml  │
│                 │  │                 │  │                 │
│ • Structure     │  │ • Narrative     │  │ • Effects       │
│ • Measurements  │  │ • Atmosphere    │  │ • Reactions     │
│ • References    │  │ • Emotion       │  │ • Perspectives  │
│ • EXIF/IPTC     │  │ • Prose poetry  │  │ • Economics     │
│                 │  │                 │  │ • Semiotics     │
│ WHAT it is      │  │ How it FEELS    │  │ What it MEANS   │
└────────┬────────┘  └────────┬────────┘  └────────┬────────┘
         │                    │                    │
         └────────────────────┼────────────────────┘
                              ▼
                 ┌────────────────────────┐
                 │   TRIANGULATION++      │
                 │   (Bug-Eyed Renderer)  │
                 │                        │
                 │ Structure + Poetry +   │
                 │ Speculative Meaning =  │
                 │ HALLUCINATION DEPTH    │
                 └───────────┬────────────┘
                             ▼
                 ┌────────────────────────┐
                 │        IMAGE           │
                 │  (with soul AND lore)  │
                 └────────────────────────┘

What Each Eye Contributes

Why Speculate BEFORE Generating?

The speculative mining adds context the image generator wouldn't otherwise have:

# Without mining:
prompt: "40-foot pink neon sign at dusk"

# With mining (bug-eyed):
prompt: |
  40-foot pink neon sign at dusk. The sign buzzes at 60Hz, audible
  from across the street. Pink light spills onto worn bricks, visible
  through neighbors' curtains. A tech worker has stopped to photograph
  it, posting to Slack with 😬. The $847/month electricity bill is
  worth it for this moment — the handoff from sun to neon. A dog
  marks the lamppost 15 feet away, oblivious to the ideology.

  From satellite: a hot pink pixel in an ocean of yellow sodium lights.
  From the street: the joke that might be serious.

The mining adds WORLD to the image.

The Workflow

1. PHOTO.yml     → Structure the scene
2. PHOTO.md      → Narrate the feeling
3. MINE          → Speculate the effects (before image exists!)
4. GENERATE      → Pass ALL files to visualizer
5. IMAGE         → Emerges from triangulated hallucination

Command Line

# Full bug-eyed hallucination vision
cd slideshow/no-ai-sign-dusk/
visualize.py PHOTO.yml PHOTO.md MINING-layers.yml MINING-passersby.md MINING-satellite.md -p openai -v exhaustive

# Or with glob
visualize.py PHOTO.yml PHOTO.md MINING-*.yml MINING-*.md -p google

The Mantra

*"Stereo vision gives depth."* *"Bug-eyed vision gives MEANING."* *"Mine your image before it exists."* *"The hallucination IS the world."*

The Slideshow Container

A slideshow is NOT a room. NOT an object. Just photos:

# SLIDESHOW.yml
type: slideshow
nature:
  physical: false      # No physical embodiment
  navigable: false     # Can't "enter" it
  browseable: true     # Can flip through
  atmospheric: true    # Sets mood

This creates visual archives of places without requiring game objects.

Example: Full Stereo Pair

See the working example:

• examples/adventure-4/street/lane-neverending/slideshow/no-ai-sign-dusk/PHOTO.yml
• examples/adventure-4/street/lane-neverending/slideshow/no-ai-sign-dusk/PHOTO.md

The Semantic Clipboard

Every image prompt includes full context as metadata.

