01The big idea: denoise your way to a picture

Imagine starting with a screen of TV static and slowly wiping away the fuzz until a clear picture appears underneath - that gradual clearing is the trick behind most image generators, and the method has a name: diffusion. A diffusion model does not paint a picture in one stroke, and it does not paste together bits of stored photos. It begins with a square of pure random noise - the TV static - and then takes many small steps, removing a little noise at each one. Every step nudges the image toward something that matches your prompt, until a clear picture emerges. Step through it below and watch the static resolve into a shape.

• Generation starts from random noise, not from a blank canvas or a stored image.
• It proceeds iteratively - many small denoising steps, each one removing a bit more noise.
• At each step the model estimates the noise to remove, moving the sample toward a coherent image.

02Training: add noise, then learn to undo it

How does a model learn to denoise? Cleverly - by being shown the problem in reverse. During training, real images are deliberately corrupted with noise, a little more at each step, until almost nothing is left. The model job is to learn the opposite direction: given a noisy image, predict the noise so it can be removed. Generation is then just running that learned reverse process, starting from fresh noise. Step through it.

• The forward process adds noise to real images; it is only used to create training examples.
• The model learns the reverse: predict the noise so it can be removed, step by step.
• Sampling is the reverse run from scratch - noise in, new image out - so outputs are original, not copies.

03Steering the result: text conditioning & guidance

Plain denoising would produce some image, but not necessarily the one you asked for. Text conditioning is what ties the picture to your words: the prompt is fed in as a condition at every denoising step, so the emerging image matches the description. A few related ideas make this work - and let you dial how literally the model follows you. Explore them below.

• The prompt conditions every step, steering the picture toward the words.
• CLIP-style text-image models give the prompt and image a shared language so guidance works.
• Classifier-free guidance trades stronger prompt-following against variety and naturalness.

04Check your understanding

05Working in latent space + practical controls

Two last pieces make modern image generation practical. First, latent diffusion: instead of denoising millions of raw pixels, the process runs in a compact, compressed "latent" space and only decodes to a full picture at the end - far cheaper, which is why high-resolution generation became widely usable. Second, a toolkit of controls that go beyond a single prompt: editing parts of an image, extending it, or pinning down its structure.

• Latent diffusion denoises a small compressed representation, then decodes to pixels - much more efficient than working on raw pixels.
• Inpainting regenerates a masked region (remove or replace an object); outpainting extends an image past its borders.
• ControlNet-style conditioning adds spatial controls - edges, pose, depth - so the output follows a structure you provide.