Startup Flags

ComfyUI accepts command-line arguments when started with python main.py. This page documents every flag defined in comfy/cli_args.py.

Windows Portable users can add flags to the .bat launch files (for example, run_nvidia_gpu.bat). See the Windows Portable guide for details.

Run python main.py --help in your ComfyUI directory for the built-in help text. Combine multiple flags as needed:

python main.py --listen 0.0.0.0 --port 8288 --disable-auto-launch --lowvram

Network & Server

Flag	Default	Description
`--listen` `[IP]`	`127.0.0.1`	IP address to listen on. Comma-separated list supported (e.g. `127.2.2.2,127.3.3.3`). If provided without a value, defaults to `0.0.0.0,::` (all IPv4 and IPv6 interfaces).
`--port`	`8188`	Port to listen on.
`--tls-keyfile` `PATH`	—	Path to TLS (SSL) key file. Enables HTTPS. Requires `--tls-certfile`.
`--tls-certfile` `PATH`	—	Path to TLS (SSL) certificate file. Enables HTTPS. Requires `--tls-keyfile`.
`--enable-cors-header` `[ORIGIN]`	disabled	Enable CORS. Optional origin, or `*` for all origins when no value is given.
`--max-upload-size`	`100`	Maximum upload size in MB.
`--enable-compress-response-body`	disabled	Enable compressing the HTTP response body.

# Listen on all interfaces (LAN access)
python main.py --listen

# Listen on a specific IP
python main.py --listen 0.0.0.0

# Custom port with HTTPS
python main.py --port 8443 --tls-keyfile key.pem --tls-certfile cert.pem

Directories

Flag	Default	Description
`--base-directory` `PATH`	ComfyUI root	Base directory for models, custom_nodes, input, output, temp, and user directories.
`--models-directory` `PATH`	ComfyUI root/models	Models directory. Overrides the models folder in `--base-directory`.
`--extra-model-paths-config` `PATH`	—	Load one or more `extra_model_paths.yaml` files. Can be specified multiple times.
`--output-directory` `PATH`	—	Output directory. Overrides `--base-directory`.
`--temp-directory` `PATH`	—	Temp directory. Overrides `--base-directory`.
`--input-directory` `PATH`	—	Input directory. Overrides `--base-directory`.
`--user-directory` `PATH`	—	User directory (absolute path). Overrides `--base-directory`. Path must exist and be readable.

Launch & Browser

Flag	Default	Description
`--auto-launch`	disabled	Automatically open ComfyUI in the default browser on startup.
`--disable-auto-launch`	disabled	Disable auto-launching the browser.
`--windows-standalone-build`	disabled	Windows portable build convenience mode. Enables auto-launch on startup (equivalent to `--auto-launch`).

--windows-standalone-build sets auto_launch to true. --disable-auto-launch overrides it. To run as a server without opening a browser, use --disable-auto-launch.

# Run server without opening browser
python main.py --disable-auto-launch

Devices & CUDA

Flag	Default	Description
`--cuda-device` `DEVICE_ID`	—	CUDA device IDs to use, comma-separated (e.g. `0` or `0,1`), or `all` to keep all currently visible devices. Other devices are hidden.
`--default-device` `ID`	—	Default device ID. All other devices remain visible.
`--cuda-malloc`	auto (torch 2.0+)	Enable cudaMallocAsync. Mutually exclusive with `--disable-cuda-malloc`.
`--disable-cuda-malloc`	—	Disable cudaMallocAsync. Mutually exclusive with `--cuda-malloc`.
`--directml` `[DEVICE]`	—	Use torch-directml. Optional device index; defaults to `-1` when no value is given.
`--oneapi-device-selector` `STRING`	—	oneAPI device selector string for Intel devices.

On Windows, ComfyUI uses GPU 0 only by default. To restore all GPUs, pass --cuda-device all.

Precision & Inference

Flags in the Global, UNET, VAE, and Text Encoder groups below are mutually exclusive within each group. Only one flag per group can be used at a time.

Global floating point

Flag	Description
`--force-fp32`	Force fp32 globally. Report if this improves GPU performance.
`--force-fp16`	Force fp16 globally. Also sets `--fp16-unet`.