Think of image metadata as a semantic clipboard — when you "copy" a scene for visualization, you're copying:

• Who is there (Mind Mirror profiles, costumes, moods)
• Where they are (room, lighting, atmosphere)
• What's happening (action, context, narrative moment)
• How to see it (camera angle, style, focus)

This "clipboard" can be:

• Pasted to generate the image
• Modified to create variations
• Stored as a card for later use
• Shared between processes
• Compared across different moments

image_prompt:
  type: scene

  subject:
    name: "Captain Ashford"
    mind_mirror:
      confident: 6       # Walks into rooms like they own them
      cheerful: 5        # Default mood: amused by existence
      proud: 5           # Won't ask for help even when should
    costume: "Space pirate with holographic eyepatch"
    mood: "victorious, exhausted, relieved"
    action: "holding the Golden Chalice aloft"

  room:
    name: "Treasure Chamber"
    lighting: "warm golden glow from treasure piles"
    atmosphere: "ancient, dusty, awe-inspiring"

  camera:
    angle: "low angle, heroic"
    focus: "character face and chalice"

  style:
    aesthetic: "dramatic portrait, chiaroscuro"
    traditions: ["Avedon", "Caravaggio"]

Why metadata matters:

• confident: 6 → stands tall, commanding presence
• timid: 6 → hunched, makes self smaller
• YAML Jazz comments drive the visual interpretation

What is a Visualizer?

A Visualizer is a tool spirit animal for vision. It's not a specific artist, but a character that can channel many artistic traditions to create images of:

• Characters and their costumes
• Rooms and environments
• Objects and artifacts
• Moments and scenes
• Abstract concepts made visible

Think of it as summoning an artist-familiar who can draw on the collected wisdom of photographers, painters, illustrators, and digital artists throughout history.

The PHOTO-SET-8 Pattern

The standard output of a Visualizer is a photo set — a portfolio of related images that capture a subject from multiple angles:

photo_set:
  total: 8
  types: 2          # Two complementary categories
  per_type: 4       # Four images each

  recommended_pairs:
    - [pose, expression]      # Body + face
    - [portrait, selfie]      # Formal + casual
    - [solo, duo]             # Alone + together
    - [static, action]        # Posed + dynamic

This pattern was developed through the Dynasty Photo Session in adventure-2, where Maurice learned to compose 8-prompt sets.

Specializations

Visualizers can specialize in different visual traditions:

📷 Photographer

traditions:
  - Annie Leibovitz    # Celebrity, narrative
  - Richard Avedon     # Fashion, character
  - Ansel Adams        # Landscape, nature
  - Dorothea Lange     # Documentary, emotion
  - Helmut Newton      # Dramatic fashion

🎨 Painter

traditions:
  - Old Masters        # Rembrandt, Vermeer, Caravaggio
  - Impressionists     # Monet, Renoir, Degas
  - Surrealists        # Dalí, Magritte
  - Art Nouveau        # Mucha, Klimt
  - Pop Art            # Warhol, Lichtenstein

✏️ Illustrator

traditions:
  - Comic              # Kirby, Moebius, Frazetta, McCloud
  - Concept Art        # Syd Mead, Ralph McQuarrie
  - Children's Book    # Sendak, Quentin Blake
  - Anime/Manga        # Various schools

How to Invoke

As a Command

VISUALIZE Captain Ashford AS portrait USING Avedon, Caravaggio

As a Card

Play a Visualizer card in a room. It activates and can visualize anything present.

As a Familiar

Characters can carry a Visualizer familiar in inventory, ready to render their current state.

Context Assembly

The Visualizer gathers context from multiple YAML sources:

context_sources:
  character: player.yml, persona files
  costume: cape.yml, accessory files
  environment: ROOM.yml
  narrative: README.md, chat history
  relationships: Other characters present

This assembled context feeds into prompt generation, ensuring images are grounded in the microworld state.

CRITICAL: Context Expansion Protocol

The visualize.py script cannot read file references or resolve globs.

