Diffusion, Denoised: How AI Turns Static Into Masterpieces

2 Просмотры

Спасибо! Поделитесь с друзьями!

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

Издатель Sep 23, 2026

The GAN won sharpness with a fragile two-player game. This episode fires the critic. Diffusion models take a clean image and destroy it with Gaussian noise, step by step, until only static remains. Then they train one network to reverse that process: predict the noise inside a corrupted image, remove a little, repeat. Generation is not a single forward pass, it is iterative refinement, and it is the pipeline behind today's text-to-image systems.

Subscribe to @neurals_ca for the full series: https://neurals.ca

What you will learn:
- The intuition: corrupt data to every noise level t, train a network to remove a LITTLE noise, then apply it many times to turn pure static into a clean sample
- The forward process: q(x_t | x_{t-1}) as a fixed noising chain, and the closed form that jumps straight to any t (x_t = root-alpha-bar x_0 + root(1-alpha-bar) eps)
- The reverse process: the true denoising chain is intractable, so a network learns p_theta(x_{t-1} | x_t) by predicting the noise
- Why diffusion training is a plain MSE loss (stable) while GAN training is a minimax duel (unstable, mode collapse), and the price diffusion pays: slow, many-step sampling
- The machinery of modern image AI: U-Net denoisers conditioned on t, latent diffusion (diffuse in the autoencoder's latent space), and text-to-image conditioning with T5/CLIP embeddings through a Diffusion Transformer
- Where it still falls short: sampling cost and trustworthiness

Chapters:
0:00 Intro
0:15 Firing the critic: from duel to denoising
1:05 Destroy the data: the noising conveyor
1:55 The closed-form shortcut
2:45 The reversal: predict the noise
3:35 Training with MSE: stable by design
4:00 Sampling: walk backward from static
4:25 The price: slow, many-step sampling
4:44 The U-Net denoiser
5:05 Latent diffusion: the autoencoder returns
5:29 Text conditioning: prompts steer the denoiser
6:00 Limits: sampling cost and trustworthiness
6:24 What is next: learning without labels

This episode comes from MIA5150 (Generative AI and LLMs) at the University of Ottawa, visualized by neurals.ca.

Links:
neurals.ca: https://neurals.ca
X / Twitter: https://x.com/neurals_ca
AI Algorithms (playlist): https://www.youtube.com/playlist?list=PLbhLSYkUZeAg

If this was useful, subscribe to @neurals_ca. Next: self-supervised and multimodal learning.

#AI #MachineLearning #AIAlgorithms #DiffusionModels #GenerativeAI #DeepLearning #StableDiffusion #TextToImage #MIA5150 #uOttawa #neuralsca