LiFT: Loop Flow Transformers

Loop Flow Transformers (LiFT) scale generative computation by repeatedly applying a shared Diffusion Transformer core through a single regression target indexed by continuous depth. On ImageNet 256x256, LiFT-L/2 achieves an FID 3.34 points lower than DiT-XL/2 while using 60% fewer parameters and 52% fewer inference FLOPs.