Architecture¶

Three-Layer Architecture¶

The installer uses a bootstrap + two-phase architecture:

Layer	File	Purpose
Bootstrap	`Install.bat` / `Install.sh`	Downloads `uv`, creates venv with Python 3.11–3.13, installs CLI
Phase 1	`src/installer/phase1.py`	System checks, venv/conda setup, tool installs (aria2, git)
Phase 2	`src/installer/phase2.py`	ComfyUI clone, junctions, pip packages, custom nodes, launchers

Junction-Based Data Separation¶

User data (models, outputs, custom nodes) lives in external folders, linked into ComfyUI via junctions/symlinks. This allows git pull updates without data loss.

install_path/
├── scripts/venv/            # Python virtual environment
├── ComfyUI/                 # Git repo (can be wiped for updates)
│   ├── models/ → ../models  # ← junction (symlink)
│   ├── output/ → ../output  # ← junction
│   └── main.py
├── models/                  # ← User data (persists)
├── output/                  # ← User data (persists)
├── logs/                    # Install and update logs
└── UmeAiRT-Start-ComfyUI.bat

Module Overview¶

graph TB
    CLI[cli.py] --> Install[installer/install.py]
    CLI --> Update[installer/updater.py]
    CLI --> Download[downloader/engine.py]
    CLI --> Scanner[utils/model_scanner.py]

    Install --> System[installer/system.py]
    Install --> Env[installer/environment.py]
    Install --> Repo[installer/repository.py]
    Install --> Deps[installer/dependencies.py]
    Install --> GPU[installer/gpu_setup.py]
    Install --> Nodes[installer/nodes.py]
    Install --> Opt[installer/optimizations.py]
    Install --> Final[installer/finalize.py]

    subgraph Utils
        Cmd[utils/commands.py]
        DL[utils/download.py]
        Log[utils/logging.py]
        Pkg[utils/packaging.py]
        GPUUtil[utils/gpu.py]
    end

    subgraph Platform
        Base[platform/base.py]
        Win[platform/windows.py]
        Lin[platform/linux.py]
        Mac[platform/macos.py]
    end

    Install --> Utils
    Install --> Platform

Configuration¶

All dependencies and URLs live in scripts/dependencies.json, validated by Pydantic models in src/config.py.

Custom nodes are managed via scripts/custom_nodes.json using an additive-only manifest: the installer never removes user-installed nodes.