LiFT repeatedly applies one shared core, increasing effective depth without adding parameters. Each loop is trained toward a different point on a straight path from the initial velocity estimate to the flow-matching target, rather than requiring every loop to predict the final target.
This supervision lets inference use substantially more loops than training. An XL/2 model trained with two loops improved from 17.07 FID at two loops to 9.31 at sixteen. On ImageNet at 256×256, LiFT-L/2 achieved an FID 3.34 points below the dense DiT-XL/2 baseline while using about 60% fewer parameters, 32% fewer training FLOPs, and 52% fewer inference FLOPs.
The result introduces recurrent depth as a second inference-compute control alongside integration steps. Its value is scale-dependent, however: the B/2 dense model remained stronger at every evaluated cost.