Command Line Arguments

ComfyUI supports numerous command line arguments to customize server behavior, hardware configuration, memory management, and performance optimizations.

Running ComfyUI with Arguments

Pass arguments when launching ComfyUI:

python main.py --listen 0.0.0.0 --port 8188 --preview-method auto

Server Configuration

--listen

string

default:"127.0.0.1"

Specify the IP address to listen on. Use 0.0.0.0 to listen on all IPv4 addresses, or provide a comma-separated list.

# Listen on localhost only (default)
python main.py --listen 127.0.0.1

# Listen on all interfaces (IPv4 and IPv6)
python main.py --listen

# Listen on specific IPs
python main.py --listen 127.0.0.1,192.168.1.100

Using --listen without an argument defaults to 0.0.0.0,:: (all IPv4 and IPv6 addresses).

--port

integer

default:"8188"

Set the listen port for the web server.

python main.py --port 3000

--tls-keyfile

string

Path to TLS (SSL) key file. Enables HTTPS access. Requires --tls-certfile.

python main.py --tls-keyfile key.pem --tls-certfile cert.pem

--tls-certfile

string

Path to TLS (SSL) certificate file. Enables HTTPS access. Requires --tls-keyfile.

--enable-cors-header

string

default:"*"

Enable CORS (Cross-Origin Resource Sharing) with optional origin.

# Allow all origins
python main.py --enable-cors-header

# Allow specific origin
python main.py --enable-cors-header "https://example.com"

--max-upload-size

float

default:"100"

Set the maximum upload size in MB.

python main.py --max-upload-size 500

--auto-launch

boolean

Automatically launch ComfyUI in the default browser on startup.

python main.py --auto-launch

--disable-auto-launch

boolean

Disable auto-launching the browser (overrides --auto-launch).

--dont-print-server

boolean

Suppress server output messages.

--enable-compress-response-body

boolean

Enable response body compression to reduce bandwidth usage.

Directory Configuration

--base-directory

string

Set the ComfyUI base directory for models, custom_nodes, input, output, temp, and user directories.

python main.py --base-directory /path/to/comfyui

--output-directory

string

Set the output directory for generated images. Overrides --base-directory.

python main.py --output-directory /path/to/outputs

--input-directory

string

Set the input directory for source images. Overrides --base-directory.

python main.py --input-directory /path/to/inputs

--temp-directory

string

Set the temporary files directory. Overrides --base-directory.

python main.py --temp-directory /tmp/comfyui

--user-directory

string

Set the user directory with an absolute path. Overrides --base-directory.

python main.py --user-directory /path/to/user

--extra-model-paths-config

string

Load one or more extra_model_paths.yaml files. Can be specified multiple times.

python main.py --extra-model-paths-config /path/to/config1.yaml --extra-model-paths-config /path/to/config2.yaml

GPU and Device Configuration

--cuda-device

integer

Set the CUDA device ID to use. Makes all other devices invisible.

python main.py --cuda-device 0

--default-device

integer

Set the default device ID while keeping other devices visible.

python main.py --default-device 0

--directml

integer

default:"-1"

Use torch-directml for AMD/Intel GPUs on Windows.

# Use default DirectML device
python main.py --directml

# Use specific DirectML device
python main.py --directml 0

--oneapi-device-selector

string

Sets the oneAPI device selector for Intel GPUs.

python main.py --oneapi-device-selector "level_zero:0"

--disable-ipex-optimize

boolean

Disables ipex.optimize when loading models with Intel’s Extension for PyTorch.

--supports-fp8-compute

boolean

Force ComfyUI to act as if the device supports FP8 compute.

