Diffusion Models from Scratch: How Noise Becomes Images (DDPM, DDIM, Stable Diffusion)

2 Просмотры

Спасибо! Поделитесь с друзьями!

Вам не понравилось видео. Спасибо за то что поделились своим мнением!

Издатель Oct 3, 2026

Standard text-to-image diffusion sampling, as in Stable Diffusion v1, DALL·E 2 and Imagen, starts from random noise. Image editing can start from a noised input image, and the original DALL·E was autoregressive. In this video we build diffusion models from the ground up and see exactly how a neural network learns to turn noise into images.

We start with the forward process, which slowly destroys an image with Gaussian noise and can jump to any noise level in a single step. Then we reverse the arrow: a time-aware U-Net learns to undo the noise, one small Gaussian step at a time. You will see why the DDPM paper's key trick, predicting the noise instead of the image, reduces training to a simple mean squared error.

Next, a second lens: score matching. We connect noise prediction to the gradient of log-probability, from Hyvärinen's score matching and Vincent's denoising score matching to Langevin dynamics, noise-conditional score networks and the SDE framework.

Then DDPM vs. DDIM: how the same trained network can use far fewer steps with a deterministic sampler. The DDIM paper reported 10-50× wall-clock speedups in its experiments, with a quality, speed trade-off, and demonstrated interpolation and approximate inversion. Finally, latent diffusion, the architecture behind Stable Diffusion v1: an autoencoder represents 512×512×3 images as 64×64×4 latents, 48× fewer tensor values, diffusion runs in latent space, cross-attention connects text, and classifier-free guidance makes the model follow your prompt.

All formulas are shown on screen, with the idea behind each one explained in plain language.