Publications

In submission

DeVA: Decoupled Video-Action Model with physical guidance for robot policy learning

Mengqi Zhang, Sahil Khose, Simar Kareer, Yuchen Song, Unnat Jain, Judy Hoffman

Transferring visual and physical priors from video models into robot policies, with strong generalization and up to 20× fewer training examples than the baseline.

In submission

WFM-Eval: Interpretable Error Diagnostics for Video World Models in Robotics

Sahil Khose, Mengqi Zhang, Prithvijit Chattopadhyay, Judy Hoffman

Diagnosing failures that standard video metrics miss: five world models, two robot datasets, and links to downstream policy performance.

CVPR 2026 · MUSI & Visual Concepts Workshops

SPOT: Structured Prompting with Object-centric Tokens for open-world scene graphs

Mengqi Zhang, Sahil Khose, Fiona Ryan, Judy Hoffman

Spatially and semantically grounded scene graphs with open-source VLMs, including cross-domain generalization and zero-shot 3D transfer.

CVPR 2026 · AI4CC Workshop Oral

Style-Instructed Mask-Free Virtual Try-On

Mengqi Zhang, Qi Li, Mehmet Saygin Seyfioglu, Karim Bouyarmane

Mask-free virtual try-on with natural-language control over garment types and wearing styles.

NeurIPS 2024 Oral

MeshFormer: High-Quality Mesh Generation with 3D-Guided Reconstruction Model

Minghua Liu^*, Chong Zeng^*, Xinyue Wei, Ruoxi Shi, Linghao Chen, Chao Xu, Mengqi Zhang, Zhaoning Wang, Xiaoshuai Zhang, Isabella Liu, Hongzhi Wu, Hao Su

High-quality textured mesh reconstruction with 3D guidance; contributed to text-conditioned multi-view generation.

CVPR 2024

HOIDiffusion: Generating Realistic 3D Hand-Object Interaction Data

Mengqi Zhang^*, Yang Fu^*, Zheng Ding, Sifei Liu, Zhuowen Tu, Xiaolong Wang

Controllable hand-object interaction synthesis with disentangled geometry and appearance, improving downstream 6D object pose estimation.

ICLR 2024

Patched Denoising Diffusion Models For High-Resolution Image Synthesis

Zheng Ding^*, Mengqi Zhang^*, Jiajun Wu, Zhuowen Tu

Resolution-agnostic diffusion for high-resolution synthesis and generation beyond training resolution.