Search papers, labs, and topics across Lattice.
Australian National University
2
0
5
Method improvements in diffusion transformers may not be reflected in both ImageNet and text-to-image tasks, revealing a critical gap in current evaluation practices.
LeapAlign unlocks efficient fine-tuning of flow matching models by collapsing long generation trajectories into two-step leaps, enabling direct gradient propagation to early generation steps without prohibitive costs.
