LiFT: Loop Flow Transformers
LiFT gets more image-generation work out of a smaller DiT by looping the same core at inference time.
The paper introduces a looped generative model that repeatedly applies a shared Diffusion Transformer core with light architectural changes. Each recurrent step is trained toward one regression target along a straight path from the initial estimate to the flow-matching target. Because the target is indexed by continuous depth, the model can run past its training depth without retraining or early-exit machinery. On ImageNet 256x256, the authors report LiFT-L/2 beating a dense DiT-XL/2 baseline by 3.34 FID while using fewer parameters and FLOPs. HF Daily Papers' note
The paper introduces a looped generative model that repeatedly applies a shared Diffusion Transformer core with light architectural changes. Each recurrent step is trained toward one regression target along a straight path from the initial estimate to the flow-matching target. Because the target is indexed by continuous depth, the model can run past its training depth without retraining or early-exit machinery. On ImageNet 256x256, the authors report LiFT-L/2 beating a dense DiT-XL/2 baseline by 3.34 FID while using fewer parameters and FLOPs. HF Daily Papers' note
score 5