LiFT: Loop Flow Transformers for image generation with flow matching
Radar: LiFT trains an image generator with flow matching and a shared transformer core, where each loop progressively refines the initial velocity estimate. According to the author, a model trained with 2 loops improves with up to 16 loops at inference.
According to the author's description, LiFT (Loop Flow Transformers) trains an image generator with flow matching using a shared transformer core. In each loop, the initial velocity estimate is progressively refined. The author states that a model trained with 2 loops improves with up to 16 loops at inference.
The work also shows how to trade inference loops for quality without retraining, which may be useful for DiTs with lower computational cost. This Radar summary does not detail metrics, datasets or comparisons. To verify, read the original text, check the reported results, and reproduce them in your own environment before adopting the technique.