Search papers, labs, and topics across Lattice.
2
0
5
Ditch the manual segmentation masks: ConceptPrism automatically disentangles concepts in personalized diffusion models by comparing images, achieving better fidelity and alignment.
Block diffusion LLMs can achieve near-lossless long-context performance with 3-4x speedups by using attention patterns learned from the first denoising step, unlocking efficient sparse attention without autoregressive approximations.
