ComfyUI has more than 100 command line arguments, and most guides still list flags that have changed meaning or no longer do anything. This cheat sheet covers every launch flag that matters in the current ComfyUI, grouped by what it does: memory and VRAM, precision, attention, caching, network access, folders, previews, custom nodes and debugging. It also shows how to add flags on the Windows portable build, a manual install and the Desktop app, and gives copy-paste presets for 6 to 8 GB, 12 GB, 16 GB and 24 GB GPUs, Macs, AMD cards and remote access.
- Where to add flags: at the end of the
python main.pyline, or inrun_nvidia_gpu.baton the Windows portable build. - Modern NVIDIA cards need few flags. ComfyUI now manages VRAM on its own (“Dynamic VRAM”), so
--lowvramdoes nothing on most NVIDIA setups. - Most useful flags:
--use-sage-attention(speed),--reserve-vram 1.5(stop OOM when other apps use VRAM),--listen(use from another device),--preview-method auto(live previews). - Out of memory: try
--reserve-vram, then--cache-none, then--novram. Smaller model files (fp8 or GGUF) help more than any flag. - A custom node broke ComfyUI? Start with
--disable-all-custom-nodes.
How to Add Launch Flags
Windows portable build
Open the ComfyUI portable folder, right-click run_nvidia_gpu.bat and choose Edit (or open it in Notepad). The first line looks like this:
.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
Add your flags to the end of that line, separated by spaces, then save and run the file:
.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --use-sage-attention --reserve-vram 1.5
Tip: copy the file to something like run_custom.bat first, so an update does not overwrite your changes.
Manual install (Windows, Linux, macOS)
Activate your virtual environment, go to the ComfyUI folder and add the flags after main.py:
cd ComfyUI python main.py --listen --port 8188 --preview-method auto
ComfyUI Desktop app
The Desktop app does not use a .bat file. Open Settings and look for the Server-Config section, which exposes the common launch options (listen address, port, VRAM mode, previews, attention and more) as switches. Restart the app after changing them.
Check that a flag worked
The console prints the active settings at startup, for example the VRAM state, the attention method (“Using sage attention”) and the device name. If a flag is misspelled, ComfyUI stops with “unrecognized arguments”, which tells you exactly which one is wrong. To see the full list for your version, run:
python main.py --help
Copy-Paste Presets
Start with the preset closest to your setup and add or remove flags from there. These assume a current ComfyUI on PyTorch 2.8 or newer.
| Situation | Flags | Why |
|---|---|---|
| NVIDIA 16 to 24 GB (RTX 4080, 5080, 4090, 5090) | --use-sage-attention --preview-method auto |
Faster attention, live previews. Memory is handled automatically. |
| NVIDIA 8 to 12 GB | --use-sage-attention --reserve-vram 1 --preview-method latent2rgb |
Keeps 1 GB free for Windows and the browser; cheap previews. |
| NVIDIA 6 GB or less | --reserve-vram 0.8 --cache-none --preview-method none |
Frees memory between runs; add --novram only if you still run out. |
| Gaming or streaming while generating | --reserve-vram 3 |
Leaves room for the game or OBS. |
| Older ComfyUI or Dynamic VRAM disabled, low VRAM | --lowvram |
Runs text encoders on the CPU to save VRAM. |
| Use from phone, laptop or another PC on your network | --listen 0.0.0.0 --port 8188 |
Opens ComfyUI on your local network. Do not expose it to the internet without a login. |
| Apple Silicon Mac | --preview-method auto (and --force-fp16 if you see errors on older macOS) |
Uses the MPS backend automatically. |
| Troubleshooting a broken install | --disable-all-custom-nodes --verbose DEBUG |
Starts clean and prints detailed logs. |
| Fully offline / private | --offline --disable-metadata |
No outside requests from the frontend; no workflow saved inside output images. |
Memory and VRAM Flags
Only one of the first five flags can be used at a time.
