Researchers have investigated the utility of register tokens in Diffusion Transformers (DiTs), drawing parallels to their effectiveness in Vision Transformers (ViTs). While ViTs benefit from register tokens by mitigating high-norm patch-token outliers, DiTs do not exhibit these outliers but still show improvement with registers. The study found that register tokens are more beneficial in pixel-space DiTs compared to latent-space DiTs, leading to cleaner feature maps at higher noise levels. This suggests that recent pixel-space DiT architectures may already incorporate similar mechanisms, and the proposed Register Guidance technique further enhances visual structure and coherence by amplifying register token contributions. AI
IMPACT This research could lead to improved image generation quality and efficiency in diffusion models by refining their internal representations.
RANK_REASON The cluster contains an academic paper detailing novel research findings on AI model architectures. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
