Model too large for the card, or several models loaded at once. Use fp8 or GGUF versions, VAE Decode (Tiled), smaller batch; see the VRAM guide.

Trends

.ai

the best open source ai models

What is ComfyUI? A beginner's guide to nodes, workflows and the first graph

Q: Nodes are red after loading a workflow

Custom nodes the workflow uses are not installed. Manager > Install Missing Custom Nodes, restart ComfyUI.

Topic: ComfyUI

By Lookout

Published 2025-12-17

ComfyUI explained for people coming from Automatic1111, Fooocus or a web app: why it is a graph, the seven nodes in the default workflow and what each does, how to load a workflow from an image, where models go, the Manager for missing nodes, and when a simpler tool or the cloud is the better choice.

Overview

ComfyUI is a free program that runs image and video models as a graph: boxes (nodes) that each do one job, connected by wires that carry the model, the prompt, the latent image and the final pixels. Where

Automatic1111

gives you one form with fixed fields, ComfyUI lets you wire the same pieces however the newest model needs, which is why

and most 2025-2026 releases ran in ComfyUI first and often only there.

The default workflow has seven nodes and that is all a beginner needs to understand: Load

Checkpoint

(the model, its text encoder and VAE), two CLIP Text Encode nodes (positive and

negative prompt

), Empty Latent Image (the size), KSampler (seed, steps,

CFG

, sampler, scheduler), VAE Decode (latent to pixels) and Save Image. Every other workflow you will download is this chain with extra nodes inserted: a Load

LoRA

between the checkpoint and the sampler, a

ControlNet

before the sampler, an

upscaler

after the decode.

ComfyUI is also where workflows are shared: a PNG saved by ComfyUI contains its whole graph, and dragging it onto the canvas rebuilds it. The ComfyUI node guide on this site documents every common node and the errors it throws. If you only want a prompt box, Fooocus or a web app like

BitVector

is a gentler start, and

PirateDiffusion

runs many of the same ComfyUI workflows on its servers under a single command.

Reference

Name	Type	What it is
Load Checkpoint	node	Outputs MODEL, CLIP and VAE from one file. For Flux, Qwen , Wan and Z-Image you use Load Diffusion Model + DualCLIPLoader/CLIPLoader + Load VAE instead.
CLIP Text Encode	node	Turns a prompt into conditioning. One for the positive, one for the negative. Weight syntax (word:1.2) works here.
Empty Latent Image	node	Width, height, batch size. Multiples of 8 (SD) or 16 (Flux, Wan).
KSampler	node	seed, control_after_generate (fixed/random), steps, cfg, sampler_name, scheduler, denoise. The heart of the graph.
VAE Decode	node	Latent to image. Use VAE Decode (Tiled) when large sizes run out of memory .
Save Image / Preview Image	node	Save writes to ComfyUI/output with the workflow embedded; Preview only shows.
Load LoRA	node	Between checkpoint and sampler; strength_model and strength_clip. LoraLoaderModelOnly for Flux-style graphs.
ComfyUI Manager	extension	Installs missing custom nodes from a workflow ("Install Missing Custom Nodes") and updates everything.

Just want a prompt box?

If the graph is more than you need today, BitVector Prism runs the same SDXL, Flux and Qwen models from a plain web app with nothing to install. Come back to ComfyUI when you want the wiring.

Try BitVector

Step by step

Install: ComfyUI Desktop (Windows, Mac) is a one-click installer; the portable zip (Windows, NVIDIA) unpacks anywhere; Linux and Mac can run from source with Python 3.12. Our ComfyUI app page lists the links.
Put a checkpoint in ComfyUI/models/checkpoints (an
SDXL
fine-tune from this site is a good first one), LoRAs in models/loras, VAEs in models/vae.
Start ComfyUI, open the browser tab it prints, and load the default workflow (Workflow menu > Browse Templates > Image generation).
Click the checkpoint name in Load Checkpoint and pick your file. Type a prompt in the top CLIP Text Encode, a short negative in the bottom one.
Set Empty Latent Image to 1024x1024 for SDXL. In KSampler set steps 28, cfg 6, dpmpp_2m, karras, seed control "randomize".
Press Queue (or Ctrl+Enter). The image appears in Save Image and is written to the output folder with the workflow inside it.
Insert a Load LoRA: double-click the canvas, type "Load LoRA", drop it, and rewire MODEL and CLIP from the checkpoint through it to the sampler and the text encoders.
Install ComfyUI Manager from its GitHub page, restart, and use "Install Missing Custom Nodes" whenever a downloaded workflow shows red nodes.

