Diffusion models generate images by learning to reverse a gradual noising process. Forward process: Gradually add Gaussian noise to image over T steps until it becomes pure noise. Defined by noise schedule β₁...βT. Reverse process: Learn to denoise at each step. Neural network predicts noise (or clean image) given noisy input and timestep. Training: Add noise to real images at random timesteps, train U-Net to predict the added noise (or original), MSE loss between predicted and actual noise. Sampling: Start from random noise → iteratively denoise using learned model → each step recovers signal → final step produces clean image. Noise schedules: Linear, cosine, learned. Affect training and sample quality. DDPM vs DDIM: DDPM (stochastic sampling, 1000 steps), DDIM (deterministic, fewer steps, faster). Architecture: U-Net with attention, residual connections, timestep conditioning. Conditioning: Class labels, text embeddings (cross-attention), other signals. Advantages over GANs: More stable training, better mode coverage, easier to control. Foundation of modern image generation (Stable Diffusion, DALL-E, Midjourney).