Haoge Deng

I am currently a AI Algorithm Engineer at Alibaba. Previously, I completed my MSc at BUPT (supervised by Prof. Yonggang Qi) and conducted research at CASIA and BAAI with Prof. Zhaoxiang Zhang and Dr. Xinlong Wang.

My research interests include generative models , with a particular focus on multimodal generation .

Email  /  Scholar  /  Github

Research

Representative papers are highlighted. * indicates equal contribution, # indicates corresponding author.

Emu3.5 is a natively multimodal world model that unifies vision and language through end-to-end next-token prediction on interleaved video-derived data, enhanced by reinforcement learning and DiDA-based parallel decoding for efficient, spatiotemporally consistent generation.

URSA is a simple yet powerful discrete framework that formulates video generation as an iterative process of global refinement over spatiotemporal tokens, enabling efficient scaling to long-duration videos.

NOVA is a non-quantized autoregressive model that enables efficient video generation by reformulating the video creation as frame-by-frame and set-by-set predictions.

See3D is a scalable visual-conditional MVD model for open-world 3D creation, which can be trained on web-scale video collections without camera pose annotations.

GeoDream is a 3D generation method that integrates explicit generalized 3D priors with 2D diffusion priors to enhance the capability of obtaining unambiguous 3D consistent geometric structures without sacrificing diversity or fidelity.

SketchKnitter is a method that achieves vectorized sketch generation by reversing the stroke deformation process using a diffusion model learned from real sketches, enabling the creation of higher quality, visually appealing sketches with fewer sampling steps.

Education

Experiences

Academic Services

• Conference Reviewer: CVPR, ICCV, ECCV, ICLR, ICASSP
  

Website Template