Score-based diffusion models learn to reverse a gradual noising process. The forward process q(x_t|x_0) adds Gaussian noise over T timesteps: dx = −½β(t)x dt + √β(t) dW, gradually destroying data structure until x_T ~ N(0,I). The score function s_θ(x,t) ≈ ∇_x log p_t(x) points toward higher probability regions. A neural network learns this score by minimizing E[‖s_θ(x_t,t) − ∇_x log q(x_t|x_0)‖²]. Generation follows the reverse SDE: dx = [−½β(t)x − β(t)s_θ(x,t)]dt + √β(t) dW̄, walking from noise back to data. The arrows show the estimated score field, and the orange particles follow the reverse diffusion trajectory.