UNET precision

Flag	Description
`--fp32-unet`	Run the diffusion model in fp32.
`--fp64-unet`	Run the diffusion model in fp64.
`--bf16-unet`	Run the diffusion model in bf16.
`--fp16-unet`	Run the diffusion model in fp16.
`--fp8_e4m3fn-unet`	Store UNet weights in fp8 (e4m3fn).
`--fp8_e5m2-unet`	Store UNet weights in fp8 (e5m2).
`--fp8_e8m0fnu-unet`	Store UNet weights in fp8 (e8m0fnu).

Each FP8 flag suffix names an 8-bit float layout from the industry FP8 spec (ONNX / OCP):

e4m3fn: 4 exponent bits, 3 mantissa bits. Good balance of precision and range for model weights on NVIDIA Ada Lovelace and newer GPUs.
e5m2: 5 exponent bits, 2 mantissa bits. Wider dynamic range but lower precision than e4m3fn. Useful when weight magnitudes vary widely.
e8m0fnu: 8 exponent bits, 0 mantissa bits. Maximum dynamic range with no fractional precision. Typically used as a block scaling factor (powers of two only).

VAE precision

Flag	Description
`--fp16-vae`	Run the VAE in fp16. May cause black images.
`--fp32-vae`	Run the VAE in full precision fp32.
`--bf16-vae`	Run the VAE in bf16.
`--cpu-vae`	Run the VAE on the CPU (not mutually exclusive with VAE precision flags).

Text encoder precision

Flag	Description
`--fp8_e4m3fn-text-enc`	Store text encoder weights in fp8 (e4m3fn).
`--fp8_e5m2-text-enc`	Store text encoder weights in fp8 (e5m2).
`--fp16-text-enc`	Store text encoder weights in fp16.
`--fp32-text-enc`	Store text encoder weights in fp32.
`--bf16-text-enc`	Store text encoder weights in bf16.

See the UNET precision section for the explanation of each FP8 format suffix.

Other inference options

Flag	Default	Description
`--fp16-intermediates`	disabled	Experimental: use fp16 for intermediate tensors between nodes instead of fp32.
`--force-channels-last`	disabled	Force channels-last memory format during inference.
`--supports-fp8-compute`	disabled	Act as if the device supports fp8 compute.
`--enable-triton-backend`	disabled	Enable Triton backend in comfy-kitchen. Disabled by default at launch.
`--disable-triton-backend`	disabled	Force-disable the comfy-kitchen Triton backend, overriding `--enable-triton-backend`.

Preview

Flag	Default	Description
`--preview-method`	`none`	Preview method for sampler nodes. Choices: `none`, `auto`, `latent2rgb`, `taesd`.
`--preview-size`	`512`	Maximum preview image size in pixels.

Cache

Cache mode flags are mutually exclusive. Only one of --cache-ram, --cache-classic, --high-ram, --cache-lru, or --cache-none should be used.

Flag	Default	Description
`--cache-ram` `[GB] [GB]`	enabled (default mode)	RAM pressure caching. First value: active-cache threshold in GB. Optional second value: inactive-cache/pin threshold in GB. When no values given: active = 10% of system RAM (min 2 GB, max 10 GB); inactive = 100% of system RAM (max 128 GB). Accepts at most two values.
`--cache-classic`	—	Use the old aggressive caching style.
`--high-ram`	disabled	Can improve performance slightly on systems with plenty of RAM or where pagefile use is preferred over reloading models. Also enables classic caching.
`--cache-lru` `N`	`0` (disabled)	LRU caching with a maximum of N node results cached. No hard upper limit; bounded by available RAM/VRAM. May use more RAM/VRAM.
`--cache-none`	—	Reduced RAM/VRAM usage; re-executes every node on each run.

ComfyUI reuses cached node results between runs in the same session, which is what makes repeat runs faster. Use --cache-none when you would rather every run start from a clean cache, for example while debugging a workflow or ruling out cached reuse between runs. Every node re-executes, so generations take longer. It is a startup flag, so restart ComfyUI for it to take effect.

Attention

Cross-attention method flags are mutually exclusive. Split and quad attention are ignored when xformers is used.

Flag	Description
`--use-split-cross-attention`	Use split cross attention optimization.
`--use-quad-cross-attention`	Use sub-quadratic cross attention optimization.
`--use-pytorch-cross-attention`	Use PyTorch 2.0 cross attention.
`--use-sage-attention`	Use Sage attention.
`--use-flash-attention`	Use FlashAttention.
`--use-ck-attention`	Use Comfy Kitchen attention. Requires hardware support for its INT8 kernels.
`--disable-xformers`	Disable xformers.
`--force-upcast-attention`	Force attention upcasting. Report if this fixes black images. Mutually exclusive with `--dont-upcast-attention`.
`--dont-upcast-attention`	Disable all attention upcasting. For debugging only.

