Step-Distillation & Synthetic Data methods (consistency distillation, adversarial diffusion distillation, and guidance distillation) compressed 30–50 step pipelines down to 1–4 steps, enabling near-real-time image generation on Serverless GPU & Inference Clouds and local ComfyUI setups.
Diffusion Models
Also known as: Latent Diffusion · Flow Matching · Denoising Diffusion
Plain English
The dominant family of generative models for images and video. They learn to create realistic visuals by starting from pure random noise and gradually sculpting it into a coherent picture over multiple steps.
Technical Overview
Diffusion and rectified flow models learn the reverse of a gradual noising process—predicting noise or velocity fields conditioned on text embeddings and timesteps. Modern systems run diffusion inside a compressed VAE latent space using DiT backbones, guided at inference by CFG and Sampling & Schedulers.
Chronological Entries (3)
Modern image and Video Generation Models (Sora, Veo, Kling, Wan) largely shifted from classic DDPM noise prediction to Rectified Flow / Flow Matching, which traces straighter trajectories from noise to data. Combined with Diffusion Transformer (DiT) backbones in models like FLUX, flow matching produces sharper outputs in fewer sampling steps.
Rombach et al. introduced Latent Diffusion Models (LDMs), moving the denoising loop from raw high-resolution pixels into a compact Latent Space & VAE representation. This breakthrough enabled consumer GPUs to generate 512×512+ images and directly spawned Stable Diffusion and Black Forest Labs.