Diffusion Transformers (DiT) are the generative image architecture that replaces the traditional U-Net backbone in latent diffusion models with a standard Vision Transformer, unlocking predictable transformer scaling laws for image generation quality and establishing the backbone behind state-of-the-art text-to-image systems.

Why Replace the U-Net?

U-Nets served latent diffusion well but have irregular architectures (encoder/decoder with skip connections) that resist clean scaling analysis. DiT showed that a vanilla ViT — with no skip connections and no convolutional layers — can match and exceed U-Net quality when scaled properly, and that image generation quality improves log-linearly with compute just like language model perplexity.

Architecture Details

• Patchification: The latent representation from a pretrained VAE encoder is divided into non-overlapping patches (typically 2x2 in latent space), each projected into a transformer token.
• Conditioning via adaLN-Zero: Instead of cross-attention, DiT injects the diffusion timestep embedding and class label through Adaptive Layer Normalization — modulating the scale and shift parameters of each LayerNorm. The "Zero" variant initializes the final modulation to output zeros, making each transformer block initially act as the identity function for training stability.
• No Decoder: The final transformer output is linearly projected back to the latent patch shape and reassembled; the pretrained VAE decoder converts the latent back to pixel space.

Scaling Behavior

Each doubling of compute yields a predictable FID improvement — a property U-Net diffusion models never cleanly demonstrated.

Practical Implications

• Infrastructure Reuse: DiT runs on the exact same FlashAttention, FSDP, and activation checkpointing infrastructure already battle-tested for LLM training. No custom U-Net kernel engineering is needed.
• VAE Quality Ceiling: DiT cannot generate details finer than what the VAE can reconstruct. A blurry or artifact-prone VAE decoder sets a hard floor on visual quality regardless of how large the transformer grows.

Diffusion Transformers are the architecture that unified language and vision scaling laws — proving that the same transformer recipe that conquered text also governs the predictable improvement of visual generation quality with compute.