A diffusion model generates images by learning to reverse a noising process. During training it takes real images, adds random noise in small steps…
Grasp diffusion as reversing a noising process.
A diffusion model rests on one elegant idea. Take a real image and repeatedly add a little random noise until it's indistinguishable from static — that's easy and needs no learning. The clever part is learning to run this backwards: to look at a noisy image and estimate what noise to remove to make it slightly cleaner.
If a model can reliably remove a bit of noise, you can start from pure random noise and apply that step over and over, gradually revealing a coherent image that was never there before. Generation is denoising, repeated.
See how the model learns to predict and remove noise.
Training uses the easy direction to teach the hard one. The model takes a real image, adds a known random amount of noise (the forward process), and is asked to predict that noise. Because we know exactly what noise we added, we can score how close its guess is and improve it.
Do this across millions of images and many noise levels, and the model becomes an expert noise-remover at every stage — from nearly clean to almost pure static. That single learned skill is all generation needs.
Turn learned denoising into images that match a prompt.
To generate, the model starts with a fresh field of random noise and applies its denoising step many times (often tens of steps), each pass removing a bit more noise until a clear image emerges. More steps generally means higher quality but slower generation.
To make the image match a prompt, the denoising is guided by the text. The prompt is encoded and fed in at each step so the model removes noise in the direction of 'a red bicycle in the rain,' not just any image. This conditioning is how text-to-image works.
Understand why diffusion won and what it doesn't do.
Diffusion models became the standard for image generation because they produce high-quality, diverse images and train stably, avoiding some instabilities of earlier approaches. The same denoising idea extends to audio and video too. The main trade-off is speed: because generation takes many steps, it's slower than a single forward pass, which is why much research focuses on cutting the number of steps.
Two to correct: a diffusion model doesn't paste together bits of training images — it generates pixels from noise using learned patterns. And it isn't retrieving a stored picture; each run from different noise gives a different image. Generation is probabilistic denoising, not copy-paste or lookup.
A diffusion model learns to reverse noising: in training it adds known noise to real images and learns to predict it, becoming an expert denoiser. To generate, it starts from pure noise and denoises step by step into a coherent image, guided at each step by a text prompt so the result matches. It won for high-quality, diverse, stable generation, extends to audio and video, and its main cost is the many steps — and it generates pixels rather than copying or retrieving images.
Explain to a friend why a diffusion model can create a picture of something that has never existed, walking through how it starts from noise, why the many steps matter, and how the text prompt steers what appears.
What is the core idea of a diffusion model?
Diffusion generates by repeatedly denoising, sculpting an image out of random noise using patterns it learned.
How is a diffusion model trained?
Training uses the easy forward noising to teach the hard reverse denoising, scored against the known added noise.
How does a diffusion model make an image match a text prompt?
Text conditioning guides each denoising step so the emerging image aligns with the prompt — the basis of text-to-image.
What is the main trade-off of diffusion models?
The multi-step process yields high quality but is slower, so much work focuses on achieving good images in fewer steps.