ComfyUI Command-Line Arguments: A Complete Reference Guide

ComfyUI exposes approximately 53 command-line arguments defined in comfy/cli_args.py that control server networking, GPU device selection, VRAM management strategies, model precision formats, and UI customization.

ComfyUI, the open-source node-based diffusion GUI maintained by Comfy-Org, provides extensive configuration flexibility through its command-line interface. Understanding the available ComfyUI command-line arguments is essential for optimizing performance across different hardware configurations, securing server deployments, and customizing directory structures. The argument parser initializes these options in comfy/cli_args.py before PyTorch loads, ensuring hardware-specific environment variables are set correctly during early initialization in main.py.

Server and Networking Configuration

The HTTP/WebSocket server behavior is controlled through several networking flags parsed in comfy/cli_args.py (lines 38-43) and implemented in server.py.

• --listen – IP address (or CSV list) for the UI to bind. Defaults to 127.0.0.1. Supplying the flag without a value binds to all IPv4/IPv6 interfaces.
• --port – TCP port for the HTTP server (default 8188).
• --tls-keyfile / --tls-certfile – Paths to TLS key and certificate files for enabling HTTPS.
• --enable-cors-header – Adds Cross-Origin Resource Sharing headers; accepts an optional origin or * for all origins.
• --max-upload-size – Maximum file upload size in MiB (default 100).

Directory and File Path Configuration

ComfyUI’s directory layout can be completely customized using arguments processed by folder_paths.py and utils/extra_config.py.

• --base-directory – Root folder containing models, custom_nodes, input, output, temp, and user subdirectories.
• --extra-model-paths-config – Load one or more extra_model_paths.yaml files to register additional model search paths.
• --output-directory, --temp-directory, --input-directory – Override specific subdirectories of the base path.
• --user-directory – Explicit path for user data (overrides --base-directory).

Hardware and Device Selection

Device selection arguments in comfy/cli_args.py (lines 52-53, 88-92) configure CUDA, DirectML, and Intel oneAPI backends before model loading.

• --cuda-device – Restrict the process to a single CUDA device index, hiding other GPUs from PyTorch.
• --default-device – Select a default CUDA device while keeping other devices visible.
• --directml – Use torch-directml for AMD/Intel GPU support on Windows; accepts optional device index (default -1).
• --oneapi-device-selector – Set Intel oneAPI device selector string for XPU acceleration.
• --disable-ipex-optimize – Disable automatic ipex.optimize for Intel extensions.
• --supports-fp8-compute – Force-enable FP8 compute support flags for testing purposes.

Precision and Model Format Options

Model precision can be fine-tuned per component (UNet, VAE, text encoder) through flags defined in comfy/cli_args.py (lines 60-86).

• --force-fp32, --force-fp16 – Force the entire pipeline to FP32 or FP16 precision.
• UNet precision – --fp32-unet, --fp64-unet, --bf16-unet, --fp16-unet, --fp8_e4m3fn-unet, --fp8_e5m2-unet, --fp8_e8m0fnu-unet.
• VAE precision – --fp16-vae, --fp32-vae, --bf16-vae, or --cpu-vae to offload VAE decoding to CPU (useful for low-VRAM GPUs).
• Text encoder precision – --fp8_e4m3fn-text-enc, --fp8_e5m2-text-enc, --fp16-text-enc, --fp32-text-enc, --bf16-text-enc.
• --force-channels-last – Request channels-last tensor layout for improved performance on some GPU architectures.

Memory Management and VRAM Optimization

VRAM handling strategies are critical for stable diffusion workloads. These arguments (lines 37-44, 145-169 in comfy/cli_args.py) configure how models are loaded and offloaded.

• VRAM strategies – --gpu-only, --highvram, --normalvram, --lowvram, --novram, --cpu define where models reside (GPU only, aggressive offloading, or CPU-only).
• --reserve-vram – Reserve a fixed amount of VRAM (in GB) for the operating system and other applications.
• Async offloading – --async-offload enables async weight offloading with 2 streams by default; --disable-async-offload turns it off.
• --disable-pinned-memory – Disable pinned-memory usage for CPU-GPU transfers.
• Memory mapping – --mmap-torch-files enables memory-mapped loading for large checkpoints; --disable-mmap forces standard file loading.
• --disable-smart-memory – Force aggressive offloading to RAM rather than keeping models in VRAM between executions.
• --deterministic – Enable deterministic PyTorch algorithms for reproducibility (slower performance).

Performance Tuning and Optimization Flags

Advanced performance options control cross-attention implementations and experimental optimizations.

• --fast – Enable experimental optimizations including fp16_accumulation, fp8_matrix_mult, cublas_ops, autotune, and dynamic_vram. Supplying no arguments enables all optimizations.
• Cross-attention selection – --use-split-cross-attention, --use-quad-cross-attention, --use-pytorch-cross-attention, --use-sage-attention, --use-flash-attention.
• --disable-xformers – Explicitly disable Xformers attention kernels if automatically detected.

Security and Custom Node Controls

These flags manage the ComfyUI-Manager and custom node security posture.

• --enable-manager – Activate the ComfyUI-Manager plugin.
• --disable-manager-ui / --enable-manager-legacy-ui – Control manager UI visibility while keeping background tasks active.
• --disable-all-custom-nodes – Prevent loading any third-party custom node directories (useful for CI/CD).
• --whitelist-custom-nodes – Comma-separated list of custom node folders to load even when globally disabled.
• --disable-api-nodes – Block loading of API-exposing nodes and prevent internet calls from the frontend.

Preview, Frontend, and Miscellaneous Options

Additional flags control UI behavior, logging, and database configuration.

• --preview-method / --preview-size – Choose preview rendering (none, auto, latent2rgb, taesd) and set maximum preview resolution.
• --auto-launch / --disable-auto-launch – Control automatic browser opening on startup.
• --multi-user – Enable per-user storage isolation (creates separate subdirectories for each OS user).
• --verbose / --log-stdout – Set logging level (DEBUG through CRITICAL) and redirect logs to stdout.
• --front-end-version / --front-end-root – Select specific frontend versions or point to local UI builds.
• --database-url – Override the SQLite database location or use in-memory storage.
• --disable-assets-autoscan – Skip automatic model asset database population at startup.
• --enable-compress-response-body – Enable HTTP response compression (gzip/deflate).
• --quick-test-for-ci – Exit immediately after successful startup for CI pipeline validation.
• --windows-standalone-build – Enable conveniences for the Windows standalone binary distribution.

Practical Usage Examples

Configure ComfyUI for different deployment scenarios using these command patterns:

# Remote server with HTTPS, custom port, and CORS

python main.py --listen 0.0.0.0 --port 8888 --tls-keyfile key.pem --tls-certfile cert.pem --enable-cors-header "*"

# Low-VRAM GPU with CPU-offloaded VAE and async loading

python main.py --lowvram --cpu-vae --async-offload

# Deterministic FP16 workflow with specific CUDA device

python main.py --cuda-device 1 --force-fp16 --deterministic

# External storage layout with disabled custom nodes for testing

python main.py --base-directory /mnt/comfy_data --output-directory /mnt/comfy_data/output --disable-all-custom-nodes

# Enable all experimental fast optimizations

python main.py --fast

# Specific frontend version with preview enabled

python main.py --front-end-version Comfy-Org/ComfyUI_frontend@latest --preview-method latent2rgb --preview-size 512

Summary

• Configuration Source: All arguments are defined in comfy/cli_args.py and parsed before PyTorch initialization in main.py.
• Hardware Control: Use --cuda-device, --directml, or --oneapi-device-selector to specify compute backends; VRAM strategies (--lowvram, --highvram, etc.) control memory residency.
• Precision Tuning: Component-specific flags (--fp16-unet, --cpu-vae, --bf16-text-enc) allow fine-grained control over model formats.
• Security: --disable-api-nodes and --whitelist-custom-nodes provide granular control over custom node execution and network access.
• Performance: The --fast flag bundles experimental optimizations, while specific attention implementations can be selected via --use-*-cross-attention flags.

Frequently Asked Questions

How do I bind ComfyUI to all network interfaces?

Use the --listen flag without specifying an IP address. According to the source code in comfy/cli_args.py (line 38), supplying --listen alone binds to all IPv4 and IPv6 addresses, while --listen 0.0.0.0 explicitly binds to all IPv4 interfaces.

What is the difference between --cuda-device and --default-device?

--cuda-device restricts the process to a single GPU by hiding other CUDA devices from PyTorch entirely, while --default-device merely sets a preferred default device index while keeping all GPUs visible to the application. Both are processed in comfy/cli_args.py (lines 52-53) before the CUDA runtime initializes.

How can I run ComfyUI on systems with very limited VRAM?

Combine --lowvram or --novram with --cpu-vae to offload the diffusion model with aggressive memory management and force VAE decoding to CPU. For extreme constraints, use --cpu to run entirely on system memory, though this significantly reduces performance.

Where does ComfyUI store its configuration and model files by default?

By default, ComfyUI uses relative directories (models, output, input, etc.) in the installation root. Override this using --base-directory to specify a new root, or use individual flags like --output-directory and --user-directory for specific paths. The folder_paths.py module handles these path resolutions during startup.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}