VRAM & Memory

Async weight offloading is on by default on Nvidia and AMD GPUs, and dynamic VRAM is enabled automatically where it is supported. The flags below override those defaults.

VRAM mode flags (--gpu-only, --highvram, --lowvram, --novram, --cpu) are mutually exclusive.

Flag	Default	Description
`--gpu-only`	—	Store and run everything on the GPU (text encoders, CLIP, etc.).
`--highvram`	—	Keep models in GPU memory instead of unloading to CPU after use.
`--lowvram`	—	No effect when dynamic VRAM is enabled. Otherwise, runs text encoders on CPU.
`--novram`	—	Minimal VRAM usage when `--lowvram` is not enough.
`--cpu`	—	Use CPU for everything (slow).
`--reserve-vram` `GB`	OS-dependent	VRAM in GB to reserve for the OS and other software.
`--vram-headroom` `GB`	`0`	Extra VRAM in GB for dynamic VRAM to keep free above the default headroom. ComfyUI tries to keep this much VRAM completely free and unused, even counting VRAM used by other apps.
`--disable-nvml-pressure`	disabled	Use CUDA instead of NVML to report dynamic VRAM memory pressure.
`--async-offload` `[NUM_STREAMS]`	enabled on Nvidia and AMD	Async weight offloading. Optional stream count (default: 2).
`--disable-async-offload`	—	Disable async weight offloading.
`--disable-dynamic-vram`	—	Disable dynamic VRAM; use estimate-based model loading.
`--enable-dynamic-vram`	auto on Nvidia, ROCm 7.14+	Enable dynamic VRAM on systems where it is not enabled by default.
`--fast-disk`	disabled	Prefer disk-backed dynamic loading over unpinned RAM. Useful with fast NVMe.
`--disable-fast-disk`	disabled	Disable disk-backed dynamic loading and offload, so unpinned RAM is used instead. Overrides `--fast-disk`.
`--disable-cuda-graphs`	disabled	Disable CUDA graphs.
`--disable-comfy-compiler`	disabled	Disable the Comfy model compiler, including its CUDA graph subfeature.
`--assert-graph-breaks`	disabled	Fail on Comfy model compiler graph breaks.
`--force-non-blocking`	disabled	Force non-blocking tensor operations. May help on non-Nvidia systems; can break some workflows.
`--disable-smart-memory`	disabled	Aggressively offload to RAM instead of keeping models in VRAM.
`--disable-pinned-memory`	disabled	Disable pinned memory use.
`--mmap-torch-files`	disabled	Use mmap when loading ckpt/pt files.
`--disable-mmap`	disabled	Do not use mmap when loading safetensors.

Performance & Debugging

Flag	Default	Description
`--fast` `[OPT...]`	disabled	Enable experimental optimizations that may affect quality or stability. `--fast` alone enables all. Specific options: `fp16_accumulation`, `fp8_matrix_mult`, `cublas_ops`, `autotune`.
`--deterministic`	disabled	Use slower deterministic PyTorch algorithms where possible. Does not guarantee identical images in all cases.
`--default-hashing-function`	`sha256`	Hash function for duplicate filename/content comparison. Choices: `md5`, `sha1`, `sha256`, `sha512`.
`--debug-hang`	disabled	Enable stack trace dumps on Ctrl-C for debugging hangs.

# Enable all fast optimizations (experimental)
python main.py --fast

# Enable specific optimizations only
python main.py --fast fp16_accumulation cublas_ops

ComfyUI Manager

See ComfyUI-Manager Installation for setup instructions.

Flag	Description
`--enable-manager`	Enable ComfyUI-Manager.
`--disable-manager-ui`	Disable Manager UI and endpoints only. Background tasks (scheduled installs, etc.) continue. Requires `--enable-manager`.
`--enable-manager-legacy-ui`	Use the legacy Manager UI. Implies `--enable-manager`. Mutually exclusive with `--disable-manager-ui`.

