ComfyUI startup reporting errors? One article to resolve common ComfyUI faults and errors.

If you're aAI drawing enthusiasts, or are exploring the advanced play of Stable Diffusion, theComfyUIIt must not be unfamiliar. As an open source node-based workflow tool, it has attracted countless fans with its flexibility and high degree of freedom, and is called "the Lego blocks of AI mapping". But there are two sides to everything, ComfyUI's power is also accompanied by a certain threshold to get started, especially in the building process, a variety of "metaphysical errors" can make people crazy. Today, as a crawler over the "old driver", I want to break open the common failures and solutions to tell you, full of dry goods, clear logic, to ensure that you can read less detour!

I. Dependency libraries "roadblocks": what about startup failures?

Phenomenon: I knocked down python main.py on the spur of the moment, but the screen showed a bunch of red characters and ModuleNotFoundError: No module named 'torch' popped up, which instantly made people doubt life.

Reason: Most of the time, this is because the Python environment is not properly configured, the dependency libraries are not fully installed, or the version is not correct, ComfyUI has strict requirements for torch, numpy and other basic libraries, and any one of them will not be able to work.

Solution Trick:

1. Open the ComfyUI directory, find requirements.txt, and directly sentence pip install -r requirements.txt to batch install.
2. Domestic network jam? Don't panic, change the Tsinghua mirror source: pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple, the speed flies.
3. If it still reports errors, try installing torch manually, e.g. pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118, and be careful to pick the CUDA version according to your graphics card.
4. Finally, check the Python version, 3.8 or 3.10 is the most stable, python --version will tell you.

Pros and cons chat: this method is simple and rough, suitable for novices to get started quickly. However, the disadvantage is that it depends on the network environment, and the mirror source is occasionally unstable, so you may need to try a few more times.

Real case: I have a friend, loaded to half stuck in torch, tossed half a day to find that is not installed CUDA driver, after a second to solve. So, don't forget to check your graphics card driver!

Second, the model file is "missing": what if the nodes are red?

Phenomenon: the workflow is built, a run, the node red like New Year's lanterns, prompting FileNotFoundError, the model file can not be found.

Reason: Either you didn't download the model, or the path is wrong, ComfyUI looks for .ckpt or .safetensors files from models/checkpoints by default, it won't recognize them if they're in the wrong place.

Solution Trick:

1. Go to Hugging Face or Civitai and drop a base model like sd-v1-4.ckpt into models/checkpoints.
2. Don't make the file name too complicated, change it to sd1.4.ckpt which is simple, avoid Chinese and special characters.
3. Right click on the node in the workflow and check if the model name and the actual file match.

Advantages and disadvantages of chatting: this trick is simple to operate, the model put in the right place basically on the line. But the disadvantage is that newbies tend to ignore the directory structure, it is recommended to follow the default honestly, do not be smart to change the path.

Case in point: I've seen people throw their models on the desktop, then look for tutorials all over the world, and finally realize that it's a path problem, and move them back to the desktop to fix it in seconds. Remember, ComfyUI is dead set in its ways, and it just won't work if the path is wrong!

Third, the video memory emergency: CUDA error reporting "heartbreaking tears"

Phenomenon: When generating pictures, the screen suddenly pops up CUDA out of memory, the graphics card directly "strike".

Reason: not enough video memory. High-resolution images, complex workflows, or too many programs running in the background can overwhelm the graphics card.

Solution Trick:

1. Lower the resolution, try 384x384 if 512x512 is too much of a stretch.
2. Starting with a parameter such as python main.py --lowvram, low video memory mode can be a life saver.
3. Turn off "background monsters" that eat video memory, such as the dozens of tabs in your browser.
4. Check the workflow, don't load a dozen LoRAs at once, streamline it.

Pros and Cons Chat: This method is practical and allows even low end graphics cards to play with ComfyUI. but the downside is that the generation speed may become slower and the image quality has to be compromised appropriately.

Real life example: I used a 3060 with 6GB of video memory, the initial run of 512x512 crashed straight away, after adding -lowvram it stabilized, it's slower but at least it works.

Fourth, the plug-in "strike": nodes do not display?

Phenomenon: I installed a popular plug-in, such as ComfyUI-Manager, but the interface didn't change, or reported Import failed.

Reason: Plugins are not installed properly or dependencies are not kept up, ComfyUI has a rich plugin ecosystem, but the installation process is a bit "wild".

Solution Trick:

1. Use ComfyUI-Manager's own function, click Install Missing Custom Nodes, it's done automatically.
2. To install manually just go to the custom_nodes folder, git clone the plugin address, and run down install.py (if available).
3. Reboot after loading, don't be lazy.

Pros and Cons Chat: Manager automation is fragrant and saves effort, but manual installation is more flexible and can install some cold plugins. The disadvantage is that when the network is not strong, cloning the repository is easy to fail.

Real case: I was stuck when I installed ControlNet plugin, but later found out that the proxy was not on, and it finished smoothly after hanging the global proxy. Newbie remember to check the network!

Fifth, the interface "stealth": localhost how can not open?

Phenomenon: After booting up and visiting http://localhost:8188, it either goes blank or fails to connect.

Reason: The port is occupied, or the firewall is messing with it. 8188 is the default port, but that doesn't mean no one is grabbing it.

Solution Trick:

1. Check ports: Windows use netstat -aon | findstr 8188, Linux use lsof -i:8188, find the occupying process to kill it.
2. Change port: add --port 8199 at startup, then access the new address.
3. Firewall off, or give ComfyUI a green light.

Advantages and disadvantages: changing ports is the most direct and has a high success rate, but you have to memorize the new address. Firewall adjustment is more thorough, but the operation is a little troublesome.

Real case: I once ran two ComfyUI, port conflict directly confused, changed to 8199 after a perfect solution. Don't fight with this port thing!

VI. Updated "disaster": what's wrong with workflow?

Phenomenon: After updating ComfyUI or plugin, the previous workflow is all red and the nodes are not recognized.

Reason: The new version changed the node names or parameters and the old workflow could not keep up.

Solution Trick:

1. Look at the changelog (GitHub has it) for change points.
2. Open the JSON file and manually change the node names or parameters.
3. Roll back if you don't want to toss it, delete the new version, unzip the old one and run it.

Pros and cons chat: manual change is accurate but time-consuming, rollback saves time but gives up new features. Depends on your needs to choose.

Real life example: my KSampler parameters changed after the update, and it took adjusting the number of sample steps and CFG to save it. It's good to update, but it's better to be cautious!

At the end of the day: the "sweetness" and "bitterness" of ComfyUI.

The advantages of ComfyUI are obvious: node operation is intuitive, the degree of freedom is overwhelming, and it can realize all kinds of play from simple drawing to complex post-processing. But it also has a "bitter": high threshold for getting started, frequent error reporting, hardware and environmental requirements are not low. However, as long as you master the experience of these potholes, 90% problems can be solved.

Tip:

• Logging is a life saver, look at the terminal more often when errors are reported.
• The community is great, Discord and B-site tutorials are worth flipping.
• Load it with a virtual environment to save fighting with other projects.

I hope this dry goods can let you less fall a few heels, smooth play ComfyUI. have any specific problems, feel free to leave a message, let's solve it together!