Diffusion Models

In a classic denoising diffusion probabilistic model, the forward process progressively corrupts training data according to a specified noise schedule. This forward process is generally fixed rather than a second neural network that must learn to add noise. A learned reverse process aims to recover structure through successive steps. One common DDPM training parameterization asks a network to estimate added Gaussian noise from a noisy example and its timestep. That estimate helps construct a reverse transition; a sampler is not simply required to subtract the raw prediction without coefficients or other terms. Noise prediction is one parameterization, not a rule for every later diffusion system. Text-conditioned latent diffusion adds control and representation choices. A text encoder can supply embeddings through cross-attention to influence denoising. The heavy denoising computation runs in a learned compressed latent representation, and a decoder reconstructs pixel-space output afterward. Compression trades representational detail against computational cost; it does not guarantee identical quality to every pixel-space alternative. Classifier-free guidance combines conditional and unconditional estimates, permitting a quality-and-diversity tradeoff without a separate guidance classifier. Classic DDPM training does not require an adversarial discriminator, but avoiding that objective does not by itself prove better diversity or reliability than all GANs. Iterative sampling can require many network evaluations. Progressive distillation research reduces a trained sampler's step count, with reported quality evaluated under specific benchmark conditions. Faster sampling and useful conditioning remain questions to assess on the actual system.