How AI Paints From Pure Noise: Diffusion Explained (series finale)

5 Просмотры

Спасибо! Поделитесь с друзьями!

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

Издатель Jul 18, 2026

That photo is fake — no camera, no artist, nothing. An AI conjured it out of pure random static. How is that even possible? The answer is diffusion, and it's the finale of our "How LLMs Actually Work" series.

Here's the twist: to learn how to CREATE images, the AI first learned how to destroy them. Take a photo, add a speck of noise, then more, then more, until it's pure television static. The AI trains on millions of photos noised step by step, learning to guess the noise at every stage. Then to generate, you start from PURE static, whisper "a cat," and the AI peels off a little noise, guided by those words — again and again — until a cat that never existed sharpens into view. It never draws a thing. It only ever answers one tiny question, thousands of times: what noise do I subtract next? Creation, as sculpture.

Chapters:
0:00 A photo conjured from static
0:32 You can't just say "draw a cat"
0:50 Training: learn to destroy
1:10 Generating: carve a cat from noise
1:25 Every image & video generator
1:35 Never drawn — carved from noise
1:43 The whole series, wrapped

Play with every idea in the series, free, at https://vibeengines.com

Vibe Engines makes intuitive, visual explainers for AI and computer science. This is #7 — the FINALE — of the "How LLMs Actually Work" series: tokens, attention, embeddings, experts, alignment, and noise.

#diffusion #stablediffusion #dalle #midjourney #aiart #howaiworks #llm #ai2026 #machinelearning #generativeai