Files
transcendedhacker 0f0f088ab2 v2.0
Introduces a production-ready, schema-driven prompt generation engine for ComfyUI, including a new 'everyday_casual' library with 290+ atomic options. Major updates include a comprehensive README, scaling and library comparison docs, a strict LICENSE (CC BY-NC 4.0), and a refactored __init__.py with improved node structure, negative prompt support, and detailed docstrings. The engine (engine/composer.py) now enforces strict separation of concerns, supports addons, intent flags, and deterministic assembly. Documentation and changelog updated to reflect the new scalable, maintainable architecture.
2026-01-19 03:09:18 +05:30

3.3 KiB

Changelog

v2.0.0 - Schema-Driven Architecture

Structural Changes

Before (v1.x):

  • 5-10 options per category
  • Full sentences in dropdown options
  • Approximately 80 total selections available
  • Tightly coupled engine and content

After (v2.0):

  • 300+ atomic options across 17 compositional blocks
  • Schema-driven architecture with strict separation of concerns
  • Alphabetically sorted dropdown navigation
  • "None" option for selective block skipping
  • Intent flag preparation for future explicit control

Architecture

Mode_personal_node/
├── engine/composer.py      → Deterministic assembly logic
├── schema/connector.json   → Structural contract and rules
└── libraries/*.json         → Content-only option datasets

Library Structure

Example library block counts (comprehensive_portrait.json):

  • subject: 5 options
  • age: 7 options
  • hair_color: 27 options
  • hair_style: 22 options
  • hair_ornament: 14 options
  • outfit: 29 options
  • outfit_detail: 10 options
  • pose: 20 options
  • mood: 15 options
  • location: 15 options
  • location_detail: 10 options
  • atmosphere: 15 options
  • lighting: 14 options
  • angle: 18 options
  • frame: 10 options
  • filter: 15 options
  • technical: 12 options

Total: 258 options in base library

Features

Alphabetical Sorting: All dropdown options auto-sorted A-Z for predictable navigation. "None" always appears first.

None Behavior: Selecting "None" completely omits that block from composition. No blank text, clean assembly.

Atomic Composition: Each option represents a single compositional element. Multiple selections compose into structured prompts.

Scalability: System architecture supports thousands of options across unlimited libraries without engine modification.

Example Composition

Input Selections:

  • subject: "Woman"
  • age: "23-27"
  • hair_color: "Platinum Blonde"
  • outfit: "Silk Slip Dress"
  • pose: "Over Shoulder"
  • (other blocks: None)

Output:

A woman 23-27 years old with platinum blonde hair, wearing red silk slip dress with spaghetti straps, (turning back to camera, looking over shoulder:1.2), photorealistic, sharp focus, highly detailed, 8k

Scaling Strategy

Current: 258 options in base library Capacity: System handles 1000+ options per library, unlimited libraries

Multi-Library Approach: For specialized contexts, create focused libraries:

  • comprehensive_portrait.json (general purpose)
  • cyberpunk_neon.json (genre-specific)
  • Additional libraries as needed

Switch library dropdown to change entire option context.

Technical Details

ComfyUI Limitation: Dropdowns cannot dynamically update based on other selections. Workaround is comprehensive libraries covering all combinations, or context-specific library switching.

Weighting: Model-specific weight multipliers applied to atmospheric and technical blocks (configurable in schema).

BREAK Tokens: Inserted at schema-defined positions for token priority control.

Caching: Option lists cached per library-block pair for performance.

Future Extensions

Content expansion planned from reference documentation:

  • Hair colors: 27 → 40+ options
  • Camera angles: 18 → 100+ options
  • Poses: 20 → 200+ options
  • Locations: 15 → 300+ options
  • Atmospheric conditions: 15 → 100+ options

All expansions require only library file edits, never engine modifications.