Designed and implemented a data-driven solution based on GenAI to help artists to create textures in game more efficiently. Trained the model on local device and deployed on Azure. The solution is already used in production. Paper preparing to be submitted to Siggraph 2025.- Worked closely with different stakeholders (artists, directors) to clarify business goals- