ComfyUI 5 Installation methods from installation entry to mastery, ComfyUI

ComfyUI 5 Installation methods from installation entry to mastery, ComfyUI

It's never the "what's a good model" that's asked the most in the background, but the first step --"I don't know anythingComfyUI How does it work?""Where's the download address?" How do you make the first drawing after filling up the nodes

Seven or eight such messages were received a day. I used to write ComfyUI, which is straight from "How to choose the model" -- and now look back, it's like I'm talking to you about how to bend and drift before I teach you how to fire. I've never been serious about the first step of putting the software on, running, making the first picture。

It's a break. It's a patch today, and it's a patch. From this point on, I'll break the ComfyUI into a whole collection:"CommyUI from installation to mastery."I don't know. This is 01, I'm not messing with..I've tried all five ways to fit ComfyUI on the market, download the address, how do you point each step, who fits, what's the hole, and write it down. You're White. Just read this。

I'M USING 4060, 8G, AND THE PIT BASICALLY STEPPED ON IT. IT'S A LITTLE LONG, AND IT'S SUGGESTED TO COLLECT IT AND PUT IT IN ONE OF THE STEPS TO COME BACK。

One, two minutes to see if it's enough

I'VE SEEN TOO MANY PEOPLE RUSHING OVER SEVEN OR EIGHT GS, DECOMPRESSING A START-UP AND MAKING A BIG MISTAKE ALL NIGHT. THREE THINGS FIRST。

First: Is there enough (decisive) to show the card。

Your configuration
What can I do
No unique / certified
It's too slow to suspect life
4G (1650/3050)
I can't run, I can't do anything
6-8G (3060/4060/5060)
Comfortable area
SDXL/Flux/Animation
12G AND ABOVE
Whatever

I'M USING 4060, 8G, WHICH BELONGS TO THE "ENOUGH, OCCASIONALLY VISIBLE". THIS SERIESEvery piece has a low-profile memoryYOU CAN FOLLOW 8G, 6G。

NUMBER TWO: HARD DRIVE SOLID, 20G。 IT'S GOT TO BE SSD, IT'S STUCK TO THE COMPUTER. THE SOFTWARE PLUS A MODEL IS 20 G, AND I'M HEADING FOR 100 G NOW。

Third: Update the graphic card drive。 901 TP3T'S START-UP BLUNDER WAS ALL HERE, AND I FELL FIRST。

1. Go to the latest stabilization drive under the NVIDIA network (nvidia.cn/drivers), completeRestart;
2. Click Win+R to enter cmd and return to the car and click in the black box line:

nvidia-smi

IT'S OKAY TO SHOW GRAPHIC CARD AND CUDA VERSIONS; THE "NOT AN INTERNAL COMMAND" INDICATES THAT THE DRIVE WAS NOT INSTALLED AND RETURNED TO RELOAD。

Invisible pit:The installation path may not include Chinese, spaces, special symbols. Like D: \AI drawing \CommyUI has an eight-nine-thirty-nine to blame. Old practical, pure English, like D:\AI\CommyUI. For the first time, I put the desktop on, and the path was long and Chinese, and started a direct crash for more than an hour。

Two, five. First look at this table

From the "most stupid" to the "most flexible" order, don't be greedy and pick the right one to look down。

Fake
For who
shortcoming
1 Official portable version
Most rookies
We need to install the manager once
2 Official desktop
Imagine the software
Update slow
3 Autumn Leaf Integration Package
Pure white is afraid of orders
Big and sometimes bad
4 Manual Installation
Progress/Developer
Order to desist
Five cloud-free
No graphic card/low configuration
It costs money

The recommendations are addressed directly to you:THERE ARE N CARDS WHO WANT TO LEARN →1 HAND-HELD EDITIONS; A HEADACHE AT THE COMMAND LINE 3 AUTUMN INTEGRATION PACK; NO GRAPHIC CARD/NUCLEO 5 CLOUDS. NEXT ONE。

III. The official hand-held version of Mode 1 (first option of 6 steps)

I recommend this most. Take an independent Python, don't fight with the system, have a problem check, always the latest。

Step 1: Compress the bag under the official pageTwo addresses, two or one

• GitHub Post: Github.com/comfyanonymous/CommyUI/releases (see CommyUI_windows_portable_nvidia.7z)
• official chinese documentation:docs.comfy.org/zh/establishment/comfyui_portable_windows

Select the pit

Your card
Which one
20/30/40/50 System
Standard nvidia version (CUDA13.0+Py3.13)
10 is or is older
c126 Old version (CUDA12.6+Py3.12)

GitHub is slow and can search the domestic mirrors of "YanWenKun ComfyUI with a handbag" with the same content. About 7G。

Step 2: Represse the English path with 7-ZipLike D: \AI\CommyUI_Windows_portable. Do not desktop, do not C, do not Chinese. Post-pressure structure:

It's not like it's a bad idea
Ideas - ComfyUI
- python_embeded
- run_cpu.bat
└-run_nvidia_gpu.bat

step 3: double-click run_nvidia_gpu.batPop the black box off and wait for it to run. Seeing this business rises:

To see the GUI go to: http://127.0.0.1:8188

Click this address manually without opening the browser. The black box is "engine" and never turn it off。

Step 4: Install Manager Plugin Manager(Unfinished node it's a key, save 80% toss). Enter ComfyUI\custom_nodes folder, address bar loses cmd back, paste:

git clone https://github.com/ltdrdata/ComfyUI-Manager.git

This post is part of our special coverage Global Voices 2011. The new version of the handbag mostly carries its own Manager, and the interface has a button to skip this step。

Step 5: Release a modelI don't know. The next basic checkpoint (.safetensors, 2-6G, where to look at Part VIII) is available at the ComfyUI\models\checkpoints。

Step 6: FigureLook at part 9 of this paper, "Get your first picture."。

Mode 2 official desktop version (like software)

You don't want to touch a compressor, command line, you want to load up with an icon, double-click, and you use it。

Download address:web site: http://www.download.comfy.org/windows/nsis/x64. it's an exeONLY N CARDS SUPPORTED.

step:Double-click exe → Get Started → Select NVIDIA → Select an installation position (suspend 50G solid, no C disc) → Load a portable version to move, completely new load skip the auto-dependencyFIRST 5-15 MINUTES AROUND 15GWait, wait, wait, wait。

Advantages: The dumbest of them have icons with their own belt manager. Shortcomings: Beta status, half-literal update。

V. Mode 3 Autumn integration package (pure white saviour)

One of the greatest benefits of integration packages made by communities in the countryAll Chinese Graphics Launcher, update, load, cut the model with a full mouse pointer, without a single line of command。

Download address:Site B tutorial bilibili.com/video/BV1Ew411776J/, Quarque Webpapan.quark.cn/s/41302f448cec. 20G+, pre-packaged with a bunch of plugin models。

step:Depression to a plain English path double-click "A Paint Starter.exe" to update the kernel extension with a Chinese and model path Zoom Point to start "Auto Open Browser."。

Advantages: All Chinese starters, one-key upgrades of nodes, openings are the most useful. Weaknesses: big, sometimes different from official, and far less transparent when packed too much。

VI. Mode 4 manual installation (step forward, freshman jump)

It is the most flexible to understand command lines, to develop them or to customize them in depth. The first three are enough。

i'm not sure I'm going to be able to do that
i'm sorry
(PyTorch for CUDA when activated)
pip install -r requirements.txt
python main.py

THE CUDA VERSION SHOULD BE ALIGNED WITH THE DRIVER (THE EASIEST TO OVERTURN):

Driver Version
PyTorch CUDA
560+
c126
570+
c128
575+
cu130

RTX 50 is required to cu128 and above, low enough to run。

VII. Mode 5 cloud-free installation (no graphic card)

The computer is not unique, the notebook is visible, the old machine can't run -- don't be hard-on, rent a machine with a good graphic card to run online, and the browser is open。

RunningHub:runninghub.ai, registered to deliver points, 40,000+ models, 8,000+ nodes, environment fit. The shared type is about 7 yuan/hour and only RRTX4090 is about 1500 yuan/month。

Advantages: High-end cards are also available without local content. Shortcomings: Costing, data in the clouds. It's for people who experience it first or who can't run。

VIII. Where does the model go? Which folder

Loading the software is just an empty shell, and a model is needed to make the picture. This is also the most frequently asked in the background。

website
Features
Liblibai liblib.art
No magic in the country, no first choice
Gitcode ai.gitcode.com/models
Domestic mirrors, download fast
Civitai (station C)
The best. Magic
HuggingFace
magic.com

No magic。 _Other Organiser

typology
Where
Big model Checkpoint
It's not like you're in the middle of something
LoRA
CofyUI\models\loras
VAE
ComfyUI\models\vae
ControlNet
ComfyUI\models\controlnet

Can not delete folder: %s: No such folder The starter's next checkpoint bottom model will be enough to pick the right one for the 8G 03。

ix. Show your first graph

The software runs, the model is set, and now the map. Don't panic. Follow me

1. Loading the default workflow: Point Load Default, produce a set of connected base nodes, the most standard of the texture flow。

2. Modelling: Load Checkpoint bottom box selects your bottom model。

3. Writing hints: two CLIP Text Encode boxes - a profile cat, sunshine, high Quality; and below, undesired, like blackry, low Quality. Freshman first in English。

4. Dimensioning_Other Organiser。SD1.5 512 x 512, SDXL/Flux 1024 x 1024, making a mistake can create a ghost map or explode。

5. Freeze the samplerKSampler default step 20, CFG 7, start with the default。

Figure: Point Queue Prompt (or Ctrl+Enter), progress strip ran out, charted in Save Image, auto-stored ComfyUI\output。

CONGRATULATIONS, THE FIRST AI CHART CAME OUT。Two more trips, and the latter complex workflows are based on this。

The interface is English? Two steps to Chinese

Click to open Manager & Language / Language Settings → to select a CREDIT CREDIT REDUCTION. It doesn't matter if you can't find it。

Eleven, I stepped on the pit and ran it against it

phenomena
How
Chinese path error
Reload the English-only path
A BUNCH OF CUDAS
Update driver + restart
Depression failed/ Double click failed
Propertiestick Unlock + Kill Soft White List
Manager List Space
address bar plus /manager/repair fix
Failed to get version
Update Manager or Tunnel Agent
CUDA out of memory
Smaller, lot 1 and visible parameters
It's downloading dependent
I'M IN 15G. DON'T WAIT 10 MINUTES

Twelve, it's finished. Then what

At this point, whatever method you use, there's a running, graphic ComfyUI on the computer. This is just the first step — the software is in hand, the first picture is out, but it looks like it's a bad idea。

Take it easy, step by step. I'll take you nextMake sure the interface is clearReally understand what every node is doing。

statement:The content of the source of public various media platforms, if the inclusion of the content violates your rights and interests, please contact the mailbox, this site will be the first time to deal with.
TutorialEncyclopedia

WorkBuddy uses the tutorial to build the whole plan with a micro-intelligence applet that can be used in a minute

2026-8-10 9:27:02

productvideo

Copy Video AI: Video Transcript Tool to convert YouTube links or MP4 videos to structured AI video generation tips

2026-7-26 9:58:58

Search