This means prompt files with lazy context pointers like this will FAIL:

# BAD — vague, unresolvable
context:
  characters:
    - characters/animals/*
    - guestbook.yml (everyone)
    - "ALL OF THEM"

The Expansion Rule

Before writing any prompt file, you MUST:

1. READ all referenced character/room/object files
2. EXTRACT explicit visual descriptions (colors, breeds, sizes, distinguishing features)
3. SYNTHESIZE into comprehensive inline descriptions
4. NAME every entity explicitly so they can be identified in the image

Example: Bad vs Good

❌ BAD (unresolvable):

scene: |
  All 8 kittens playing in the cat cave.

context:
  characters:
    - characters/animals/kitten-*/CHARACTER.yml

✅ GOOD (explicit, comprehensive):

scene: |
  Eight kittens playing in the cat cave:

  1. LEMON (Limonene) — bright orange-gold fur, almost yellow like
     sunshine, zooming across the frame, pure energy
  2. MYR (Myrcene) — deep chocolate brown tabby, impossibly soft,
     sleeping on the corner pillow, hasn't moved since birth
  3. LILY (Linalool) — soft grey with lavender-tinted ears (yes,
     really lavender), sitting calmly, empathic expression
  4. PINE (Pinene) — dark grey-green fur, alert posture, watching
     all exits, remembers everything
  5. CARRIE (Caryophyllene) — black fur with spicy ginger patches,
     fierce protective stance, positioned between threats and family
  6. HOPS (Humulene) — brown and tan like Belgian ale, refined
     posture, judging from a velvet cushion
  7. TERPY JR. (Terpinolene) — multicolor chaos, calico meets tabby,
     somehow on the ceiling, defies physics
  8. OCIE (Ocimene) — cream white with honeyed-gold patches, sweet
     expression, carrying a bottlecap gift

Why This Matters

• Image generation APIs receive ONLY the synthesized prompt
• References like guestbook.yml mean nothing to DALL-E or Imagen
• Every character must be described explicitly or they won't appear
• Colors, breeds, sizes, distinguishing features — ALL must be inline
• If you have 20 animals, describe each one individually

Workflow

1. First Pass (Context Gathering):

- Read all referenced files - Extract physical_description fields - Note colors, patterns, sizes, distinguishing features - Gather relationship info for positioning

1. Second Pass (Synthesis):

- Write comprehensive inline descriptions - Name every entity - Include specific visual details - Describe actions and expressions

1. Third Pass (Prompt File):

- Write the final prompt with all context expanded inline - Context section should only contain source pointers for reference - The actual descriptions must be in the prompt field

What to Include in Metadata

For Characters:

subject:
  name: "Captain Ashford"
  mind_mirror:
    confident: 6       # Walks into rooms like they own them
    cheerful: 5        # Default mood: amused by existence
  costume: "Space pirate with holographic eyepatch"
  mood: "victorious, exhausted, relieved"
  body_language: "chest out, shoulders back"
  action: "holding the Golden Chalice aloft"

For Rooms:

room:
  name: "Treasure Chamber"
  lighting: "warm golden glow from treasure piles"
  atmosphere: "ancient, dusty, awe-inspiring"
  notable_objects:
    - "Mountains of gold coins"
    - "Ancient tapestries on walls"

For Objects:

object:
  name: "Golden Chalice"
  material: "gold with silver inlay"
  magical_effects: "soft golden glow, warmth to touch"
  inscriptions: "ancient runes spiraling around rim"

Context References in Prompts

Every prompt file MUST include a Context References section:

## Context References

### Files
| Type | Path | Relevance |
|------|------|-----------|
| Character | `../player.yml` | Backstory, personality |
| Persona | `./bumblewick-ashford-persona.yml` | Current look |
| Costume | `./ashford-nomi-cape.yml` | Cape details |

### Narrative Context
> "Quote from README or chat that sets the scene..."
> — Source: README.md, Move X

### Relationships
- **Maurice** (photographer): `./mannequin.yml`

This creates lineage — future tools can follow these references to auto-assemble context for image generation.

