Preprint2026
UltraViT: Latency-Optimized On-device Vision Encoder for Large Vision-Language Models
Ioannis Maniadis Metaxas, Adrian Bulat, Alberto Baldrati, et al.
UltraViT is a latency-optimized vision encoder for LVLMs using a pyramidal architecture with heterogeneous spatial mixers and a two-stage generative pre-training strategy, achieving 1.7x speedup on-device.
0Jul 25, 2026MultimodalAlignment
arXiv