LiFT: Loop Flow Transformers Cut Inference Compute 52% While Improving FID

Researchers from the University of Amsterdam introduce LiFT, which repeatedly applies a shared DiT core to cut parameters by about 60% and inference compute by 52% while actually improving FID by 3.34 points.

2026-10-06 ~ 2026-10-06 · 2 related posts

• LiFT loops a DiT at inference: beats DiT-XL/2 with 52% less inference compute — cgmsnoek · 2026-10-06
• LiFT: looped DiT cuts params 60% while improving FID by 3.34 on ImageNet — VISLab-Amsterdam · 2026-10-06