Looped Diffusion Transformer: Reusing Blocks Boosts Efficiency

Looped Diffusion Transformer (Looped DiT) runs shared transformer blocks multiple times within each denoising step, instead of scaling model size. The authors fixed naive looping with deep supervision and self-modulating attention. A 260M looped model outperforms models 6.5x larger on several benchmarks with 4.9x less inference compute; at a fixed budget, more loop depth helps more than more denoising steps. User tests show depths 1 and 4 take ~23-24s and ~43-47s per image, with depth adjustable at sampling time.