Configuration

Everything you can tune at runtime — CLI flags, environment variables, and the extra_model_paths.yaml file. The authoritative source is comfy/cli_args.py.

CLI flags by category

Network and TLS

Flag	Default	Purpose
--listen [IP[,IP]]	127.0.0.1	Listen address. Without arg → 0.0.0.0,::
--port N	8188	Listen port
--tls-keyfile, --tls-certfile	-	Enable HTTPS
--enable-cors-header [ORIGIN]	-	CORS. Without arg → *
--enable-compress-response-body	off	Gzip JSON/text responses
--max-upload-size MB	100	Max upload size in MB

Directories

Flag	Purpose
--base-directory PATH	Set the base for models/, custom_nodes/, input/, output/, temp/, user/
--output-directory PATH	Override output directory
--temp-directory PATH	Override temp directory
--input-directory PATH	Override input directory
--user-directory PATH	Override user directory
--extra-model-paths-config PATH...	Load one or more extra_model_paths.yaml files

Devices

Flag	Purpose
--cpu	Run everything on CPU
--cuda-device DEVICE_ID	Restrict to a single CUDA device
--default-device DEFAULT_DEVICE_ID	Default device while keeping all visible
--directml [DEVICE]	Use torch-directml on Windows for AMD/Intel
--oneapi-device-selector STR	oneAPI device selector for Intel XPU
--disable-ipex-optimize	Disable Intel IPEX optimizations
--cuda-malloc / --disable-cuda-malloc	Force enable / disable cudaMallocAsync

Memory

Flag	Purpose
--gpu-only	Keep everything (text encoders, CLIP, etc.) on GPU
--highvram	Don't unload to CPU after use
--normalvram	Force NORMAL_VRAM state
--lowvram	Split-load the unet
--novram	When --lowvram isn't enough
--reserve-vram GB	Reserve GPU memory for OS/other apps
--async-offload [N]	Async weight offload with N streams (default 2). Default on for NVIDIA
--disable-async-offload	Disable async offload
--enable-dynamic-vram / --disable-dynamic-vram	Force enable/disable DynamicVRAM
--disable-smart-memory	Aggressively offload to CPU between models

Dtypes

Flag	Purpose
--force-fp32 / --force-fp16	Coarse override
--fp32-unet / --fp16-unet / --bf16-unet / --fp64-unet	Diffusion model dtype
--fp8_e4m3fn-unet / --fp8_e5m2-unet / --fp8_e8m0fnu-unet	FP8 storage dtypes for the unet
--fp16-vae / --fp32-vae / --bf16-vae	VAE dtype
--cpu-vae	Run VAE on CPU
--fp16-text-enc / --fp32-text-enc / --bf16-text-enc / --fp8_e4m3fn-text-enc / --fp8_e5m2-text-enc	Text encoder dtype
--fp16-intermediates	Use fp16 for inter-node tensors
--force-channels-last	Channels-last memory layout
--supports-fp8-compute	Pretend the device supports FP8 compute

Attention

Flag	Purpose
--use-pytorch-cross-attention	torch SDPA
--use-flash-attention	FlashAttention
--use-sage-attention	Sage attention
--use-split-cross-attention	Chunked attention
--use-quad-cross-attention	Sub-quadratic
--disable-xformers	Don't use xformers even if installed
--force-upcast-attention / --dont-upcast-attention	Override attention upcasting heuristics

Caching

Flag	Purpose
--cache-classic	Aggressive cache (drop ASAP)
--cache-lru N	Bounded LRU
--cache-ram [GB]	RAM-pressure cache; default 25% of system RAM (4-32 GB)
--cache-none	No cache

Performance

Flag	Purpose
--fast [features...]	Enable experimental optimizations: fp16_accumulation, fp8_matrix_mult, cublas_ops, autotune. With no args, all are enabled.
--mmap-torch-files	mmap when loading ckpt/pt files
--disable-mmap	Don't mmap safetensors
--disable-pinned-memory	Disable pinned host memory pool
--deterministic	Use deterministic torch algorithms
--force-non-blocking	Use non-blocking tensor ops everywhere

Frontend / manager

Flag	Purpose
--front-end-version OWNER/REPO@VERSION	Resolve a specific frontend release. Default comfyanonymous/ComfyUI@latest (uses bundled)
--front-end-root PATH	Use a pre-extracted frontend bundle
--enable-manager	Enable ComfyUI-Manager
--disable-manager-ui	Hide the Manager UI but keep background features
--enable-manager-legacy-ui	Use the legacy Manager UI
--auto-launch / --disable-auto-launch	Open the browser at startup

Custom nodes

Flag	Purpose
--disable-all-custom-nodes	Skip loading custom nodes
--whitelist-custom-nodes NAME...	Re-enable specific custom nodes
--disable-api-nodes	Disable comfy_api_nodes/

Database / assets

Flag	Purpose
--database-url URL	SQLAlchemy URL. Default sqlite:///<repo>/user/comfyui.db. sqlite:///:memory: works
--enable-assets	Turn on the assets API + DB sync + scanner

Previews

Flag	Purpose
--preview-method none|auto|latent2rgb|taesd	Preview style during sampling
--preview-size N	Max preview size in pixels (default 512)

Misc

Flag	Purpose
--disable-metadata	Don't embed prompt metadata in saved files
--multi-user	Per-user storage
--default-hashing-function md5|sha1|sha256|sha512	Hash for duplicate detection. Default sha256
--comfy-api-base URL	API base for paid integrations. Default https://api.comfy.org
--verbose [LEVEL]	Logging level. Default INFO. With no arg → DEBUG
--log-stdout	Send logs to stdout instead of stderr
--dont-print-server	Suppress server output
--quick-test-for-ci	Exit immediately after init (CI smoke)
--windows-standalone-build	Behavior tweaks for the Windows portable build

Environment variables

main.py sets these before importing torch:

Variable	Set when
HF_HUB_DISABLE_TELEMETRY=1	Always
DO_NOT_TRACK=1	Always
MIMALLOC_PURGE_DELAY=0	On Windows
TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1	Always (ROCm AOT Triton experimental)
CUBLAS_WORKSPACE_CONFIG=:4096:8	When --deterministic and not already set
CUDA_VISIBLE_DEVICES, HIP_VISIBLE_DEVICES, ASCEND_RT_VISIBLE_DEVICES	When --cuda-device or --default-device is set
ONEAPI_DEVICE_SELECTOR	When --oneapi-device-selector is set
OCL_SET_SVM_SIZE=262144	On ROCm (per AMD's request)

User-set env vars also matter:

Variable	Effect
HSA_OVERRIDE_GFX_VERSION	AMD: spoof gfx version for unsupported cards
PYTORCH_TUNABLEOP_ENABLED=1	AMD: enable TunableOp (slow first run, faster after)

extra_model_paths.yaml

Copy extra_model_paths.yaml.example to extra_model_paths.yaml (or pass via --extra-model-paths-config) and edit. Each entry adds search paths for a folder type:

a1111:
  base_path: /path/to/stable-diffusion-webui/
  checkpoints: models/Stable-diffusion
  vae: models/VAE
  loras: |
    models/Lora
    models/LyCORIS

The runtime merges these into folder_paths.folder_names_and_paths at startup. See utils/extra_config.py for the loader.

comfy.options

comfy/options.py is a tiny module with one boolean: args_parsing. Setting it to True (which main.py does) lets comfy.cli_args parse sys.argv. Tests import comfy.options and leave args_parsing False to get default args.