What's this about?
The paper shows that register tokens — a known fix for Vision Transformers — also matter for diffusion transformers that generate images directly in pixel space, cleaning up high-norm token outliers that degrade feature quality. (HF 16↑)
What exactly is a "register token"?
Vision Transformers develop weird high-magnitude tokens that act as internal scratch space, polluting the feature map. Registers are extra blank tokens added to the input purely to give the model somewhere to park that computation — like adding a junk drawer so the kitchen counters stay clean. This work carries the trick into pixel-space image generation.
Why does it matter?
Pixel-space diffusion (skipping the usual latent-space compression step) is a candidate for simpler, higher-fidelity image generators; small architectural hygiene like this is what makes it viable.