diffusion transformer
**Diffusion Transformer (DiT)** is the **architecture that replaces the traditional U-Net backbone in diffusion models with a pure Transformer design** — using self-attention over patched latent representations to generate images, video, and other media with superior scaling properties compared to convolutional U-Nets, where scaling model size and compute directly improves generation quality following predictable scaling laws, making DiT the architecture behind state-of-the-art systems like DALL-E 3, Stable Diffusion 3, and Sora.
**Why Replace U-Net with Transformers**
- Traditional diffusion (DDPM, Stable Diffusion 1/2): U-Net with conv layers + cross-attention.
- U-Net limitations: Fixed spatial structure, hard to scale beyond ~2B parameters, convolution is local.
- Transformers: Scale smoothly from millions to hundreds of billions of parameters.
- DiT insight: Treat image patches as tokens → apply standard Transformer → better scaling.
**DiT Architecture**
```
Input latent z (e.g., 32×32×4 from VAE)
↓
[Patchify]: Split into p×p patches → sequence of tokens
↓
[Positional embedding + timestep embedding]
↓
[DiT Block 1]: LayerNorm → Self-Attention → MLP (with adaptive conditioning)
[DiT Block 2]: ... (repeated N times)
...
[DiT Block N]
↓
[Unpatchify]: Reconstruct spatial dimensions
↓
Predicted noise ε (or velocity v)
```
**Adaptive Layer Norm (adaLN-Zero)**
- Standard transformers: LayerNorm has fixed learnable scale/shift.
- DiT: Scale and shift parameters are **predicted** from timestep and class label.
- adaLN-Zero: Initialize the final layer to predict zeros → model starts as identity → stable training.
- This is the key conditioning mechanism — how DiT tells the network what timestep and what class to generate.
**Scaling Properties**
| Model | Parameters | FID-50K (ImageNet 256) |
|-------|-----------|------------------------|
| DiT-S/2 | 33M | 68.4 |
| DiT-B/2 | 130M | 43.5 |
| DiT-L/2 | 458M | 23.3 |
| DiT-XL/2 | 675M | 9.62 |
| DiT-XL/2 + cfg | 675M | 2.27 |
- Clear log-linear scaling: Doubling parameters consistently improves FID.
- U-Net scaling: Plateaus around ~1B parameters (architecture bottleneck).
**DiT in Practice**
| System | Architecture | Scale |
|--------|-------------|-------|
| Stable Diffusion 3 (Stability AI) | MM-DiT (multimodal DiT) | ~3B |
| DALL-E 3 (OpenAI) | DiT variant | ~12B (estimated) |
| Sora (OpenAI) | Spacetime DiT | Unknown (large) |
| PixArt-α/Σ | DiT with T5 text encoder | 600M |
| Flux (Black Forest Labs) | DiT variant | ~12B |
**DiT vs. U-Net**
| Property | U-Net | DiT |
|----------|-------|-----|
| Architecture | Conv + attention | Pure transformer |
| Scaling | Saturates ~2B | Scales to 100B+ |
| Training efficiency | Good at small scale | Better at large scale |
| Spatial inductive bias | Strong (convolution) | Weak (learned) |
| Hardware utilization | Mixed ops | Uniform matmul → GPU-optimal |
The Diffusion Transformer is **the architectural evolution that enabled diffusion models to scale into the frontier generative AI era** — by replacing the U-Net's convolutional backbone with Transformers, DiT unlocked the same scaling laws that made LLMs powerful, allowing image and video generation models to improve predictably with more compute and data, making it the standard architecture for all major generative AI systems from 2024 onward.