VRAM Management

--gpu-only

boolean

Store and run everything (text encoders, CLIP, VAE, etc.) on the GPU. Requires high VRAM.

python main.py --gpu-only

--highvram

boolean

Keep models in GPU memory after use instead of unloading to CPU. Requires ~12GB+ VRAM.

python main.py --highvram

--normalvram

boolean

Force normal VRAM usage mode (used if lowvram is automatically enabled).

python main.py --normalvram

--lowvram

boolean

Split the UNet model to use less VRAM. For GPUs with 4-8GB VRAM.

python main.py --lowvram

--novram

boolean

Extreme memory saving mode. Use when lowvram isn’t enough.

python main.py --novram

--cpu

boolean

Run everything on the CPU. Very slow but works without a GPU.

python main.py --cpu

--reserve-vram

float

Reserve amount of VRAM (in GB) for OS/other software.

python main.py --reserve-vram 2.0

--disable-smart-memory

boolean

Force aggressive offloading to RAM instead of keeping models in VRAM when possible.

--disable-dynamic-vram

boolean

Disable dynamic VRAM management and use estimate-based model loading.

--async-offload

integer

default:"2"

Use async weight offloading with specified number of streams. Enabled by default on NVIDIA.

# Use default 2 streams
python main.py --async-offload

# Use 4 streams
python main.py --async-offload 4

--disable-async-offload

boolean

Disable async weight offloading.

Precision and Data Types

Global Precision

--force-fp32

boolean

Force FP32 precision globally. Use if your GPU has issues with FP16.

python main.py --force-fp32

--force-fp16

boolean

Force FP16 precision globally.

python main.py --force-fp16

UNet/Diffusion Model Precision

--fp32-unet

boolean

Run the diffusion model in FP32 precision.

--fp16-unet

boolean

Run the diffusion model in FP16 precision.

--bf16-unet

boolean

Run the diffusion model in BF16 precision.

--fp64-unet

boolean

Run the diffusion model in FP64 precision (very slow, for debugging).

--fp8_e4m3fn-unet

boolean

Store UNet weights in FP8 E4M3FN format.

--fp8_e5m2-unet

boolean

Store UNet weights in FP8 E5M2 format.

--fp8_e8m0fnu-unet

boolean

Store UNet weights in FP8 E8M0FNU format.

VAE Precision

--fp16-vae

boolean

Run the VAE in FP16. May cause black images on some GPUs.

python main.py --fp16-vae

--fp32-vae

boolean

Run the VAE in full precision FP32. Use if you get black images.

python main.py --fp32-vae

--bf16-vae

boolean

Run the VAE in BF16 precision.

--cpu-vae

boolean

Run the VAE on the CPU instead of GPU. Slower but uses less VRAM.

python main.py --cpu-vae

Text Encoder Precision

--fp8_e4m3fn-text-enc

boolean

Store text encoder weights in FP8 (E4M3FN variant).

--fp8_e5m2-text-enc

boolean

Store text encoder weights in FP8 (E5M2 variant).

--fp16-text-enc

boolean

Store text encoder weights in FP16.

--fp32-text-enc

boolean

Store text encoder weights in FP32.

--bf16-text-enc

boolean

Store text encoder weights in BF16.

CUDA Configuration

--cuda-malloc

boolean

Enable cudaMallocAsync (enabled by default for PyTorch 2.0+).

--disable-cuda-malloc

boolean

Disable cudaMallocAsync. Use if you get CUDA errors.

python main.py --disable-cuda-malloc

--force-channels-last

boolean

Force channels-last memory format when inferencing models.

--force-non-blocking

boolean

Force non-blocking operations for all tensors. May improve performance on non-NVIDIA systems.

--disable-pinned-memory

boolean

Disable pinned memory usage.

Attention Mechanisms

--use-split-cross-attention

boolean

Use split cross attention optimization. Ignored when xformers is used.

--use-quad-cross-attention

boolean

Use sub-quadratic cross attention optimization. Ignored when xformers is used.

--use-pytorch-cross-attention

boolean

Use PyTorch 2.0+ native cross attention function.

python main.py --use-pytorch-cross-attention

--use-sage-attention

boolean

Use Sage Attention implementation.

--use-flash-attention

boolean

Use FlashAttention for faster attention computation.

--disable-xformers

boolean

Disable xformers optimization.

--force-upcast-attention

boolean

Force enable attention upcasting. Use if you get black images.

--dont-upcast-attention

boolean

Disable all attention upcasting. For debugging only.

