Behind most of the striking image and video models of 2026 sits a single architectural choice: the Diffusion Transformer, or DiT. It became the production default not because it was novel, but because it scaled better than the alternatives — more quality per training dollar, and a cleaner path to bigger models.
Diffusion, meet the transformer
Diffusion models generate by starting from noise and denoising step by step toward a coherent sample. Early diffusion models used convolutional backbones. DiT swaps that for a transformer — the same architecture that scaled language models — to do the denoising. The payoff is the transformer's well-understood scaling behavior: pour in more data and compute, and quality improves predictably.
The lesson of the last decade repeats: when in doubt, use a transformer and scale it. Generative media just learned it too.
Why it matters for cost
Predictable scaling is an economic advantage, not just a research one. It means labs can plan investments — know that a bigger model trained on more data will pay off — rather than gamble on a finicky architecture. That reliability is why DiT dominates production: it turns "make it better" into "make it bigger," which is a problem money and engineering can solve.
The broader convergence
DiT is one more example of the transformer swallowing a field. Language, vision, audio, and now generative media increasingly run on variations of the same backbone. That convergence has a practical upside: advances in training, serving, and hardware for transformers transfer across all of them. The architecture that generates your text and the one that generates your video are, more and more, the same idea.
