Training teaches one skill: look at a noisy image and predict the noise in it. Generation runs that skill in a loop, subtracting a little predicted noise each step while guidance leans each prediction toward your prompt. The image is not assembled. It is uncovered.
Left: x_t, the noisy image. Middle: ε, the noise the model predicts. Right: x_0, what the model currently believes the final image is. Watch this one first. Bottom: the spread of pixel values in x_t, starting as a bell curve and ending as the image's own shape.
A real model learns its noise prediction from billions of images. Here the true target image plays that part (an oracle), blurred in proportion to the noise and mixed with an unguided guess that only knows the average of all three pictures. That way the process is visible without a trained network. The cosine schedule (T = 1000), the forward equation, the DDIM update, the guidance formula, and the clamp are the real math.
Switch the prompt mid-run and watch the right pane pivot. Set guidance to 0 for mush and 10 for clipped, overcooked contrast. Compare 10 steps (grainy, blotchy) with 50 (clean). Drag the scrub slider to run the forward process by hand.