Step by Step: How Denoising Builds an Image

The last lesson established that a diffusion model generates images by repeatedly removing noise from a canvas of static. This lesson slows that process down and walks through what's actually happening on screen, step by step, if you could watch a generation happen in slow motion.

If you've ever watched a Midjourney or Stable Diffusion generation with progress previews turned on, you've seen this exact sequence play out: a blurry, formless blob resolving into rough shapes, then into something recognizable, then into a sharp, finished image. This lesson explains why it happens in that specific order.

What You'll Learn

• What each denoising step is actually doing to the image
• Why shapes and composition appear before fine detail
• What "sampling steps" means and why more steps isn't always better
• A mental model for the whole process you can carry through the rest of this course

Watching Noise Become an Image

Picture a generation as a strip of film with, say, 30 frames. Frame 1 is pure random static, indistinguishable from TV snow. Frame 30 is the finished image. Here's roughly what the frames in between look like:

• Early frames (roughly frames 1-8): Still mostly noise, but faint large-scale patterns start to appear, a vague sense of where light areas and dark areas will be, and roughly how the canvas is divided. Nothing recognizable yet, just the loosest hint of composition.
• Middle frames (roughly frames 9-20): Rough shapes emerge. If the prompt was "a golden retriever sitting in a field," you might start to make out a blob shaped vaguely like a dog against a blob of green. No fur texture, no facial features yet, just silhouette and placement.
• Late frames (roughly frames 21-30): Fine detail fills in. Individual strands of fur, the texture of grass, the shine in an eye, the folds of the birthday hat. The broad structure was already locked in by the middle frames; these final steps are refinement, not reinvention.

This coarse-to-fine order isn't a coincidence, it falls directly out of how noise works. A small amount of noise mostly obscures fine detail while leaving large-scale structure mostly visible, similar to how a slightly out-of-focus photo blurs texture but you can still tell it's a dog in a field. A large amount of noise obscures almost everything, including structure. So as the model removes noise from most-noisy toward least-noisy, structure is the first thing recoverable, and fine detail is the last.

Each Step Is a Small, Repeatable Decision

Every single one of those frames is produced by the exact same operation repeated: the model looks at the current (noisy) image, makes a prediction about what noise is present, and subtracts a portion of that predicted noise to produce the next, slightly cleaner frame.

It's worth being precise about one detail here: the model doesn't jump straight from noise to the final image in one leap, and it also doesn't remove 100% of its noise prediction in a single step. Early attempts at this kind of AI tried to predict the whole clean image directly in one shot, and the results were blurry and low-quality, because predicting an entire complex image from scratch in a single guess is a much harder problem than predicting "what does this look like with slightly less noise." Removing noise gradually, in small increments, turns one very hard problem into dozens of much easier ones, and that's the core insight that makes diffusion work well.

Sampling Steps: A Dial You (or the Tool) Can Turn

The number of denoising steps in a generation, often called "sampling steps" in tool settings, is a real, adjustable number. Fewer steps mean faster generation but a rougher, sometimes less coherent result, since the model has fewer opportunities to refine the image. More steps mean a slower generation with, up to a point, better quality and more refined detail.

That's "up to a point" because the relationship isn't a straight line. A jump from 10 to 25 steps usually produces a noticeably better image. A jump from 25 to 100 steps often produces a much smaller improvement for a lot more waiting. Most consumer tools pick a sensible default, often somewhere between 20 and 50 steps, that balances speed and quality so you don't have to think about it. If you've ever seen a "quality" or "steps" slider in an image generation tool's advanced settings, this is exactly what it's controlling.

Why This Matters for Understanding "AI Mistakes"

This step-by-step structure also explains a pattern you've probably noticed in AI-generated images: bigger, structural mistakes (an extra finger, a strange extra limb, garbled text) tend to be locked in early and hard to fix, while texture-level flaws are comparatively rare and minor by comparison.

That's because structure is decided in the early, coarse steps, when the model is working with the least information and the most noise to interpret. If the model misjudges roughly where a hand should be during those early steps, later steps can only refine and add detail to that already-wrong structure, they can't go back and re-decide "wait, should there be five fingers or six?" That's also part of why some tools offer an "inpainting" or "regenerate this region" feature, it's often easier to redo a small structural mistake by rerunning the noise-to-image process on just that patch than to fix it after the fact.

Key Takeaways

• A generation runs through dozens of small denoising steps, each producing a slightly clearer version of the image than the last.
• Structure and composition emerge in the early-to-middle steps; fine detail and texture fill in during the later steps, because noise obscures detail before it obscures large-scale shape.
• The model works by predicting and removing a little noise at a time, rather than guessing the whole finished image in one shot, which is what makes the results coherent.
• "Sampling steps" is a real, adjustable setting: more steps generally means better quality up to a point of diminishing returns, and fewer steps means faster but rougher results.
• Structural errors (odd hands, garbled text) get locked in during early steps and are harder to fix than texture-level flaws, which is why they're the most common visible mistakes in AI-generated images.