Welcome to Junjie Fei’s Homepage!

I am a third-year Ph.D. student in Computer Science at King Abdullah University of Science and Technology (KAUST), advised by Prof. Mohamed Elhoseiny. My research studies long-context intelligence in multimodal systems, with a focus on long videos. I view long context not only as a context-window scaling problem, but fundamentally as a memory problem: determining which visual evidence to compress, retain, and retrieve as tasks and interactions unfold. My broader work spans vision-language generation and grounding.

I previously worked as a Research Scientist Intern at Meta AI, mentored by Dr. Chenchen Zhu, where I developed SVLM-based compressors for efficient long video modeling.

Before starting my Ph.D., I received my B.Eng. and M.Eng. degrees from Chongqing University and Xiamen University, respectively. I also worked as a Research Assistant and Visiting Scholar at the SUSTech VIP Lab and as a Visiting Scholar at KAUST Vision CAIR. Please see my CV for more details.

💡 Open to collaboration: I am currently seeking Summer 2027 research internship opportunities and welcome discussions and collaborations in multimodal learning. Feel free to contact me at junjiefei@outlook.com or junjie.fei@kaust.edu.sa.

News

[2026/06] Tempo has been accepted to ECCV 2026! 🎉
[2026/04] Project Tempo from my Meta AI internship is publicly released!
[2025/09] One paper has been accepted by NeurIPS 2025!
[2025/09] Joined Meta AI as a Research Scientist Intern!
[2025/06] Two papers have been accepted by ICCV 2025!
[2025/02] One paper has been accepted by CVPR 2025!
[2024/08] Joined KAUST as a PhD student!
[2023/07] One paper has been accepted by ICCV 2023!
[2023/04] Project Caption Anything is publicly released!

Experience

	Meta AI Research Scientist Intern \| Sep. 2025 - Feb. 2026
	Vision CAIR Research Group, KAUST Visiting Scholar \| Jan. 2024 - May 2024
	VIP (Visual Intelligence & Perception) Lab, SUSTech Visiting Scholar / Research Assistant \| Oct. 2022 - Jan. 2024

Research

(* equal contribution)

	Small Vision-Language Models are Smart Compressors for Long Video Understanding Junjie Fei, Jun Chen, Zechun Liu, Yunyang Xiong, Chong Zhou, Wei Wen, Junlin Han, Mingchen Zhuge, Saksham Suri, Qi Qian, Shuming Liu, Lemeng Wu, Raghuraman Krishnamoorthi, Vikas Chandra, Mohamed Elhoseiny, Chenchen Zhu ECCV, 2026 project / code / paper / demo Tempo is a highly efficient, query-aware framework that leverages a Small Vision-Language Model (SVLM) as an intelligent temporal compressor. It adaptively distills long-form video content into semantic-rich tokens in a single forward pass, achieving SOTA performance on LVBench while significantly reducing the computational overhead of processing hour-long videos.
	Kestrel: 3D Multimodal LLM for Part-Aware Grounded Description Mahmoud Ahmed, Junjie Fei, Jian Ding, Eslam Mohamed Bakr, Mohamed Elhoseiny ICCV, 2025 project / paper Kestrel is a part-aware point grounding 3D MLLM, capable of comprehending and generating language and locating the position of the object and its materials at the part level.
	WikiAutoGen: Towards Multi-Modal Wikipedia-Style Article Generation Zhongyu Yang, Jun Chen, Dannong Xu, Junjie Fei, Xiaoqian Shen, Liangbing Zhao, Chun-Mei Feng, Mohamed Elhoseiny ICCV, 2025 project / code / paper WikiAutoGen is a novel system for automated multimodal Wikipedia-style article generation, retrieving and integrating relevant images alongside text to enhance both the depth and visual appeal of the generated content.
	Document Haystacks: Vision-Language Reasoning Over Piles of 1000+ Documents Jun Chen, Dannong Xu, Junjie Fei, Chun-Mei Feng, Mohamed Elhoseiny CVPR*, 2025 code / paper / benchmark The Document Haystack Benchmarks aim to evaluate the performance of VLMs on large-scale visual document retrieval and understanding.
	Transferable Decoding with Visual Entities for Zero‑Shot Image Captioning Junjie Fei, Teng Wang, Jinrui Zhang, Zhenyu He, Chengjie Wang, Feng Zheng ICCV, 2023 code / paper Improving the transferability of zero-shot captioning for out-of-domain images by addressing the modality bias and object hallucination that arise when adapting pre-trained vision-language models and large language models.
	Caption Anything: Interactive Image Description with Diverse Multimodal Controls Teng Wang, Jinrui Zhang, Junjie Fei, Hao Zheng, Yunlong Tang, Zhe Li, Mingqi Gao, Shanshan Zhao arXiv*, 2023 code / technical report / demo Caption Anything is an interactive image‑to‑text generative tool that can generate diverse descriptions for any user-specified object within an image, providing a variety of language styles and visual controls to cater to diverse user preferences.