| Flag | What it does | When to use it |
|---|---|---|
--gpu-only |
Stores and runs everything, including text encoders, on the GPU | Big cards (24 GB+) that can hold the whole workflow |
--highvram |
Keeps models in VRAM after use instead of moving them to system RAM | Large cards, to skip reload time between runs |
--lowvram |
No effect when Dynamic VRAM is active. Otherwise runs the text encoders on the CPU | Older installs, AMD or other setups without Dynamic VRAM |
--novram |
Most aggressive offloading, for when lowvram is not enough | Very small GPUs; much slower |
--cpu |
Runs everything on the CPU | No usable GPU; very slow |
--reserve-vram 1.5 |
Keeps this many GB of VRAM free for your OS and other apps | Out-of-memory errors when a browser, game or second app uses the GPU |
--vram-headroom 1 |
Extra free VRAM that Dynamic VRAM keeps unused, counting other apps | Dynamic VRAM setups that still hit OOM |
--disable-smart-memory |
Moves models to system RAM after each use instead of keeping them in VRAM when possible | Sharing the GPU with other AI apps |
--async-offload [N] / --disable-async-offload |
Streams weights between RAM and VRAM in the background (default 2 streams, on by default on NVIDIA) | Disable only if you suspect it causes crashes |
--fast-disk / --disable-fast-disk |
Loads and offloads from disk instead of unpinned RAM | Fast NVMe drive and limited system RAM |
--disable-pinned-memory |
Turns off pinned (page-locked) RAM | Low system RAM or driver problems |
--disable-mmap |
Reads safetensors files fully instead of memory-mapping them | Network drives or slow first loads on some systems |
--cpu-vae |
Runs the VAE on the CPU | OOM at the very end (during VAE decode); slow, try a tiled VAE decode node first |
What Dynamic VRAM Changed
Older guides tell everyone with 8 GB to add --lowvram. That advice is outdated. Current ComfyUI includes Dynamic VRAM, which loads only the parts of a model that are needed and streams the rest from system RAM as it runs. According to the ComfyUI source:
- It is on by default on NVIDIA GPUs, and on AMD with ROCm 7.14 or newer.
- It needs PyTorch 2.8 or later (2.12+ is recommended) and the
comfy-aimdopackage. If either is missing, ComfyUI prints a warning and falls back to the old loader. - It is turned off if you use
--highvram,--gpu-only,--novramor--cpu, or--disable-dynamic-vram. - While it is on,
--lowvramdoes nothing.
Look for the line “DynamicVRAM support detected and enabled” in the console at startup. If you see “Falling back to legacy ModelPatcher” instead, update PyTorch and run pip install -r requirements.txt in the ComfyUI folder.
For real numbers on what fits in 16 GB, see our ComfyUI VRAM test.
Precision Flags (fp8, fp16, bf16)
These force a data type. In most cases ComfyUI picks the right one from the model file, so only use them to fix a problem or save memory.
| Flag | Effect |
|---|---|
--force-fp32 / --force-fp16 |
Force all models to fp32 or fp16 |
--bf16-unet / --fp16-unet / --fp32-unet |
Run the diffusion model in bf16, fp16 or fp32 |
--fp8_e4m3fn-unet / --fp8_e5m2-unet / --fp8_e8m0fnu-unet |
Store diffusion model weights in fp8, roughly halving their VRAM use |
--fp16-vae / --bf16-vae / --fp32-vae |
Set VAE precision. fp16 VAE can cause black images on some models |
--fp8_e4m3fn-text-enc / --fp8_e5m2-text-enc / --fp16-text-enc / --bf16-text-enc / --fp32-text-enc |
Set text encoder precision |
--fp16-intermediates |
Experimental: fp16 for data passed between nodes |
--supports-fp8-compute |
Tell ComfyUI the GPU supports fp8 maths even if it is not detected |
--force-upcast-attention / --dont-upcast-attention |
Fix (or debug) black images caused by attention precision |
A simpler way to save memory is to download an fp8 or GGUF version of the model instead of forcing it at load time. Our GGUF quantization guide explains which file to pick, and the bf16 speed guide for RTX 30 and 40 covers precision and speed on older cards.
Attention Flags (Sage, Flash, xformers)
Attention is the slowest part of most image and video models, so this is where flags make the biggest speed difference.
| Flag | What it uses | Notes |
|---|---|---|
| (none) | PyTorch’s built-in attention (SDPA) on most setups | Safe default |
--use-sage-attention |
SageAttention | Often the fastest on NVIDIA, especially for video models. Install the sageattention package (and Triton) into ComfyUI’s Python first |
--use-flash-attention |
FlashAttention | Needs the flash-attn package built for your PyTorch and CUDA |
--use-pytorch-cross-attention |
PyTorch 2 SDPA explicitly | Useful to override other choices |
--use-ck-attention |
Comfy Kitchen attention | Newer built-in option |
--use-split-cross-attention / --use-quad-cross-attention |
Memory-saving attention | For very low VRAM or old GPUs; slower |
--disable-xformers |
Stops ComfyUI using xformers even if installed | Fixes crashes from an xformers build that does not match your PyTorch |
To install SageAttention on the Windows portable build, run pip with the embedded Python from the portable folder, for example .\python_embeded\python.exe -m pip install sageattention. Windows also needs a matching Triton build (the triton-windows package). If the package is missing, remove the flag or ComfyUI will fail when it tries to use it.
