Prompt Description

A schematic diagram, presented in a clean, academic style, illustrates the distinction between the training and inference processes of a diffusion model applied to time series data. The left side depicts the training phase: beginning with real, clean data, Gaussian noise is iteratively added through a forward diffusion process. The model is trained using ground-truth noisy samples. Smooth, consistent arrows indicate a stable data distribution. The right side illustrates the inference (sampling) phase: starting from pure noise, a reverse diffusion process is performed, where the model iteratively uses its own prior predictions as inputs. Minor, accumulated prediction errors are visually emphasized, resulting in a shift in the input distribution relative to the training distribution. "Training Distribution" and "Inference Distribution" are clearly labeled, and their mismatch is depicted using diverging arrows or offset trajectories. A minimal, flat design is employed, featuring a white background and soft colors (blue for training, orange or red for inference).