Search papers, labs, and topics across Lattice.
1
0
3
15
Ditching the vision encoder actually *improves* multimodal understanding at scale, proving that pixel embeddings alone can achieve state-of-the-art results in unified multimodal models.
