Xu Luo

I am currently pursuing a Ph.D. at UESTC, supervised by Prof. Jingkuan Song. My primary research interest lies in advancing the state-of-the-art for generative AI models, including diffusion models and (multimodal) large language models (LLMs). Beyond these, I am fervently committed to actualizing the potential of these models, shaping them into autonomous agents capable of performing intricate tasks in real-world scenarios.

Email  /  CV  /  Google Scholar  /  Github

[2024/09] Two papers accepted to NeurIPS 2024.
[2024/05] Thrilled to introduce our latest work, Lumina-T2X, a text-to-any-modality generation framework! Our models can generate high-resolution images & 720p videos of any aspect ratio, 3D multiview images, and audio conditioned on text.
[2023/10] Regret to say that FSL is no longer a valuable research topic. My research focus has shifted from Few-Shot Learning (FSL) to the burgeoning field of generative AI. I am currently an intern at Shanghai AI Lab.
[2023/04] Our systematic study on few-shot learning has been accepted at ICML'23! This project demanded a significant investment of time and revealed numerous unexpected findings. Check it out now!
[2022/05] Our paper, which uncovers fundamental problems inherent in few-shot learning, was accepted at ICML 2022.
[2021/09] My first top conference paper at NeurIPS 2021! The paper is not that good, but a good experience(:

Text-to-any-modality models that generate images, videos, audio, and 3D multiview images conditioned on text in a flow-based diffusion framework, using novel Flag-DiT architectures with up to 5B parameters and 128K context windows.

Uncovering and analyzing extreme feature redundancy phenomenon of pretrained vision models when transferring to few-shot tasks.

Proposing DETA--a framework that solves potential data/label noise in downstream few-shot transfer tasks.

Empirically proving the disentanglement of training and adaptation algorithms in few-shot classification, and performing interesting analysis of each phase that leads to the discovery of several important observations.

Revealing a strong bias caused by the centroid of features in each few-shot learning task. A simple method is designed to rectify this bias by removing the dimension along the direction of task centroid from the feature space.

Revealing and analyzing the channel bias problem that we found critical in few-shot learning, through a simple channel-wise feature transformation applied only at test time.

Identifying image background as a shortcut knowledge ungeneralizable beyond training categories in Few-Shot Learning. A novel framework, COSOC, is designed to tackle this problem.

Applying contrastive learning to Few-Shot Learning, with views generated in a learning-to-learn fashion.

Conference reviewer

This well-designed template is borrowed from this guy