Search papers, labs, and topics across Lattice.
Affiliation:
4
0
4
16
OmniUE achieves an unprecedented 83.7% improvement in visual-interactive benchmarks, setting a new standard for multimodal embedding performance.
Frequency-dependent discrepancies in diffusion models reveal that fixed correction rules fail, but a lightweight spectral alignment method can effectively reduce exposure bias.
MMHNet proves you can train a video-to-audio model on short clips and have it generalize to generate coherent audio for videos over 5 minutes long.
