Haodong Li (李浩东)

PhD Student @ UC San Diego | Research Scientist Intern @ Adobe

I'm a PhD student at UC San Diego, working with Prof. Manmohan Chandraker. I'm also a research scientist intern at Adobe Research, working with Dr. Shaoteng Liu and Dr. Zhe Lin. I'm interested in generative world modeling in both pixels and 3D.

Previously, I got my MPhil degree from HKUST, working with Prof. Ying-Cong Chen. I got my BEng degree from Zhejiang University. I spent a wonderful summer at Tencent Hunyuan as a research scientist intern. Feel free to reach out if you have anything to discuss! 🚀

Email: hal211@ucsd.edu, haodongl@adobe.com
Wechat: lihaodong-2000

Google Scholar  /  CV  /  Github  /  Twitter (X)  /  Linkedin  /  Huggingface

Selected Publications

arXiv 2026
arXiv / Paper / Project Page / Code / Demo / Gallery / Slides

Rolling Sink aims to bridge the gap between limited-horizon training and open-ended testing in autoregressive video diffusion. Built on Self Forcing (trained on only 5s clips), Rolling Sink effectively scales the autoregressive video synthesis to ultra-long durations (e.g., 5-30 minutes) at test time, with consistent subjects, stable colors, and smooth motions.

ICLR 2026
arXiv / Paper / Project Page / Code / Demo / Data / Slides

Powered by large-scale training data curated from our panoramic data curation engine, and the SphereViT for addressing the spherical distortions in panoramas, DA² is able to predict dense, scale-invariant distance from a single 360° panorama in an end-to-end manner, with remarkable geometric fidelity and strong zero-shot generalization.

arXiv 2025
arXiv / Paper / Project Page / Code / Demo (D) / Demo (N)

Lotus-2 is an advanced two-stage deterministic framework for monocular geometric dense estimation built upon FLUX. By effectively analyzing the DiT-based rectified-flow formulation and leveraging pre-trained generative model as a deterministic world prior, Lotus-2 achieves SoTA performance while producing significantly finer details.

ICLR 2025
arXiv / Paper / Project Page / Code / Demo (D) / Demo (N) / ComfyUI

Lotus is a diffusion-based visual foundation model with a simple yet effective adaptation protocol, aiming to fully leverage the pre-trained diffusion's powerful visual priors for dense prediction. With minimal training data, Lotus achieves SoTA performance in two key geometry perception tasks, i.e., zero-shot monocular depth and normal estimation.

ICLR 2025
arXiv / Paper / Project Page / Code / Demo

Characterized by its emphasis on the interpretation of subject-essential attributes, the proposed DisEnvisioner effectively identifies and enhances the subject-essential feature while filtering out other irrelevant information, enabling exceptional image customization without cumbersome tuning or relying on multiple reference images.

SIGGRAPH Asia 2024
Paper / Video

DIScene is capable of generating complex 3D scene with decoupled objects and clear interactions. Leveraging a learnable Scene Graph and Hybrid Mesh-Gaussian representation, we get 3D scenes with superior quality. DIScene can also flexibly edit the 3D scene by changing interactive objects or their attributes, benefiting diverse applications.

CVPR 2024 Highlight
arXiv / Paper / Code / Demo / Video

We present LucidDreamer, a text-to-3D generation framework, to distill high-fidelity textures and shapes from pretrained 2D diffusion models with a novel Interval Score Matching objective and an advanced 3D distillation pipeline. Together, we achieve superior 3D generation results with photorealistic quality in a short training time.

Academic Service

Education & Working