principles of diffusion models

• revolutionized text-to-image synthesis by producing photorealistic, controllable images from textual prompts ex- Stable diffusion, DallE-3, Midjourney
• used to generate 3d molecular structures and protein conformations by modeling spatial distributions of atoms. ex- DiffDock, RFdiffusion
• video diffusion models extend image diffusion temporally learning dynamics and frame consistency. They can generate videos from text or still images. ex- Lumiere
• diffusion models are applied to robot policy learning generating action trajectories from observations rather than pixel-wise noise
• diffusion models generate realistic music and speech maintaining temporal coherence and style - ex stable audio, audio LDM

Evolution in Deep Generative methods:

• EBM (Energy Based Methods)
• VAE (Variational Auto-encoders)
• NF (Normalizing Flows)
• DPM (Diffusion Probabilistic Models)
• NODE (Neural ODE)
• NCSN (Noise Conditional Score Network)
• DDPM (Denoising Diffusion Probabilistic Models)
• Score SDE → FM (Flow Matching)

Core Perspective on Diffusion Models

Ref:

• https://arxiv.org/pdf/2510.21890

Links:

202609030934