diffusion-models #

Shows the diffusion workflow in three stages: (1) a forward time-indexed Gaussian noising Markov chain that gradually corrupts a simple blocky “image” according to a noise schedule, (2) a learned denoiser/score model εθ(x_t,t) visualized as a vector field acting on the noisy sample, and (3) the reverse iterative generative process that uses εθ to denoise step-by-step from pure noise back to data.

canvasclick to interact

⏮◀◀▶▶STEP0.25x1xZOOM

t=0s

practical uses #

01.Text-to-image generation (DDPM/latent diffusion sampling loop)
02.Audio or signal generation via iterative denoising
03.Image editing/inpainting by noising then reversing with conditioning

technical notes #

Pure Canvas2D. Uses a deterministic pseudo-random function for stable “Gaussian-like” noise per cell and time step. Animation cycles through forward→learn→reverse in ~4.2s using provided cubic ease(). Grid-aligned snapping and rectangular pixels enforce a blocky retro aesthetic; green highlights indicate the active time step t and stage.

← stochastic-gradient-descent token-embeddings →