LiFT loops part of a diffusion transformer to scale inference depth

LiFT makes part of a diffusion transformer recurrent so it can spend more compute inside each sampling step. The team reports a 3.34 lower ImageNet FID than dense DiT-XL/2, with 60% fewer parameters, 32% fewer training FLOPs and 52% fewer inference FLOPs. A supervisor says it can loop beyond its training depth at inference without retraining, early exits or other changes.