BLIP3o-NEXT: Next Frontier of Native Image Generation
* Equal contribution
An advance in native multimodal generation that improves visual fidelity, prompt alignment, and unified image understanding and creation.
Zhiyang Xu
I am a Senior Research Scientist at the MBZUAI Institute of Foundation Models (IFM) in Silicon Valley, working on video world model pretraining. I earned my Ph.D. from Virginia Tech, advised by Prof. Lifu Huang, with a dissertation titled Towards Unified and Generalizable Multimodal Foundation Models, after completing my B.S. at Rensselaer Polytechnic Institute. Previously, I interned at Meta GenAI and Apple AIML.
Selected research
Projects
Selected work in multimodal pretraining, understanding, and generation.
Vision-Flan: Scaling Human-Labeled Tasks in Visual Instruction Tuning
A large-scale visual instruction-tuning effort built from diverse human-labeled tasks to improve general multimodal instruction following.
MultiInstruct: Improving Multi-Modal Zero-Shot Learning via Instruction Tuning
* Equal contribution
Outstanding Paper Award
A multimodal instruction-tuning benchmark and training framework that improves zero-shot transfer across diverse vision-language tasks.