comfyui-workflow（comfyui 工作流程）

ComfyUI Workflows — Agent Usage Guide

⚠️ READ-ONLY WARNING: The scripts in this skill (comfy_run.py, comfy_api.py) are production-validated and must not be modified. They handle 33+ workflow JSONs with complex LiteGraph→API conversion, subgraph expansion, bypass resolution, and multi-format output downloading. Any modification risks breaking all workflows. Only read and use the scripts — never edit them.

Requirements

Before using this skill, ensure your environment meets these requirements:

Install Python dependencies:

pip install websocket-client requests

Configuration Options

Credentials

Personal Configuration Required

Before using this skill, ensure these items are configured for your environment:

Quick Setup

1. Start ComfyUI

Start your ComfyUI server manually before executing any workflows. The skill expects ComfyUI to be running at the configured host.

2. Configure Connection

Create scripts/config.json:

{
  "comfy_host": "127.0.0.1:8188",
  "workflow_dirs": []
}

Or set environment variables:

export COMFY_HOST="127.0.0.1:8188"

3. Install Dependencies

pip install websocket-client requests

4. Add Your Workflows

Place workflow JSON files in workflows/ directory, organized by category:

workflows/
├── Image-Text/
│   ├── workflow1.json
│   └── workflow2.json
├── Image-Edit/
│   └── ...
└── Video/
    └── ...

Shell Variables (use in all commands)

VENV=python3  # or path to your venv
SCRIPT=/path/to/comfyui-workflow-skill/scripts/comfy_run.py

End-to-End Workflow (Steps 0–5)

Step 0 — Ensure ComfyUI Is Running

Start ComfyUI manually. Verify it's accessible at http://127.0.0.1:8188.

Step 1 — List Available Workflows

$VENV $SCRIPT --list

Step 2 — Inspect the Workflow

$VENV $SCRIPT -w "workflow_name" --inspect

Step 3 — Prepare Inputs

--prompt "your prompt"           # Text prompts
--image /path/to/image.jpg       # Image inputs
--audio /path/to/audio.wav       # Audio inputs
--video /path/to/video.mp4       # Video inputs
--width 1024 --height 1024       # Resolution
--steps 50 --cfg 4.0 --seed 42   # Sampler settings
--override '{"node_id": {"key": value}}'  # Advanced overrides

Step 4 — Execute

$VENV $SCRIPT -w "workflow_name" --prompt "..." -o /tmp

Step 5 — Collect Outputs

Outputs are saved to the -o directory with auto-generated prefixes.

CLI Reference

Usage: comfy_run.py [-w WORKFLOW] [options]

Modes:
  --list              List all available workflows
  --inspect           Human-readable input/output inspection
  --inspect-json      Machine-readable JSON inspection
  --dry-run           Print API JSON without executing
  --queue             Submit to queue and return immediately
  --status [IDs...]   Check queue state

Input customization:
  --prompt, -p TEXT        Positive prompt
  --negative, -n TEXT      Negative prompt
  --image, -i PATH [...]   Input image(s)
  --audio, -a PATH [...]   Input audio file(s)
  --video, -v PATH [...]   Input video file(s)

Generation settings:
  --width, -W INT          Output width
  --height, -H INT         Output height
  --steps INT              Sampling steps
  --cfg FLOAT              CFG scale
  --seed INT               Random seed

Advanced:
  --override JSON          Node input overrides
  --config PATH            YAML/JSON config file
  --server HOST:PORT       ComfyUI server
  --output-dir, -o PATH    Output directory
  --output-prefix TEXT     Output filename prefix
  --no-ensure              Skip auto-starting ComfyUI
  --no-validate            Skip input validation

Troubleshooting

Knowledge Cache Pattern

MODELS_INVENTORY.md and WORKFLOWS_SUMMARY.md serve as the agent's persistent knowledge base for the user's actual ComfyUI setup. Use them to cache learned information across sessions:

⚠️ DO NOT REFERENCE THE EXAMPLE CONTENT: The default content in these files describes sanitized example workflows that will NOT work. Agents MUST update them with the user's actual data before relying on them. Using outdated example data will cause workflow failures.

Initial Setup (First Time)

When the user first provides their own workflow files:

1. Clean the example content — Remove the sanitized example entries from both files
2. Populate with user's data — Run --inspect on each workflow, document findings in WORKFLOWS_SUMMARY.md
3. Inventory models — Scan user's model directories, update MODELS_INVENTORY.md with their actual models

Ongoing Maintenance

• After using a new workflow → Add its documentation to WORKFLOWS_SUMMARY.md
• After downloading models → Update MODELS_INVENTORY.md status
• When learning model-specific quirks → Document tips/tricks in the relevant section

Session Workflow

1. Read MODELS_INVENTORY.md → know available models
2. Read WORKFLOWS_SUMMARY.md → know workflow capabilities
3. Select appropriate workflow without re-inspecting
4. If new workflow discovered → document it for future sessions

This pattern lets the agent skip repeated discovery and work efficiently with cached knowledge.

Reference Documents