The Role of Registers in Pixel-Space DiT

yresearch · hf · 2026-07-16

This work, **Register Matter for Pixel-Space Diffusion Transformers**, investigates whether register tokens remain useful in pixel-space DiTs. Key points: - ViT register tokens are commonly used to mitigate patch-token outliers; however, the authors found that pixel-space DiTs **do not exhibit the same outliers**. - Despite this, DiTs **still benefit from registers**, and they are even more effective in pixel space than in latent space. - Through intermediate representation analysis, the authors discovered that registers produce cleaner feature maps during high-noise stages, potentially explaining their effectiveness. - The paper proposes **Register Guidance**: amplifying the contribution of register tokens responsible for enhancing visual structure and consistency. - The authors further note that some recent pixel-space DiT architectures already implicitly incorporate register-like mechanisms, which may partly explain their strong performance.

Related event: Register Tokens Prove Effective in Pixel-Space DiT(2 posts)→

Original post →

More from coding & agent

coding & agent channel →