Caching Strategies

--cache-classic

boolean

Use the old-style aggressive caching (default behavior).

python main.py --cache-classic

--cache-lru

integer

default:"0"

Use LRU caching with maximum N node results cached. May use more RAM/VRAM.

python main.py --cache-lru 10

--cache-none

boolean

Disable caching to reduce RAM/VRAM usage. Executes every node for each run.

python main.py --cache-none

--cache-ram

float

default:"4.0"

Use RAM pressure caching with specified headroom threshold (in GB). Removes large items when RAM is low.

# Use default 4GB threshold
python main.py --cache-ram

# Use 8GB threshold
python main.py --cache-ram 8.0

Preview Configuration

--preview-method

enum

default:"none"

Default preview method for sampler nodes. Options: none, auto, latent2rgb, taesd.

python main.py --preview-method auto

--preview-size

integer

default:"512"

Set the maximum preview size for sampler nodes.

python main.py --preview-size 768

File Loading Optimizations

--mmap-torch-files

boolean

Use memory mapping when loading .ckpt/.pt files.

python main.py --mmap-torch-files

--disable-mmap

boolean

Don’t use memory mapping when loading safetensors files.

--default-hashing-function

enum

default:"sha256"

Hash function for duplicate filename/contents comparison. Options: md5, sha1, sha256, sha512.

python main.py --default-hashing-function sha512

Performance Optimizations

--fast

string[]

Enable experimental optimizations that may reduce quality. Use without arguments to enable all, or specify individual features.Available optimizations:

fp16_accumulation - Use FP16 accumulation
fp8_matrix_mult - Use FP8 matrix multiplication
cublas_ops - Use cuBLAS operations
autotune - Enable autotuning

# Enable all optimizations
python main.py --fast

# Enable specific optimizations
python main.py --fast fp16_accumulation fp8_matrix_mult

These are experimental features that may crash ComfyUI or reduce image quality.

--deterministic

boolean

Make PyTorch use slower deterministic algorithms when possible.

python main.py --deterministic

This may not make images fully deterministic in all cases.

Custom Nodes

--disable-all-custom-nodes

boolean

Disable loading all custom nodes.

python main.py --disable-all-custom-nodes

--whitelist-custom-nodes

string[]

Specify custom node folders to load even when --disable-all-custom-nodes is enabled.

python main.py --disable-all-custom-nodes --whitelist-custom-nodes "ComfyUI-Manager" "custom-node-2"

--disable-api-nodes

boolean

Disable loading all API nodes. Also prevents the frontend from communicating with the internet.

python main.py --disable-api-nodes

ComfyUI Manager

--enable-manager

boolean

Enable the ComfyUI-Manager feature.

python main.py --enable-manager

--disable-manager-ui

boolean

Disable only the ComfyUI-Manager UI and endpoints. Background tasks still operate. Requires --enable-manager.

python main.py --enable-manager --disable-manager-ui

--enable-manager-legacy-ui

boolean

Enable the legacy UI of ComfyUI-Manager. Requires --enable-manager.

python main.py --enable-manager --enable-manager-legacy-ui

Frontend Configuration

--front-end-version

string

default:"comfyanonymous/ComfyUI@latest"

Specify the frontend version to use. Format: [owner]/[repo]@[version]

# Use latest frontend
python main.py --front-end-version Comfy-Org/ComfyUI_frontend@latest

# Use specific version
python main.py --front-end-version Comfy-Org/[email protected]

# Use legacy frontend
python main.py --front-end-version Comfy-Org/ComfyUI_legacy_frontend@latest

--front-end-root

string

Local filesystem path to frontend directory. Overrides --front-end-version.

python main.py --front-end-root /path/to/custom/frontend

Multi-User and Database

--multi-user

boolean

Enable per-user storage.

python main.py --multi-user

--database-url

string

default:"sqlite:///user/comfyui.db"

Specify the database URL.

# Use in-memory database
python main.py --database-url "sqlite:///:memory:"

# Use PostgreSQL
python main.py --database-url "postgresql://user:pass@localhost/comfyui"

--disable-assets-autoscan

boolean

Disable asset scanning on startup for database synchronization.