Custom Nodes & API Nodes

Flag	Default	Description
`--disable-all-custom-nodes`	disabled	Disable loading all custom nodes.
`--whitelist-custom-nodes` `FOLDER...`	—	Custom node folders to load even when `--disable-all-custom-nodes` is set.
`--disable-partner-nodes`	disabled	Disable loading Partner Nodes (the paid API nodes). The frontend keeps its internet access.
`--offline`	disabled	Prevent the frontend from communicating with the internet by adding a restrictive Content-Security-Policy. Also disables Partner Nodes.
`--disable-api-nodes`	disabled	Deprecated alias for `--offline`. Use `--offline`, or `--disable-partner-nodes` to only disable Partner Nodes.
`--disable-metadata`	disabled	Disable saving prompt metadata in output files.

--offline also changes what the browser is allowed to load. The flag adds a Content-Security-Policy that keeps the page on resources served by ComfyUI itself, so images, fonts, media, frames, and network requests from other origins are blocked, and Partner Nodes (API nodes) are not registered at all. If a preview or a thumbnail that normally loads goes missing after a restart, check the launch arguments for this flag before troubleshooting the node. --disable-partner-nodes only stops Partner Nodes from loading and leaves the rest of the frontend’s internet access intact.

# Troubleshoot custom node issues
python main.py --disable-all-custom-nodes

# Allow only specific custom nodes
python main.py --disable-all-custom-nodes --whitelist-custom-nodes ComfyUI-Manager

Frontend & API

Flag	Default	Description
`--front-end-version`	`comfyanonymous/ComfyUI@latest`	Frontend version in `[owner]/[repo]@[version]` format. Requires internet to download from GitHub releases. Version can be `latest` or a semver (e.g. `1.0.0`).
`--front-end-root` `PATH`	—	Local filesystem path to the frontend directory. Overrides `--front-end-version`.
`--comfy-api-base`	`https://api.comfy.org`	Base URL for the ComfyUI API.
`--database-url`	`sqlite:///<ComfyUI>/user/comfyui.db`	Database URL. Use `sqlite:///:memory:` for in-memory.
`--enable-assets`	disabled	Enable the assets system (API routes, database sync, background scanning).
`--enable-asset-hashing`	disabled	Compute blake3 content hashes when scanning assets. Enables future asset portability features such as deduplication and cross-machine model resolution, at the cost of startup time and per-output hashing on large model directories. Off by default.
`--feature-flag` `KEY[=VALUE]`	—	Set a server feature flag. Bare `KEY` sets true. Can be repeated. Booleans and numbers are auto-converted.
`--list-feature-flags`	—	Print known CLI-settable feature flags as JSON and exit.

# List available feature flags
python main.py --list-feature-flags

# Set feature flags
python main.py --feature-flag show_signin_button=true

Logging & Misc

Flag	Default	Description
`--verbose` `[LEVEL] [FILE]`	`INFO`	Logging level. Choices: `DEBUG`, `DETAIL`, `INFO`, `WARNING`, `ERROR`, `CRITICAL`. `--verbose` alone sets `DEBUG`. Add a level followed by a file path to write that level to a file as well; the flag can be repeated to configure the console and file outputs together.
`--log-stdout`	disabled	Send normal process output to stdout instead of stderr.
`--dont-print-server`	disabled	Do not print server output to the console.
`--multi-user`	disabled	Enable per-user storage.
`--quick-test-for-ci`	disabled	Quick startup test for CI. Exits immediately after initialization.

This reference is based on ComfyUI comfy/cli_args.py. When upgrading ComfyUI, run python main.py --help or compare your local cli_args.py against this page to check for new or changed flags.

Develop with Comfy

Deploy ComfyUI

Comfy Router

Comfy CLI

Specifications

Network & Server

Directories

Launch & Browser

Devices & CUDA

Precision & Inference

Global floating point

UNET precision

VAE precision

Text encoder precision

Other inference options

Preview

Cache

Attention

VRAM & Memory

Performance & Debugging

ComfyUI Manager

Custom Nodes & API Nodes

Frontend & API

Logging & Misc

​Network & Server

​Directories

​Launch & Browser

​Devices & CUDA

​Precision & Inference

​Global floating point

​UNET precision

​VAE precision

​Text encoder precision

​Other inference options

​Preview

​Cache

​Attention

​VRAM & Memory

​Performance & Debugging

​ComfyUI Manager

​Custom Nodes & API Nodes

​Frontend & API

​Logging & Misc

Network & Server

Directories

Launch & Browser

Devices & CUDA

Precision & Inference

Global floating point

UNET precision

VAE precision

Text encoder precision

Other inference options

Preview

Cache

Attention

VRAM & Memory

Performance & Debugging

ComfyUI Manager

Custom Nodes & API Nodes

Frontend & API

Logging & Misc