FULL STORY

Looped-DiT: Recycled Transformer Blocks Beat Bigger Models

Looped-DiT from SenseTime, Tsinghua and NTU recycles the same transformer blocks across denoising steps instead of stacking deeper networks, letting a small text-to-image model outperform models 6.5x its size.

2026-10-01 ~ 2026-10-04 · 2 episodes · 6 posts

Episode 1 · Looped-DiT: Small Model Beats 6.5x Larger Ones by Reusing Transformer Blocks (2026-10-01, 2 posts)

A new paper, Looped Diffusion Transformer (Looped-DiT), repeatedly reuses a shared Transformer block at each denoising step instead of stacking deeper layers. A 260M-parameter model outperforms one 6.5x larger while using 4.9x less compute.

Episode 2 · Looped-DiT: Recycled Transformer Layers Beat Models 6.5x Larger (2026-10-03, 4 posts)

Researchers from SenseTime, Tsinghua and NTU propose Looped-DiT, the first looped Transformer for text-to-image diffusion. With only 260M parameters, reusing shared Transformer blocks within each denoising step beats models 6.5x larger while cutting inference compute by 4.9x.