用Comfy CLI运行生成式媒体工作流

CLI for using ComfyUI

comfy-cli: A Command Line Tool for ComfyUI comfy-cli is a command-line tool for installing, running, and extending ComfyUI — the open-source generative-media engine. Set up ComfyUI, install custom nodes and models, run workflows, and call hosted partner image models, all from your terminal. Demo Features 🚀 One-command ComfyUI install and launch 🎨 Direct calls to partner image and video nodes (Flux, Ideogram, DALL·E, Recraft, Reve, Gemini/nano-banana, Kling, Luma, Runway, Pika, Vidu, Hailuo, Seedance, …) via comfy generate, no workflow JSON required 🔧 Custom node management — install, update, snapshot, bisect 📦 Fast dependency resolution with uv (--fast-deps, --uv-compile) 🗄️ Model downloads from CivitAI, Hugging Face, and direct URLs 🎬 Run workflows against a local ComfyUI server, including auto-conversion of UI-format JSON 🧪 Test ComfyUI and frontend pull requests with one flag 💻 Cross-platform: Windows, macOS, Linux ☁️ Route any workflow to Comfy Cloud with --where cloud (no GPU required) 🤖 Agent-friendly: every command emits structured --json envelopes 📚 Bundled skills teach Claude / Cursor to drive comfy natively Quick Start pip install comfy-cli comfy setup comfy setup walks you through everything — local or cloud routing, authentication, and agent skill installation — in one interactive wizard. Pass -y for non-interactive (CI/scripted) installs. Run a workflow in the cloud No local GPU? Route any API-format workflow to Comfy Cloud with --where cloud: comfy cloud login # sign in via your browser (OAuth) comfy run --workflow ./workflow.json --where cloud # submits, prints a prompt_id, returns immediately comfy jobs wait <prompt_id> --where cloud # block until the job finishes (or: jobs status for a one-shot check) comfy download <prompt_id> --where cloud -o ./outputs # save the results locally comfy run submits asynchronously and prints the prompt_id you feed to jobs/download; add --wait to block inline instead. Check your sign-in anytime with comfy cloud whoami. Export the workflow JSON from ComfyUI via File → Export (API) (UI-format JSON is auto-converted). Set cloud as your default target so you can drop the flag: comfy set-default --where cloud. Credits: cloud generation (comfy run --where cloud, comfy generate) consumes Comfy Cloud credits and needs an active subscription. Discovery and inspection commands — comfy cloud whoami, comfy cloud status, comfy jobs status/ls, comfy templates ls, comfy generate list — don't. Check your balance and tier with comfy cloud status. Installation (Recommended) Activate a virtual environment (venv or conda). Install with pip (requires Python 3.10+): pip install comfy-cli Shell Autocomplete Install shell completion so comfy <TAB> expands commands and options: comfy --install-completion Usage Installing ComfyUI To install ComfyUI using comfy, simply run: comfy install This command will download and set up the latest version of ComfyUI and ComfyUI-Manager on your system. If you run in a ComfyUI repo that has already been setup. The command will simply update the comfy.yaml file to reflect the local setup comfy install --skip-manager: Install ComfyUI without ComfyUI-Manager. To use a custom Manager fork or specific version, skip the default installation and install your own into the workspace venv: comfy install --skip-manager # Then install your custom Manager: pip install -e /path/to/your-manager-fork # editable install # or pip install comfyui-manager==4.1b8 # specific version comfy --workspace=<path> install: Install ComfyUI into <path>/ComfyUI. comfy install --fast-deps: Use uv instead of pip for faster dependency resolution during initial ComfyUI installation. comfy-cli's built-in resolver compiles all requirements (core + custom nodes) into a single lockfile and installs from it. Also handles GPU-specific PyTorch wheel selection automatically. For comfy install, if no path specification like --workspace, --recent, or --here is provided, it will be implicitly installed in <HOME>/comfy. Python environment handling When you run comfy install, comfy-cli picks a Python environment for ComfyUI dependencies using the following precedence: An active virtualenv or conda environment (VIRTUAL_ENV / CONDA_PREFIX) is used as-is. An existing .venv or venv directory inside the workspace is reused. Otherwise the choice depends on how comfy-cli was installed: pip install comfy-cli (global / system Python): dependencies go directly into the same Python environment. This is the typical Docker setup. pipx install comfy-cli or uv tool install comfy-cli (isolated tool environment): a .venv is created inside the ComfyUI workspace. Use comfy launch to start ComfyUI with the correct Python. Updating ComfyUI comfy update brings an existing workspace up to date: comfy update (or comfy update comfy): pull the branch the workspace is currently on and reinstall requirements.txt. comfy update all: also update every installed custom node. comfy update cli: upgrade comfy-cli itself. By default comfy update all exits 0 even when the custom-node update step fails — the error is printed, but the exit code says success, so scripts and wrappers can't tell. Pass --exit-on-fail (the same flag name and default as comfy node install --exit-on-fail) to have that failure exit non-zero instead. The flag only changes the all target; comfy update comfy and comfy update cli already exit non-zero when they fail, and accept the flag as a no-op so it can be forwarded unconditionally. One limit worth knowing: --exit-on-fail can only surface what ComfyUI-Manager's cm-cli update reports. cm-cli currently handles a single pack failing to update by printing ERROR: ... and carrying on, still exiting 0, so that particular case stays invisible until ComfyUI-Manager propagates it. Unlike comfy node install, the flag is not forwarded to cm-cli — only its install subcommand accepts --exit-on-fail; its update subcommand has no such option and would reject it. When the flag does fire, the exit code is cm