#diffusion

Iterative denoising / score-based generation.

16 chapters, oldest era first: one idea traced across the books.

From Noise to Image 2015–2020 Ch. 8: Diffusion Foundations Forward noising, the ELBO reduced to a simple denoising loss, and score functions — the math under everything modern.
From Noise to Image 2020–2021 Ch. 9: Scores, SDEs, and Faster Sampling The SDE/ODE unification, probability flow, and DDIM's deterministic shortcut.
From Frames to Worlds 2020–2023 Ch. 17: The Third Dimension NeRF, 3D from 2D priors, and Gaussian splatting — where 3D meets video and worlds.
From Noise to Image 2021–2022 Ch. 10: Guidance Classifier guidance, the classifier-free guidance formula, and the dial between fidelity and diversity.
From Noise to Image 2021 Ch. 12: DALL-E and GLIDE Discrete-token text-to-image versus text-conditioned diffusion — the fork in the road.
From Noise to Image 2021–2022 Ch. 13: Latent Diffusion and Stable Diffusion Perceptual compression, cross-attention conditioning, and the open-weights detonation of August 2022.
From Noise to Image 2021–2023 Ch. 17: Editing SDEdit, attention-map surgery, instruction-based editing, and inpainting.
From Noise to Image 2022 Ch. 14: The Year Images Went Mainstream DALL-E 2, Imagen, Parti, and Midjourney: unCLIP, giant text encoders, cascades, and autoregression at scale.
From Noise to Image 2022–2024 Ch. 19: Flow Matching and Transformer Backbones Velocity-field regression, rectified flow, and DiT replacing the U-Net — the SD3/FLUX lineage.
From Frames to Worlds 2022 Ch. 5: Video Diffusion Models Factorized spatio-temporal attention, cascades, and joint image–video training.
From Noise to Image 2023 Ch. 15: Control ControlNet's zero-convolution trick, adapters, and reference images — from prompting to specification.
From Frames to Worlds 2023 Ch. 6: Latent Video Diffusion Temporal layers bolted onto image LDMs, AnimateDiff, and Stable Video Diffusion's open weights.
From Frames to Worlds 2023 Ch. 7: Video-to-Video and Editing Structure-preserving synthesis — Runway's Gen-1 becomes a product before the model race begins.
From Frames to Worlds 2024 Ch. 9: Sora and Spacetime Patches Diffusion transformers for video, spacetime patches, and the video-models-as-world-simulators claim.
From Frames to Worlds 2024–2025 Ch. 11: Open Video Models The Stable-Diffusion moment for video: CogVideoX, HunyuanVideo, Mochi, and Wan.
From Frames to Worlds 2024 Ch. 12: Long Video and Consistency Character and scene persistence, AR–diffusion hybrids, and streaming generation.