Diffusion Transformer

A Diffusion Transformer is a neural network that draws a finished image or video by removing noise from a blurry, static-filled screen little by little, over many steps. It's a bit like wiping a foggy window with your hand again and again, slowly revealing the scenery outside — not in one wipe, but through repeated passes that make the picture clearer each time.

The name combines two ideas. "Diffusion" refers to the step-by-step process of removing noise, while "Transformer" refers to the architecture that figures out how each part of an image or sentence relates to the others. Put together, running the noise-removal process through a Transformer structure lets the model generate not just images, but also video and audio.

The catch is that more repetitions mean more time to produce a result. That's why a wave of acceleration techniques has emerged recently, cutting the noise-removal process down from dozens of steps to as few as 4 or 5 while still preserving quality.

An article explains that "video generation models typically start from a screen full of noise and run a Diffusion Transformer (DiT) repeatedly to produce the finished video." One thing that's easy to get wrong: DiT isn't a specific company's product name — it's a general term for a neural network architecture shared across many video and image generation models, such as MiniMax H3.