ZeroShotMind

Series

Diffusion Models: From Noise to Image

Score matching, DDPM, latent diffusion, DiT, and flow matching — the full arc of how diffusion models became the dominant generative architecture.

Fundamentals
diffusiongenerative-modelsimage-generation
  1. 1

    Score Matching: The Math Behind Diffusion

    Diffusion models learn to reverse a noise process. The key insight is that you don't need to know the data distribution — you only need to learn its score function, the gradient of the log-density.

    2024-06-19

  2. 1

    Score Matching: The Math Behind Diffusion

    Diffusion models learn to reverse a noise process. The key insight is that you don't need to know the data distribution — you only need to learn its score function, the gradient of the log-density.

    2024-06-19

  3. 2

    DDPM: The Diffusion Process, Forward and Reverse

    DDPM defines a fixed forward process that gradually destroys an image into noise, then trains a neural network to reverse it. The math is tractable because each step is Gaussian.

    2024-06-19

  4. 2

    DDPM: The Diffusion Process, Forward and Reverse

    DDPM defines a fixed forward process that gradually destroys an image into noise, then trains a neural network to reverse it. The math is tractable because each step is Gaussian.

    2024-06-19

  5. 3

    DDIM: Deterministic Sampling in Fewer Steps

    DDPM needs 1000 steps to generate a sample. DDIM reframes the reverse process as an ODE and gets the same quality in 50. The model weights are identical — only the sampling procedure changes.

    2024-06-19

  6. 3

    DDIM: Deterministic Sampling in Fewer Steps

    DDPM needs 1000 steps to generate a sample. DDIM reframes the reverse process as an ODE and gets the same quality in 50. The model weights are identical — only the sampling procedure changes.

    2024-06-19

  7. 4

    Latent Diffusion: Why You'd Compress Before You Denoise

    Running DDPM in pixel space at 512×512 is expensive. Latent diffusion compresses the image into a small latent space first, runs the diffusion process there, and decodes back. The same quality, a fraction of the compute.

    2024-06-19

  8. 4

    Latent Diffusion: Why You'd Compress Before You Denoise

    Running DDPM in pixel space at 512×512 is expensive. Latent diffusion compresses the image into a small latent space first, runs the diffusion process there, and decodes back. The same quality, a fraction of the compute.

    2024-06-19

  9. 5

    Classifier-Free Guidance: Steering Diffusion with a Signal

    Conditioning a diffusion model on text gives you text-to-image generation. Classifier-free guidance makes that conditioning much stronger — at the cost of some diversity.

    2024-06-19

  10. 5

    Classifier-Free Guidance: Steering Diffusion with a Signal

    Conditioning a diffusion model on text gives you text-to-image generation. Classifier-free guidance makes that conditioning much stronger — at the cost of some diversity.

    2024-06-19

  11. 6

    DiT: Replacing the U-Net with a Transformer

    DDPM, DDIM, and latent diffusion all use a U-Net backbone. DiT replaces it with a transformer — and finds that diffusion scales with model size the same way language models do.

    2024-06-19

  12. 6

    DiT: Replacing the U-Net with a Transformer

    DDPM, DDIM, and latent diffusion all use a U-Net backbone. DiT replaces it with a transformer — and finds that diffusion scales with model size the same way language models do.

    2024-06-19

  13. 7

    Flow Matching: The Cleaner Generalization of Diffusion

    Diffusion models learn to reverse a specific noise process. Flow matching learns to transport any source distribution to any target distribution along straight paths — simpler math, faster sampling, and better training signal.

    2024-06-19

  14. 7

    Flow Matching: The Cleaner Generalization of Diffusion

    Diffusion models learn to reverse a specific noise process. Flow matching learns to transport any source distribution to any target distribution along straight paths — simpler math, faster sampling, and better training signal.

    2024-06-19