Diffusion Transformers & Rectified Flow
The U-Net is not the secret of diffusion. Replace it with a transformer, swap the noise schedule for a straight-line flow, and suddenly you have SD3, FLUX, and every 2026 text-to-image model. Trace the evolution from U-Net DDPM (Lesson 10) to Diffusion Transformer (DiT), MMDiT (SD3), and single+double-stream DiT (FLUX). Explain rectified flow: why a straight-line trajectory between noise and data lets models sample in 20 steps instead of 1000. Implement a tiny DiT block and a rectified-flow training loop, both under 100 lines. Distinguish model variants (SD3, FLUX.1-dev, FLUX.1-schnell, Z-Image, Qwen-Image) by architecture, parameter count, and licensing. Lesson 10 built a DDPM with a U-Net denoiser. That recipe dominated 2020-2023: U-Net + beta schedule + noise-prediction loss. It produced Stable Diffusion 1.5 and 2.1 and DALL-E 2. Every 2026 state-of-the-art text-to-image model has moved past it. Stable Diffusion 3, FLUX, SD4, Z-Image, Qwen-Image, Hunyuan-Image — none use a U-Net. They use Diffusion Transformers (DiT). SD3 and FLUX also swap the DDPM noise schedule for rectified flow, which straightens the path from noise to data and enables 1-4 step inference with consistency or distilled variants. The shift matters because it is the reason diffusion-based image generation became controllable, prompt-accurate (SD3/SD4 solved text rendering), and production-fast. Understanding DiT + rectified flow is understanding the 2026 generative-image stack. DiT (Peebles & Xie,…
Diffusion Transformers & Rectified Flow: The U-Net is not the secret of diffusion. Replace it with a transformer, swap the noise schedule for a straight-line…
This free lesson is part of the AI Engineering from Scratch curriculum. Read the full explanation, run the lesson code, and verify the result in the interactive reader or from the repository source.
Browse the complete course catalog or open this lesson on GitHub.