	Small Vision-Language Models are Smart Compressors for Long Video Understanding Junjie Fei, Jun Chen, Zechun Liu, Yunyang Xiong, Chong Zhou, Wei Wen, Junlin Han, Mingchen Zhuge, Saksham Suri, Qi Qian, Shuming Liu, Lemeng Wu, Raghuraman Krishnamoorthi, Vikas Chandra, Mohamed Elhoseiny, Chenchen Zhu ECCV, 2026 project / code / paper / demo Tempo is a highly efficient, query-aware framework that leverages a Small Vision-Language Model (SVLM) as an intelligent temporal compressor. It adaptively distills long-form video content into semantic-rich tokens in a single forward pass, achieving SOTA performance on LVBench while significantly reducing the computational overhead of processing hour-long videos.
	Kestrel: 3D Multimodal LLM for Part-Aware Grounded Description Mahmoud Ahmed, Junjie Fei, Jian Ding, Eslam Mohamed Bakr, Mohamed Elhoseiny ICCV, 2025 project / paper Kestrel is a part-aware point grounding 3D MLLM, capable of comprehending and generating language and locating the position of the object and its materials at the part level.
	WikiAutoGen: Towards Multi-Modal Wikipedia-Style Article Generation Zhongyu Yang, Jun Chen, Dannong Xu, Junjie Fei, Xiaoqian Shen, Liangbing Zhao, Chun-Mei Feng, Mohamed Elhoseiny ICCV, 2025 project / code / paper WikiAutoGen is a novel system for automated multimodal Wikipedia-style article generation, retrieving and integrating relevant images alongside text to enhance both the depth and visual appeal of the generated content.
	Document Haystacks: Vision-Language Reasoning Over Piles of 1000+ Documents Jun Chen, Dannong Xu, Junjie Fei, Chun-Mei Feng, Mohamed Elhoseiny CVPR*, 2025 code / paper / benchmark The Document Haystack Benchmarks aim to evaluate the performance of VLMs on large-scale visual document retrieval and understanding.
	Transferable Decoding with Visual Entities for Zero‑Shot Image Captioning Junjie Fei, Teng Wang, Jinrui Zhang, Zhenyu He, Chengjie Wang, Feng Zheng ICCV, 2023 code / paper Improving the transferability of zero-shot captioning for out-of-domain images by addressing the modality bias and object hallucination that arise when adapting pre-trained vision-language models and large language models.
	Caption Anything: Interactive Image Description with Diverse Multimodal Controls Teng Wang, Jinrui Zhang, Junjie Fei, Hao Zheng, Yunlong Tang, Zhe Li, Mingqi Gao, Shanshan Zhao arXiv*, 2023 code / technical report / demo Caption Anything is an interactive image‑to‑text generative tool that can generate diverse descriptions for any user-specified object within an image, providing a variety of language styles and visual controls to cater to diverse user preferences.

Junjie Fei

Welcome to Junjie Fei’s Homepage!

News

Experience

Research

Academic Services

Conference Reviewer

Journal Reviewer