Dianbing Xi

I'm a final-year Ph.D. student in the Advanced Rendering and Creativity (ARC) Group at the State Key Laboratory of CAD&CG, Zhejiang University, supervised by Prof. Rui Wang and Prof. Yuchi Huo. I have been fortunate to collaborate with Dr. Jiepeng Wang at TeleAI, Prof. Yuan Liu at the Hong Kong University of Science and Technology, and Prof. Yuliang Xiu at Westlake University.

Email  /  Scholar  /  Github

Research

My research primarily focuses on Computer Graphics and Artificial Intelligence Generated Content (AIGC). I currently work on Video Generation, which I see as a promising new paradigm for future rendering. I also have prior experience in 3D Avatar Generation, Inverse Rendering, and Neural Rendering.

!!! I am expected to complete my Ph.D. in June 2026 and am actively seeking research positions in industry or postdoctoral opportunities.

Feel free to contact me at db.xi@zju.edu.cn regarding relevant positions or collaboration opportunities.

Selected Publications

CtrlVDiff unifies forward and inverse video generation within a single model, enabling the extraction of all modalities in a single pass. It provides layer-wise control over appearance and structure, facilitating applications such as material editing and object insertion.

OmniVDiff enables controllable video generation and understanding in a unified video diffusion framework.

PFAvatar reconstructs and edits personalized avatars from OOTD photos using pose-aware diffusion models and 3D-SDS, overcoming previous method limitations and supporting editing and animation.

AniTex generates high-quality, temporally consistent PBR materials for animated 3D objects using a two-stage diffusion pipeline, achieving realistic multi-view textures and outperforming existing static-focused methods.

IntrinsicControlNet uses intrinsic images and cross-domain control to generate photorealistic yet explicitly controllable images, combining rendering-level precision with diffusion-model realism while bridging gaps between synthetic and real data.

We propose a dual-band neural GI framework using object-centric feature grids and single-bounce queries to fuse low- and high-frequency illumination, enabling high-quality multi-frequency dynamic reflections beyond prior GI and denoising methods.

We propose MIRReS, a two-stage inverse rendering framework that recovers explicit geometry, materials, and lighting using multi-bounce path tracing with reservoir-sampled gradients, achieving accurate intrinsic decomposition and SOTA performance.

We introduce I^2-SDF, a neural SDF–based inverse rendering framework that reconstructs indoor geometry, materials, and radiance via differentiable MC ray tracing, enabling high-quality reconstruction, relighting, and editing with SOTA performance.

We propose a single-image inverse rendering framework using differentiable MC ray tracing and uncertainty-aware out-of-view lighting prediction, enabling accurate geometry, lighting, and material recovery with photorealistic edits beyond prior methods.

Experience

©Dianbing Xi. Last updated: .