TL;DR: A new DiT variant loops shared transformer blocks within each denoising step, letting builders trade inference compute for image quality at sampling time.
Summary: Researchers released a Looped Diffusion Transformer that reuses a shared group of transformer blocks multiple times per denoising step instead of scaling parameters. Deep supervision on every loop and self-modulating attention make naive looping effective; a 260M looped model reportedly beats a 6.5x larger model with 4.9x less inference compute. The MIT-licensed B/16 weights allow loop depth to be chosen at sampling without retraining, with an M5 Pro test showing roughly 23–24 seconds per image at depth 1 versus 43–47 seconds at depth 4.
Why it matters: Builders can tune image quality against latency per request by adjusting loop depth, rather than retraining or switching models. Try the open weights to benchmark depth/quality tradeoffs on your own prompts.
Source: reddit