Architecture

한국어 버전

This project is a small Python/ComfyUI bridge for OpenToonz-inspired line preparation. It intentionally avoids OpenToonz C++ linkage.

Module Map

ComfyUI-OpenToonzLineTools/
  __init__.py                       ComfyUI custom-node entrypoint
  opentoonz_line_tools/
    cleanup.py                      blue rough line extraction
    autoclose.py                    endpoint-based gap closing
    regions.py                      fillable region labeling
    nodes.py                        ComfyUI node classes
  tests/
    test_line_tools.py              smoke tests for core algorithms
  docs/
    architecture.md / .ko.md
    visual_pipeline.md / .ko.md

Design Rules

Keep core algorithms independent from ComfyUI tensors.
Keep ComfyUI wrappers thin: tensor conversion, parameter binding, JSON output.
Prefer Python/OpenCV reimplementation over direct C++ binding.
Emit visual previews for every destructive-looking preprocessing step.
Preserve JSON artifacts so the panel splitter pipeline can inspect or replay decisions.

Data Flow

flowchart TD
    A[ComfyUI IMAGE tensor] --> B[nodes.py tensor adapter]
    B --> C[cleanup.py blue line mask]
    C --> D[clean line RGB]
    D --> E[autoclose.py gap closure]
    E --> F[regions.py label map]
    F --> G[preview IMAGE tensor]
    F --> H[regions JSON]

Implementation Notes

Cleanup

cleanup.py uses HSV blue classification plus morphology and connected-component despeckling. This is the practical counterpart of OpenToonz cleanup/color-line processing for blue rough manuscripts.

AutoClose

autoclose.py uses a small Zhang-Suen thinning implementation, endpoint detection, local direction estimation, and distance/angle filtering. This mirrors the useful OpenToonz idea: close small gaps before paint/fill operations.

Region Map

regions.py labels connected fillable regions outside the line mask and emits simple palette-style metadata. It is not a full Toonz Raster palette-index format yet, but it creates the data boundary needed for one.

Extension Plan

flowchart LR
    A[v0 smoke nodes] --> B[v1 batch panel workflow]
    B --> C[v2 palette JSON correction]
    C --> D[v3 reference color matching]
    D --> E[v4 ComfyUI reassemble integration]

Verification

Current verification:

/Users/iwongyeong/AI/ComfyUI/.venv/bin/python -m unittest discover -s tests -v

The tests cover:

blue line extraction with despeckling,
small horizontal gap closure,
closed-region labeling.

2.5 KiB Raw Permalink Blame History