Search papers, labs, and topics across Lattice.
2
0
4
5
Seemingly innocuous photometric variations between training pairs cripple low-level vision models, but a simple closed-form alignment loss can unlock significant gains across diverse tasks and architectures.
By steering diffusion trajectories through semantically meaningful waypoints, WiT sidesteps the entanglement issues plaguing pixel-space diffusion models, leading to faster training and better image quality.
