How AI Actually Makes Images Diffusion

5 Просмотры

Спасибо! Поделитесь с друзьями!

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

Издатель Jun 14, 2026

AI image generators feel like magic — type a sentence, get a photo. But Stable Diffusion isn't a painter, and it isn't pulling pictures from a database. Under the hood, it's a machine trained to do one weird thing: remove noise.

In this video we open the black box and draw the whole pipeline, end to end: prompt → CLIP text encoder → random noise → a U-Net (or, in newer models, a DiT transformer) that denoises 20–50 times → a VAE decoder → your image. We explain why it's called "latent" diffusion (the trick that made it run on a home GPU instead of a datacenter), the real parameter counts (SD ~983M → SDXL 3.5B → SD 3.5 Large 8B → FLUX.2 32B), the dataset it learned from (LAION-5B, ~5.85 billion image-text pairs), and how the 2026 tools — Stable Diffusion / FLUX, Midjourney V8, DALL·E/GPT Image, and Google's Nano Banana — actually differ.

Then we get concrete about where this actually lives in 2026 — Adobe Firefly (24B+ images generated), Canva Magic Studio (10B+ AI uses), and the 86% of creators who now use generative AI — across design, ads, e-commerce shots, game art, stock, archviz, and film pre-viz. And we look at what's next: real-time and on-device models, consistent characters, 3D & world-models, and the regulation wave (the EU AI Act's Article 50 + C2PA Content Credentials) — plus the new bottleneck, which isn't quality anymore but provenance and trust.

We also do the honest part the hype skips: these models learned from billions of images scraped without consent, the Getty v. Stability AI lawsuit, the Nov 2025 UK ruling, and why "legal" and "settled" aren't the same thing.

Every number is cross-checked against ≥2 sources (the DDPM and CFG papers, LAION's paper, NVIDIA's docs, Wikipedia, Black Forest Labs, and court filings). Where the hype outruns reality, we say so.

⏱ CHAPTERS
0:00 The $600k machine that made images free
0:52 What generative AI actually is
1:38 GANs vs VAEs vs Diffusion vs Autoregressive
2:35 The core trick: learning to undo noise
3:34 The full pipeline: prompt → pixels
4:34 Why it's "latent" (and why it runs at home)
5:31 The real numbers: params, steps, data
6:21 Five tools, five philosophies (SD/FLUX · Midjourney · GPT Image · Nano Banana)
7:18 Where it's used: inside the apps you already open
8:12 What's next + the bottleneck (real-time, on-device, EU AI Act)
9:12 The catch: images nobody licensed
10:07 The moat: math is free, data isn't
10:54 Outro