ComfyUI 1-Click Installer With Sage Attention on Windows: Install, Launch and Update
By Furkan Gözükara · Updated Nov 21, 2026
The SECourses ComfyUI 1-click installer sets up ComfyUI on Windows in its own Python virtual environment with Sage Attention, Triton, xFormers and Flash Attention already installed, on Torch 2.13 and CUDA 13. Its launcher starts ComfyUI with Sage Attention and the Triton backend switched on, so you get the speed-up without compiling anything or editing files. The installer is available to members on Patreon; this guide covers what it installs, the install order, the launch flags and safe updates.
What the installer sets up
According to the ComfyUI auto installer post and the backend notes in the SwarmUI post, one run gives you:
- ComfyUI with Torch 2.13 and CUDA 13, including RTX 5000 series support.
- Speed libraries: Sage Attention, Triton, xFormers, Flash Attention, MSLK and TorchAO. The SwarmUI post notes that xFormers won't work properly without MSLK.
- Extras: FaceID, IP-Adapter, InsightFace, Reactor, DeepSpeed, Segment Anything, SAM2 and SAM3.
- Every custom node the presets use. Since V152,
Windows_Install_Or_Update_ComfyUI.batinstalls them all: no bundle menu and no separate node installer. Older tutorials show a custom node bundles installer with bundle 100; current versions do it in the main bat file. - Ready presets for image, video, lip sync, audio and music models. V144 had 106 ComfyUI presets plus 3 API presets, covering models such as FLUX 2, FLUX 3 Klein, Qwen Image, Wan 2.1, Wan 2.2, LTX 2.3, MiniMax H3 and Z-Image. V154 (18 November 2026) adds a 3-step LongCat-Video-Avatar 1.6 talking video preset.
- Other platforms: installers for RunPod, Massed Compute, SimplePod and Linux come in the same zip.
As the original announcement describes, ComfyUI goes into an isolated Python virtual environment (venv).
Before you install: Windows requirements
The installer post lists the Windows requirements as a one-time step: Python, CUDA and the C++ tools must be in place before ComfyUI. It points to the 1-click requirements installer and its one-time setup tutorial; the AI requirements installer guide explains what it does. If you prefer to fix a setup by hand, use the CUDA and C++ build errors guide.
Install ComfyUI on Windows
- Download the latest zip (ComfyUI_V154.zip at the time of writing) from the ComfyUI installer post and extract it into a folder.
- Run
Windows_Install_Or_Update_ComfyUI.bat. It installs ComfyUI, Python, Torch with CUDA, the speed libraries and all custom nodes. Since V151 this single bat file installs and updates everything. - Download models with the SwarmUI model downloader. The installer post recommends it for the core bundles so presets find their files. Point it at your ComfyUI models folder and pick the ComfyUI folder structure; the SwarmUI install guide covers the downloader in detail.
- Start ComfyUI with a launcher bat file, open a preset from the Workflows sidebar, add your prompt or files and press Run.
The ComfyUI beginner tutorial and its video follow exactly this route, from the download post to a first FLUX 3 Klein image, and then cover edits, layouts, batches, LoRAs and shared model folders.
Sage Attention: on by default, off when a model needs it
The launcher bat files start ComfyUI with --use-sage-attention and --enable-triton-backend, so Sage Attention is active without any edits. The NVFP4 with CUDA 13 tutorial post explains that the Flash Attention, Sage Attention and xFormers builds were compiled for Windows and Linux with all CUDA archs, covering GPUs from the GTX 1650 series to the RTX 5000 series, and that the compiled Sage Attention works with Qwen Image and Wan 2.2 without black images or videos.
Sage Attention can still hurt some models:
- The SwarmUI post warns that Sage Attention may rarely reduce quality or cause black output on some models, and names
--use-pytorch-cross-attentionas the highest-quality attention. - The MOSS-Image presets set attention to PyTorch themselves: in an RTX 3090 test this prevented black images with the BF16 files when ComfyUI starts with Sage Attention.
If one workflow gives black or broken images, switch that workflow or the launcher to PyTorch attention and compare the same seed.
Launch flags for VRAM and stability
The installer post lists these flags for the launcher bat files:
| Flag | When to use it |
|---|---|
--reserve-vram 3 | Your GPU starts using shared VRAM; keeps 3 GB free for other tasks |
--gpu-only | High VRAM GPU: keeps everything on the GPU |
--highvram | High VRAM GPU: keeps models on the GPU |
--cache-none | You use dual models or keep switching models |
--novram | Out-of-memory errors on long videos such as LTX 2; works with --cache-none |
--disable-smart-memory and --disable-comfy-compiler | ComfyUI gets stuck or runs out of memory |
--lowvram | Saves VRAM, but has no effect while dynamic VRAM is on, the default on NVIDIA GPUs |
Edit one argument at a time. The beginner tutorial shows adding --reserve-vram 2 to the default bat file and explains that a saved bat file only takes effect after a server restart. Windows_Run_VRAM_Optimized.bat offers a memory menu instead: since V143, normal VRAM mode starts ComfyUI with its default memory management (ComfyUI 0.22.0 removed --normalvram), and pressing Enter in the cache menu keeps ComfyUI's RAM-pressure caching.
For speed beyond attention, the beginner tutorial measured HQ INT8 ConvRot models at up to 2.3x faster on an RTX 3090 and 1.82 to 2.18x on an RTX 5090, for warm transformer steps only.
Update without reinstalling
Each update post repeats the same four steps:
- Download the latest zip from the post and extract it over your installer folder, overwriting the older files.
- Close ComfyUI, and SwarmUI too if it uses this ComfyUI, then run
Windows_Install_Or_Update_ComfyUI.bat. - Download any new core bundle a preset needs with the model downloader.
- Restart ComfyUI and refresh the browser tab.
The bat file ignores your personal pip settings since V143, so a custom package index set for another project no longer stops it. Since V152, a moved, renamed or copied ComfyUI folder gets a working environment again. The update video shows the full order when you also use SwarmUI: ComfyUI first, then models, SwarmUI and presets.
Cloud GPUs and Linux
The same zip installs on rented GPUs. Since V151, RunPod and SimplePod users follow RunPod_SimplePod_ComfyUI_Instructions_READ.txt, and Massed Compute and local Linux PCs follow Massed_Compute_Instructions_READ.txt. Two public tutorials show the cloud route: the Massed Compute and RunPod tutorial, with multi-GPU SwarmUI backends, and the RunPod tutorial on an RTX PRO 6000, which also connects ComfyUI to SwarmUI. To pick a provider, see RunPod vs Massed Compute vs SimplePod.
FAQ
Does the ComfyUI installer support RTX 50 series GPUs?
Yes. The installer post lists RTX 5000 series support with Torch 2.13 and CUDA 13, and the attention libraries were compiled with all CUDA archs from the GTX 1650 series up to the RTX 5000 series.
Why do I get black images with Sage Attention?
Sage Attention can rarely cause black output or lower quality on some models. Use PyTorch attention for that workflow; --use-pytorch-cross-attention is named as the highest-quality option. Some presets, such as MOSS-Image, already switch to PyTorch attention.
Do I need to install CUDA and Python myself first?
You need the Windows requirements once per PC. The 1-click requirements installer covers them, and the ComfyUI post links its tutorial as the one-time step.
Do I have to reinstall ComfyUI for every new version?
No. Extract the new zip over the old folder, run Windows_Install_Or_Update_ComfyUI.bat and restart ComfyUI. The beginner tutorial ends with exactly this update, done while preserving your work.
Can I use this ComfyUI as the SwarmUI backend?
Yes. The SwarmUI post calls this ComfyUI installer the mandatory backend and warns that some of the newest features may fail without it. The SwarmUI install guide shows how to connect it.