Estimate for a 16 GB card like the RTX 5080: SageAttention can cut video generation time by roughly 20 to 40 percent compared with the default attention, with smaller gains for short image jobs. This is an estimate, not a measured result.
Caching Flags
ComfyUI caches node results so that changing one setting does not rerun the whole workflow. These flags change how much it keeps.
| Flag | Behaviour |
|---|---|
--cache-ram [GB] [GB] |
The default. Caches results until system RAM gets tight. The first value is the active-cache headroom (default 10% of RAM, between 2 and 10 GB); the optional second value is for inactive caches |
--cache-classic |
The old, aggressive caching |
--cache-lru 10 |
Keeps the last N node results. Faster when you switch between workflows, uses more memory |
--cache-none |
Caches nothing. Lowest RAM and VRAM use, but every node reruns every time |
--high-ram |
Can help a little on machines with lots of RAM, or where using the pagefile is better than reloading models |
Network and Remote Access Flags
| Flag | What it does |
|---|---|
--listen |
With no value, listens on all addresses (IPv4 and IPv6) so other devices can connect. Default is 127.0.0.1 (this PC only) |
--listen 192.168.1.20 |
Listens on one address; separate several with commas |
--port 8190 |
Changes the port from 8188. Use a different port for each copy if you run two instances |
--tls-keyfile key.pem --tls-certfile cert.pem |
Serves ComfyUI over HTTPS. Both are needed |
--enable-cors-header [ORIGIN] |
Allows other web pages to call the API; with no value it allows all origins |
--max-upload-size 500 |
Upload limit in MB (default 100). Raise it for long videos |
--enable-compress-response-body |
Compresses API responses |
--multi-user |
Separate settings and workflows per user |
--auto-launch / --disable-auto-launch |
Open (or do not open) the browser on start |
ComfyUI has no password of its own. If you use --listen, keep it on your home network or behind a VPN or tunnel with authentication. Never forward port 8188 on your router.
Folder and Path Flags
| Flag | What it sets |
|---|---|
--base-directory PATH |
The base folder for models, custom_nodes, input, output, temp and user |
--models-directory PATH |
Just the models folder, for example a big second drive |
--extra-model-paths-config FILE |
Loads extra_model_paths.yaml files, to share models with Forge, A1111 or another ComfyUI |
--output-directory, --input-directory, --temp-directory, --user-directory, --previews-directory |
Each folder separately; these override the base directory |
--database-url |
Where ComfyUI keeps its database (default comfyui.db in the user folder) |
--front-end-version Comfy-Org/ComfyUI_frontend@latest |
Downloads and uses a specific frontend version (needs internet) |
--front-end-root PATH |
Uses a frontend from a local folder |
Not sure which subfolder a file belongs in? Our ComfyUI model folder cheat sheet lists every one.
Preview Flags
By default, current ComfyUI shows no live preview while sampling. --preview-method sets the default:
none: no preview (the default, fastest).latent2rgb: a quick, blurry colour preview with almost no cost.taesd: a sharper preview using small TAESD decoders. Put the decoder files (for exampletaesd_decoder.pth,taesdxl_decoder.pth,taef1_decoder.pth) inComfyUI/models/vae_approx/.auto: uses TAESD when the decoder is present, otherwise latent2rgb.
--preview-size 512 sets the maximum preview size. The preview mode can also be changed in the ComfyUI settings menu without restarting.
Custom Nodes, Manager and Offline Flags
| Flag | What it does |
|---|---|
--disable-all-custom-nodes |
Starts without loading anything from custom_nodes. The first step when ComfyUI will not start |
--whitelist-custom-nodes ComfyUI-GGUF |
With the flag above, still loads the listed folders |
--disabled-nodes-config FILE |
A YAML file listing node IDs to disable |
--enable-manager |
Turns on the built-in ComfyUI-Manager. Needs pip install -r manager_requirements.txt first |
--enable-manager-legacy-ui / --disable-manager-ui |
Use the old Manager interface, or hide the Manager interface while background tasks still run |
--offline |
Stops the frontend talking to the internet and disables the paid partner (API) nodes |
--disable-partner-nodes |
Only hides the paid API nodes |
--disable-metadata |
Does not save the workflow inside output PNGs (useful before sharing images) |
--disable-product-tips |
Hides Comfy-Org product tips in the console |
If a node pack broke your install, our ComfyUI errors and fixes guide walks through the usual causes.
