denoising

**Denoising Diffusion Probabilistic Models (DDPM)** is **a generative model class that iteratively denoises corrupted data samples over a series of diffusion steps — learning to reverse a forward diffusion process and enabling high-quality generation of diverse samples from learned distributions**. Denoising Diffusion Probabilistic Models provide an alternative to adversarial and autoregressive approaches for generative modeling, based on thermodynamics-inspired diffusion processes. The forward diffusion process gradually adds Gaussian noise to data samples over a fixed number of timesteps until the data becomes pure noise. The reverse diffusion process learns to denoise step-by-step, gradually reconstructing meaningful samples from noise. The key insight is that this reverse process can be parameterized as a neural network that predicts either the noise added at each step or the original data itself. The loss function is simple: the network is trained via mean-squared error to predict the added noise given the noisy sample and timestep. DDPM training is stable and doesn't require adversarial losses or mode collapse concerns affecting GANs. The diffusion process naturally gives rise to a hierarchical representation of data at different scales of noise, providing useful inductive biases for learning. Sampling involves starting from pure noise and applying the learned denoising network iteratively for many steps, typically 1000 or more. This many-step sampling is computationally expensive compared to single-forward-pass generative models, motivating research into accelerated sampling schedules. Guidance mechanisms like classifier guidance enable conditional generation, where a classifier provides gradients steering the diffusion process toward specific classes. Unconditional DDPMs have achieved state-of-the-art image generation quality, and conditioning mechanisms enable diverse applications from text-to-image generation to inpainting. The DDPM framework connects to score-matching and energy-based models, providing theoretical understanding. Variants like denoising score-based generative models use continuous diffusion processes rather than discrete timesteps, enabling continuous control of generation quality. DDPM has been successfully applied to audio, 3D shapes, and protein structure generation, demonstrating generality beyond images. The connection between diffusion models and consistency distillation enables faster sampling while maintaining sample quality. **Denoising diffusion probabilistic models represent a stable, scalable, and theoretically grounded approach to generative modeling with state-of-the-art quality and broad applicability across modalities.**