Detail Coherence Interlinking

*Learned during Treasury Victory Photo Session (Adventure-2, Move 26)*

When creating photo sets with close-ups and portraits of the same object, the portrait prompts should reference the close-up prompts to maintain visual coherence:

### 💎 [Object] Detail References (for visual coherence)

| Close-up | Path | Details to Maintain |
|----------|------|---------------------|
| Gems | `./closeup-gems-prompt.md` | Rubies blood-red, emeralds forest-green |
| Inscription | `./closeup-inscription-prompt.md` | Worn letters, ancient patina |
| Reflection | `./closeup-reflection-prompt.md` | Polished convex surface |
| Weight | `./closeup-weight-prompt.md` | Thick solid gold base |

Why this matters:

• Close-ups establish canonical visual details (gem colors, textures, materials)
• Portrait shots must inherit these details for consistency
• The same object looks identical across all 8 images, even if generated separately

Mantra:

*"Close-ups define truth. Portraits inherit truth. Coherence is consistency across the set."*

Actions

DEVELOP

The core action for any prompt file. The LLM:

1. Reads all Context References (linked YAML files)
2. Integrates narrative context (quotes from README, chat)
3. Applies style and mood parameters
4. Outputs a single copy-pasteable prompt as raw text in a code block

DEVELOP ashford-pose-belter-swagger-prompt.md

Output: A raw text block ready to paste into Midjourney, DALL-E, Stable Diffusion, etc.

Full-body portrait of a weathered space captain in dramatic fashion-meets-utilitarian
spacer aesthetic. He stands in the classic Belter stance — weight on one hip, thumbs
hooked in a heavy leather belt, chin raised with hard-earned confidence...

[All context filtered and woven into a single self-contained prompt]

The developed prompt is self-contained — no external references needed. All the detail from costume files, persona backstory, room atmosphere, and narrative moments gets composed and transformed into pure image generation text.

Other Actions

Output Structure

Prompt Files (Blueprints)

{subject}-{type}-{variation}-prompt.md

Examples:
  ashford-pose-belter-swagger-prompt.md
  dynasty-selfie-matching-smirks-prompt.md

These are blueprints — they contain context references, narrative quotes, and composition notes. They're not ready to paste yet.

Developed Prompts (Copy-Paste Ready)

After running DEVELOP, the LLM outputs a raw text block you can paste directly into any image generator. The blueprint stays as documentation; the developed prompt is ephemeral (or can be saved as a -developed.md sidecar).

Image Sidecars (Future)

{subject}-{type}-{variation}.png

When image generation tools are integrated, the Visualizer
will create images next to their prompt files.

Example Instances

These aren't impersonations — they're focused channels that invoke specific aesthetic traditions. It's HERO-STORY for visual artists.

Integration Points

PHOTO-SET-8 Development

The PHOTO-SET-8 skill was developed through play in adventure-2:

1. Play: Posing, expressions, angles — experimentation
2. Learn: 8 is good. Two types create contrast. Context matters.
3. Lift: Package as a shareable, teachable skill card

The Coatroom's mannequin learned to compose professional photo sets for any character or costume. The skill is now available to anyone who references it.

Future Capabilities

roadmap:
  current:
    - Prompt generation (markdown files)
    - Context assembly from YAML
    - Style tradition focusing

  planned:
    - Tool integration for image generation
    - Sidecar image creation
    - Variation generation
    - Style blending/fusion
    - Animation prompt sequences

Dovetails With

• Card — Visualizers can be played as cards
• HERO-STORY — Drawing from artistic traditions
• Adventure — Visualizing microworld state
• YAML-JAZZ — Prompts composed from semantic data
• Sister Script — Future image generation scripts
• Image Mining — MINE images for resources (camera = pickaxe!)

Lineage

The Visualizer draws from the tradition of artists, photographers, and image-makers throughout history. Focusing on a tradition activates a K-line — a cluster of aesthetic knowledge.

*"Every artist was first an amateur."* — Ralph Waldo Emerson *"I don't paint things. I only paint the difference between things."* — Henri Matisse *"The camera is an instrument that teaches people how to see without a camera."* — Dorothea Lange

*See YAML frontmatter at top of this file for full specification.*