Examples

The default seven-node chain

Load Checkpoint -> CLIP Text Encode (positive) + CLIP Text Encode (negative) -> KSampler <- Empty Latent Image

KSampler -> VAE Decode (VAE from Load Checkpoint) -> Save Image

Adding a LoRA

Load Checkpoint -> Load LoRA (strength_model 0.8, strength_clip 0.8) -> MODEL to KSampler, CLIP to both CLIP Text Encode nodes

Flux dev graph

Load Diffusion Model (flux1-dev-fp8) + DualCLIPLoader (clip_l, t5xxl_fp8) + Load VAE (ae) -> FluxGuidance (3.5) -> KSampler (euler, simple, 24 steps, cfg 1) -> VAE Decode

Tips

Drag any ComfyUI-made PNG onto the canvas to load its workflow; the model page examples on this site often come from such files.
Ctrl+Shift+V pastes nodes with their connections; Ctrl+B bypasses a node (useful for toggling a LoRA); Ctrl+M mutes it.
Red nodes mean a missing custom node; a red border on a widget means a missing file (the model name does not match anything in the folder).
Keep the models folder shared with
Forge
or A1111 through extra_model_paths.yaml instead of copying gigabytes.
Use the Preview Image node while iterating and Save Image only for keepers, or the output folder fills with hundreds of files.
The node guide on this site explains every node named here with its common errors; the ComfyUI troubleshooting guide covers install and CUDA problems.

Troubleshooting

Nodes are red after loading a workflow

Why it happens

Custom nodes the workflow uses are not installed.

How to fix it

Manager > Install Missing Custom Nodes, restart ComfyUI.

"Value not in list" on a model widget

Why it happens

The workflow names a file you do not have, or it is in a different folder.

How to fix it

Click the widget and pick your own file; download the named model from its page on this site.

Nothing happens when I press Queue

Why it happens

A required input is unconnected, or the browser lost the websocket.

How to fix it

Look for a highlighted node with a missing input; check the terminal window for the error line.

Out of memory

Why it happens

Model too large for the card, or several models loaded at once.

How to fix it

Use fp8 or GGUF versions, VAE Decode (Tiled), smaller batch; see the

VRAM

guide.

Images look washed out

Why it happens

Missing VAE in a split-loader graph, or an SDXL fp16 VAE issue.

How to fix it

Load the family VAE with Load VAE and wire it to VAE Decode.

ComfyUI workflows without the install

PirateDiffusion runs ComfyUI workflows on its own GPUs: /workflow /run:name plus your prompt in Telegram, with the models, LoRAs and custom nodes already in place. Unlimited runs on a fixed price.

Try PirateDiffusion

Questions

Is ComfyUI harder than Automatic1111?

The first hour, yes. After that it is easier to understand what is happening, and it is the only tool that runs everything new on day one.

Does ComfyUI work on Mac?

Yes, on

Apple Silicon

through PyTorch MPS; slower than NVIDIA, and some nodes need fp32. The Mac guide on this site covers it.

Can I run ComfyUI without a GPU?

It runs on CPU for small

SD 1.5

jobs, minutes per image. The practical answer is the cloud: PirateDiffusion runs ComfyUI workflows server-side under /workflow commands.

Where do I find workflows?

Model pages on this site link to the authors' workflow files; ComfyUI's own template browser has one per family; any ComfyUI PNG carries its workflow.

What is the difference between ComfyUI Desktop and portable?

Desktop is an installer with automatic updates and a bundled Python; portable is a zip with the same program you update by script. Desktop for most people.