Earngenix Logo
Skip to main content

Glossary · ComfyUI Basics

What Is a Diffusion Model? (Explained Simply)

By Earngenix Team ·

⚡ Quick Answer

A diffusion model is an AI technique for making images. It starts with a canvas of random noise (think TV static) and removes a little bit of that noise at a time, over many steps, until a clear image appears — guided the whole way by your text prompt. It is not a single file; it is the method that every checkpoint you load in ComfyUI is built on.

The name comes from a real physics idea — particles spreading out, or "diffusing", until they're evenly mixed. AI researchers found that if you teach a model to reverse that spreading process, run backwards, you can turn pure noise into a photo.

How It Actually Works, Step by Step

During training, the model is shown millions of real images with noise added on top, a little more at each step, until the image is pure static. It learns to predict exactly how much noise was added at each step — and, importantly, how to remove it.

When you generate a new image, that process runs backwards. ComfyUI starts with a canvas of random noise (this is what the Empty Latent Image node creates) and the model predicts and subtracts a bit of noise, over and over, each time nudged by your prompt so the noise resolves into what you asked for rather than something random.

Tip: This is exactly what's happening every time you watch the preview image sharpen up while a generation runs — each flicker is one denoising step.

Where You'll See It in ComfyUI

You never load "a diffusion model" by itself as a labeled thing — you load a checkpoint (a trained, saved version of one) through the Load Checkpoint node. The actual step-by-step denoising is carried out by the KSampler node, which repeats the noise-removal process for however many steps you set.

Quick Example

Stable Diffusion, SDXL, FLUX.1, and Qwen-Image are all diffusion models — different architectures and training data, but the same core idea of denoising noise into an image. LTX-2 applies the same technique to video, denoising a sequence of frames instead of just one image.

Rule of thumb: "diffusion model" is the technique. The .safetensors file you download and load is a checkpoint — one specific, trained result of that technique.

Frequently Asked Questions

Yes. "Stable Diffusion" is a family of diffusion models. A specific file like sd_xl_base_1.0.safetensors is a checkpoint — one trained version of that diffusion model.

Most modern ones do, including Stable Diffusion, SDXL, FLUX, and Qwen-Image. Older tools used a different technique called a GAN (Generative Adversarial Network), which trains two networks against each other instead of denoising step by step.

Each step only removes a little bit of noise. Removing it all at once produces a blurry, low-quality result, so the model repeats the process 20–50 times, refining the image a little more each pass.

No — the number of denoising steps stays the same either way. A faster GPU just completes each step quicker, so the whole generation finishes sooner.

See It In Action

Ready to see a diffusion model run?

Our beginner's guide walks through your first text-to-image generation, step by step.

Published: 2026-09-09 · Last updated: 2026-09-09

Discussion

Join the discussion

Sign in to leave a comment or reply

💬

No comments yet

Be the first to share your thoughts!