Research3 min read
Mean Mode Screaming: Researchers solve collapse in 1000-layer Diffusion Transformers
Tags AI · Research
arXiv·
Researchers identified and solved Mean Mode Screaming (MMS), a structural vulnerability causing deep Diffusion Transformers to collapse into mean-dominated states. The proposed Mean-Variance Split (MV-Split) Residuals enable stable training of a 1000-layer DiT. The work addresses a fundamental barrier to scaling diffusion model depth.