How to Build Your First Workflow in ComfyUI

November 22, 2026

ComfyUI is a node-based interface for running image and video diffusion models locally. You connect a checkpoint loader to a text encoder, a sampler and a save node, then press queue. There is no prompt box on a blank canvas; the graph is the prompt.

The first hour is confusing because nodes appear by typing a name, and a graph with one missing link runs without error and outputs nothing. Below is the smallest working graph and the three mistakes that stop it.

What is the smallest graph that produces an image

Four nodes:

  1. Load Checkpoint: pick a model file from your models folder.
  2. CLIP Text Encode: two of these, one positive prompt, one negative.
  3. KSampler: the sampling node, takes model, conditioning and a latent.
  4. Empty Latent Image: sets width and height, for example 1024×1024.
  5. VAE Decode then Save Image: turns the latent into a file.

Connect checkpoint model to KSampler, both CLIP encoders to KSampler’s positive and negative inputs, latent image to KSampler’s latent input, KSampler output to VAE Decode, and the decode output to Save Image. Then queue.

Why does nothing happen when I press queue?

Three usual causes:

  • A node is not connected. ComfyUI highlights missing inputs in red; look for the red outline before queueing.
  • The model file is missing from the expected folder. The checkpoint node will show a filename you do not have on disk.
  • The queue processed but the save node writes to the output folder, not the browser. Check the file path shown in the node.

A fourth, less obvious cause: two KSamplers sharing one latent, where only one is connected to a decode. The second silently wastes time.

If your graph runs instantly and produces an empty file, the sampler had no conditioning. Check that both encoders are wired to the sampler, not just one.

How much VRAM do you need

Resolution Rough VRAM
512×512 6GB
1024×1024 8 to 10GB
1024×1024 with two control layers 12GB+
Video, short clip 16GB+

On 8GB, a 1024 square with several control nodes will spill to system memory and take minutes per image. Start at 512 to confirm the graph works, then scale up.

What should you learn second

Saving your graph as a JSON workflow and reusing it. Then the node groups: an upscale group and a face-detail group are the two most common additions, and both are copied from shared workflow files rather than built from scratch.

ComfyUI’s ecosystem is the advantage and the trap: thousands of custom nodes, some broken after an update. Keep a working workflow file before adding anything.

Where does a hosted tool make more sense

If you want images without managing models, a hosted generator removes the setup. Cleanup.pictures covers the removal tasks that follow generation, and if you need a designed output rather than a raw image, a design tool with templates is a better starting point than a node graph.

Who should not start here

Anyone who needs one image quickly. The setup is hours, and a hosted generator answers in seconds. ComfyUI earns its place when you need control over every step and repeatable pipelines.

Quick Summary

  • A minimal ComfyUI graph is checkpoint, two CLIP encodes, sampler, latent, decode, save.
  • Red outlines mean an unconnected input; that is why queue appears to do nothing.
  • Reset your starting resolution to 512 on low VRAM before scaling up.
  • Save the workflow JSON before installing custom nodes.
  • Building here is worth it for repeatable pipelines, not one-off images.
Tags: