Hang Li

I am a Research Scientist at Meta AI in London, working on generative AI. Before joining Meta, I conducted my doctoral research in Computer Science at the University of Munich, under the supervision of Volker Tresp, with industry collaboration at Siemens AG. I earned my M.Sc. from the Technical University of Munich, where I was advised by Volker Tresp and Stephan Günnemann.

Email  /  Google Scholar  /  Github  /  Twitter  /  LinkedIn

My research aims to build generative models that people can steer at scale: systems whose outputs faithfully reflect what was asked of them, and whose internals can be understood and aligned. This work spans the diffusion and video foundation models themselves, the multimodal representations that ground them, and the interpretable mechanisms that make them accountable.

A zero-shot reference-to-video generation framework trained only on video-text pairs, outperforming methods trained with explicit reference-video-text triplets on OpenS2V-Eval.

A model-agnostic attention regularization that guides target queries to the correct reference region, reducing fine-grained texture distortion and achieving state-of-the-art virtual try-on and pose transfer.

Previous work interprets vectors in an interpretable latent space of diffusion models as semantic concepts. However, existing approaches cannot discover directions for arbitrary concepts, such as those related to inappropriate concepts. In this work, we propose a novel self-supervised approach to find interpretable latent directions for a given concept. With the discovered vectors, we further propose a simple approach to mitigate inappropriate generation.

We explore two types of large-scale multimodal generative models, image-to-text and text-to-image. The image-to-text model generates abstract descriptions of an image, whereas the text-to-image model decodes the text into low-level visual pixel features. These two models are closely related but their relationship is little understood. In this work, we study if large multimodal generative models understand each other. Specifically, if Flamingo describes an image in text, can DALLE reconstruct an image similar to the input image from the text?

Leveraging generative hard negative samples, we significantly enhance VLMs' performance in tasks involving multimodal compositional reasoning.

In this paper, we take inspiration from attributes of the brain, to develop a computational framework to find the optimal low cost path between a source node and a destination node in a generalized graph.

We present a unified computational theory of an agent’s perception and memory. Episodic memory and semantic memory evolved as emergent properties in a development to gain a deeper understanding of sensory information, to provide a context, and to provide a sense of the current state of the world.

We find that Graphhopper outperforms state-of-the-art scene graph reasoning model on both manually curated and automatically generated scene graphs by a significant margin.

We propose a novel method that approaches the VQA task by performing context-driven, sequential reasoning based on the objects and their semantic and spatial relationships present in the scene.

Last updated: May 2026
Website template from Jon Barron