The –fast Flag and Other Speed Options
--fast turns on experimental optimizations that may reduce quality or crash some workflows. With no value it enables all of them; you can also list the ones you want:
fp16_accumulation: faster fp16 matrix maths on supported NVIDIA cards, with a small accuracy cost.fp8_matrix_mult: uses fp8 maths on GPUs that support it (RTX 40 and 50 series) for fp8 models.cublas_ops: uses cuBLAS operations where available.autotune: lets PyTorch benchmark and pick faster kernels.
python main.py --fast fp16_accumulation fp8_matrix_mult
Related flags: --enable-triton-backend / --disable-triton-backend for the Comfy Kitchen Triton backend, --disable-cuda-graphs and --disable-comfy-compiler if the built-in model compiler causes trouble, --force-channels-last, and --deterministic for slower but more repeatable results.
Estimate for a 16 GB card like the RTX 5080: --fast fp8_matrix_mult with an fp8 model may speed up sampling by around 10 to 30 percent. This is an estimate, not a measured result. Compare images before and after, because quality changes differ by model.
GPU Selection, AMD, Intel and Mac
| Flag | Use |
|---|---|
--cuda-device 1 |
Use only the GPU with this ID (or a list like 0,1). Run one ComfyUI per GPU with different ports |
--default-device 1 |
Make this GPU the default while keeping others visible |
--cuda-malloc / --disable-cuda-malloc |
Turn the CUDA async allocator on or off (on by default). Disable it if you see allocator errors |
--directml [ID] |
Windows DirectML backend for AMD or Intel GPUs without ROCm |
--oneapi-device-selector |
Choose Intel GPUs through oneAPI |
--force-non-blocking |
May help performance on some non-NVIDIA systems |
On Apple Silicon, ComfyUI uses the MPS backend with no flag. Because Macs share memory between CPU and GPU, VRAM flags matter less; keep enough free RAM and close other heavy apps.
Logging and Debugging Flags
--verbose DEBUG: more detailed console logs. You can also add a log file, for example--verbose DEBUG comfy.log.--log-stdout: send output to stdout instead of stderr.--dont-print-server: hide server messages.--debug-hang: print stack traces when you press Ctrl+C, to find where ComfyUI is stuck.--default-hashing-function sha256: hash used to compare duplicate files (md5, sha1, sha256 or sha512).--feature-flag KEY=VALUEand--list-feature-flags: set or list server feature flags.
FAQ
Should I use –lowvram on an 8 GB NVIDIA card?
Usually no. On a current install with Dynamic VRAM, it has no effect. Use --reserve-vram 1 and smaller model files instead. It is still useful on older installs or AMD setups without Dynamic VRAM.
What is the difference between –lowvram and –novram?
--lowvram moves the text encoders to the CPU. --novram offloads much more and is for when lowvram is not enough. Both are slower than keeping models on the GPU.
How do I open ComfyUI from my phone?
Start it with --listen, find your PC’s local IP address (for example 192.168.1.20) and open http://192.168.1.20:8188 on the phone while on the same Wi-Fi.
How do I run two ComfyUI instances?
Give each one a different port, and on a multi-GPU PC a different --cuda-device, for example --port 8188 --cuda-device 0 and --port 8189 --cuda-device 1.
Why did ComfyUI say “unrecognized arguments”?
The flag is misspelled or does not exist in your version. Check with python main.py --help and update ComfyUI if it is missing.
Do flags work in the ComfyUI Desktop app?
Yes, through Settings, Server-Config, which sets the same options.
Final Thoughts
On a modern NVIDIA card, ComfyUI needs very few launch flags. The ones worth knowing are --use-sage-attention for speed, --reserve-vram to stop out-of-memory errors when other apps share the GPU, --listen for remote use, --preview-method for live previews and --disable-all-custom-nodes for troubleshooting. Old advice to add --lowvram everywhere no longer applies on current installs.
More ComfyUI reference guides:
- ComfyUI Model Folder Cheat Sheet
- ComfyUI Errors and Fixes
- ComfyUI Samplers and Schedulers Explained
- How Much VRAM Do You Need for ComfyUI?
- GGUF Quantization Explained
Sources
- ComfyUI source: comfy/cli_args.py (all flags and their help text)
- ComfyUI source: main.py (Dynamic VRAM requirements)
- GitHub: comfyanonymous/ComfyUI
