A diffusion model learns how to recover structured data from noise. During generation it begins with noise or a partially corrupted input and repeatedly predicts a cleaner state, producing images, video, audio, 3D representations or other outputs.
Traditional sampling can require many denoising steps, which adds latency. Research improves speed through better samplers, distillation, sparse computation or parallel generation blocks. Fewer steps can be much faster, but the method must preserve quality and temporal or spatial consistency.