--comfy-api-base

string

default:"https://api.comfy.org"

Set the base URL for the ComfyUI API.

python main.py --comfy-api-base "https://custom-api.example.com"

Metadata and Output

--disable-metadata

boolean

Disable saving prompt metadata in output files.

python main.py --disable-metadata

Logging

--verbose

enum

default:"INFO"

Set the logging level. Options: DEBUG, INFO, WARNING, ERROR, CRITICAL.

python main.py --verbose DEBUG

--log-stdout

boolean

Send process output to stdout instead of stderr.

Development and Testing

--quick-test-for-ci

boolean

Quick test mode for continuous integration.

--windows-standalone-build

boolean

Enable features for Windows standalone build (auto-launch, etc.).

Example Configurations

Low VRAM Setup (4GB GPU)

python main.py --lowvram --fp16-vae --preview-method auto

High Performance (24GB+ GPU)

python main.py --highvram --gpu-only --preview-method taesd --fast

CPU Only

python main.py --cpu --preview-method auto

Remote Server

python main.py --listen 0.0.0.0 --port 8188 --preview-method auto

Development Mode

python main.py --verbose DEBUG --auto-launch --preview-method taesd

Shared Models with A1111

python main.py --extra-model-paths-config ./extra_model_paths.yaml --preview-method auto

Intel Arc GPU

python main.py --use-pytorch-cross-attention --preview-method auto

AMD ROCm

python main.py --use-pytorch-cross-attention --preview-method auto

Get Started

Core Concepts

Supported Models

Advanced Features

Configuration

Command Line Arguments

Running ComfyUI with Arguments

Server Configuration

Directory Configuration

GPU and Device Configuration

VRAM Management

Precision and Data Types

Global Precision

UNet/Diffusion Model Precision

VAE Precision

Text Encoder Precision

CUDA Configuration

Attention Mechanisms

Caching Strategies

Preview Configuration

File Loading Optimizations

Performance Optimizations

Custom Nodes

ComfyUI Manager

Frontend Configuration

Multi-User and Database

Metadata and Output

Logging

Development and Testing

Example Configurations

Low VRAM Setup (4GB GPU)

High Performance (24GB+ GPU)

CPU Only

Remote Server

Development Mode

Shared Models with A1111

Intel Arc GPU

AMD ROCm

Build docs developers (and LLMs) love

Get Started

Core Concepts

Supported Models

Advanced Features

Configuration

​Running ComfyUI with Arguments

​Server Configuration

​Directory Configuration

​GPU and Device Configuration

​VRAM Management

​Precision and Data Types

​Global Precision

​UNet/Diffusion Model Precision

​VAE Precision

​Text Encoder Precision

​CUDA Configuration

​Attention Mechanisms

​Caching Strategies

​Preview Configuration

​File Loading Optimizations

​Performance Optimizations

​Custom Nodes

​ComfyUI Manager

​Frontend Configuration

​Multi-User and Database

​Metadata and Output

​Logging

​Development and Testing

​Example Configurations

​Low VRAM Setup (4GB GPU)

​High Performance (24GB+ GPU)

​CPU Only

​Remote Server

​Development Mode

​Shared Models with A1111

​Intel Arc GPU

​AMD ROCm

Build docs developers (and LLMs) love

Running ComfyUI with Arguments

Server Configuration

Directory Configuration

GPU and Device Configuration

VRAM Management

Precision and Data Types

Global Precision

UNet/Diffusion Model Precision

VAE Precision

Text Encoder Precision

CUDA Configuration

Attention Mechanisms

Caching Strategies

Preview Configuration

File Loading Optimizations

Performance Optimizations

Custom Nodes

ComfyUI Manager

Frontend Configuration

Multi-User and Database

Metadata and Output

Logging

Development and Testing

Example Configurations

Low VRAM Setup (4GB GPU)

High Performance (24GB+ GPU)

CPU Only

Remote Server

Development Mode

Shared Models with A1111

Intel Arc GPU

AMD ROCm