ML Researcher | Video Diffusion · 3D/4D Generation/Reconstruction · VLM
I am a Machine Learning Researcher at Huawei's 3D Vision Team. My current focus is on video diffusion models — specifically DiT-based architectures like Wan2.1, and HunyuanVideo for various tasks such as object removal, video inpainting, and model distillation. I work with both bidirectional (full-attention) and autoregressive generation paradigms, and I implement flow matching at the training level, not only as an inference scheduler.
On the 3D side, I continue to work with 3D and 4D Gaussian Splatting methods (4D-GS, SC-GS, SpacetimeGaussians) and feed-forward reconstruction pipelines (DUSt3R, MASt3R, InstantSplat). My earlier work at Ubisoft La Forge focused on geometry-aware texture synthesis and differentiable shading for digital avatars.
I completed my Master's at York University (BioMotion Lab), advised by Prof. Niko Troje and Prof. Marcus A. Brubaker.
- Video Diffusion Models: DiT-based architectures, bidirectional and autoregressive generation, flow matching, distillation (consistency models, flow shortcuts), ControlNet-style conditioning with pose/depth/segmentation signals
- 3D / 4D Reconstruction: 3D and 4D Gaussian Splatting, feed-forward reconstruction, novel view synthesis
- Generative AI for Faces & Avatars: Talking head generation, face reenactment, geometry-aware texture generation, differentiable shading
- Vision-Language Models: CLIP, SigLIP, Florence — used as conditioning or semantic supervision in generation pipelines
- Other CV: Anomaly detection, face swapping
-
Geometry-Aware Texture Generation for 3D Head Modeling with Artist-driven Control CVPRW 2025 | Project Page | Paper
-
MoSAR: Monocular Semi-Supervised Model For Avatar Reconstruction Using Differentiable Shading CVPR 2024 | Project Page | Paper
Additional work in video diffusion currently under review (2026).
- Languages: Python, C/C++, MATLAB
- Deep Learning: PyTorch, TensorFlow, JAX, Keras
- Video & 3D: Diffusers, 3DGS, DUSt3R / MASt3R, ControlNet, LoRA / adapter fine-tuning
- Training Infrastructure: Multi-GPU distributed training with FSDP and DeepSpeed
- Libraries: OpenCV, NumPy, Pandas, Scikit-learn
- Tools: Git, LaTeX, Linux
