Denoising Diffusion Probabilistic Models (DDPM)

Generates images by learning to reverse a forward diffusion process that gradually adds Gaussian noise to images over T timesteps. The model trains a neural network (typically a U-Net with attention mechanisms) to predict noise at each reverse step, then samples new images by starting from pure noise and iteratively denoising through learned reverse steps. This approach enables stable, high-quality image synthesis without adversarial training or autoregressive decoding.

Trains a U-Net architecture with sinusoidal positional embeddings of the diffusion timestep to predict Gaussian noise added at each step. The network uses skip connections, multi-scale feature processing, and optional cross-attention layers for conditioning on external signals (text, class labels). Timestep information is injected via learned embeddings that modulate network activations, enabling the same model to handle all T timesteps without separate models per step.

Trains the diffusion model by optimizing a score-matching objective, which is equivalent to predicting the noise added at each timestep. The score function (gradient of log probability) is approximated by the neural network, and the training objective minimizes the L2 distance between predicted and actual noise. This connection to score-based generative modeling provides theoretical grounding and enables efficient training without explicit likelihood computation.

Trains the diffusion model by optimizing a variational lower bound (ELBO) on the log-likelihood of the data. The training objective decomposes into a sum of KL divergence terms between the forward and reverse processes at each timestep, which simplifies to an L2 loss on noise prediction when using a fixed linear noise schedule. This principled probabilistic framework ensures stable convergence without adversarial losses or careful discriminator tuning.

Implements a Markov chain that gradually adds Gaussian noise to images over T timesteps using a fixed linear or cosine noise schedule. At each step t, noise is added according to q(x_t | x_0) = sqrt(alpha_bar_t) * x_0 + sqrt(1 - alpha_bar_t) * epsilon, where alpha_bar_t is a cumulative product of noise levels. This enables efficient one-shot sampling of noisy images at any timestep without sequential application, critical for efficient training.

Generates images by iteratively denoising from pure Gaussian noise through T reverse steps, where each step applies the learned reverse process p_theta(x_{t-1} | x_t) = N(x_{t-1}; mu_theta(x_t, t), Sigma_t). The mean is predicted by the U-Net, while variance can be fixed (using forward process variance) or learned. Sampling is deterministic at t=0 (no noise added) and stochastic at earlier steps, enabling controlled generation with optional temperature scaling.

Enables conditional image generation (e.g., text-to-image) by training the model on both conditioned and unconditional samples, then guiding the reverse process toward the conditioned distribution during sampling. At each denoising step, the predicted noise is adjusted as epsilon_guided = epsilon_uncond + w * (epsilon_cond - epsilon_uncond), where w is a guidance scale. This approach avoids training a separate classifier and enables flexible control over condition strength.

Enables fast approximate sampling by reducing the number of denoising steps from T (typically 1000) to a smaller number (e.g., 50) using techniques like DDIM (Denoising Diffusion Implicit Models) or DPM-Solver. These methods reformulate the reverse process as an ODE or use higher-order solvers to skip timesteps while maintaining sample quality. The key insight is that the reverse process doesn't require stochasticity; deterministic sampling with larger steps can approximate the full diffusion trajectory.

Enables image inpainting by conditioning the reverse diffusion process on known pixels while allowing the model to generate missing regions. During sampling, at each step, known pixels are replaced with their noisy versions at that timestep (computed via the forward process), while unknown pixels are denoised by the model. This approach requires no special training; any trained diffusion model can be adapted for inpainting by masking during sampling.

Enables image super-resolution by conditioning the reverse diffusion process on a low-resolution image. The low-resolution image is upsampled (via interpolation or learned upsampling) and used as conditioning at each denoising step, guiding the model to generate high-resolution details consistent with the low-resolution input. This approach can be implemented via concatenation, cross-attention, or other conditioning mechanisms, and requires training on paired low/high-resolution images.

Applies diffusion in a learned latent space (via a VAE encoder) rather than pixel space, enabling efficient generation of high-resolution images. The VAE compresses images to a lower-dimensional latent representation (e.g., 4x-8x spatial compression), then diffusion operates on latents. This approach reduces computational cost by ~50-100x (due to quadratic scaling with spatial dimensions) while maintaining generation quality, enabling 512x512+ generation on consumer GPUs.