Looped Diffusion Transformer
Looped-DiT reuses shared Transformer blocks within each denoising step, letting a 260M-parameter text-to-image model beat one 6.5x larger at 4.9x lower compute.
Looped Diffusion Transformer scales compute by repeatedly running shared transformer blocks within each denoising step, increasing depth while keeping parameter count fixed. The authors stabilize naive looping with deep supervision across intermediate loops and self-modulating attention. A 260M-parameter looped model surpasses a 6.5x larger model across multiple text-to-image benchmarks while requiring 4.9x lower inference compute, and deeper loops yield larger gains than adding denoising steps under a fixed budget.
- Looped computation increases depth without adding parameters.
- Deep supervision and self-modulating attention fix naive looping failures.
- 260M looped model beats a 6.5x larger model at 4.9x less inference compute.
- Deeper loops correct earlier mistakes, suggesting latent reasoning behavior.
From the original article
1 min read at arxiv.orgImproving text-to-image models has traditionally relied on increasing model size or the number of denoising steps. In this work, we explore an alternative way to scale computation by repeatedly running shared Transformer blocks within each denoising step, effectively increasing computational depth while keeping the parameter count fixed.
By Yong Xien Chng · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.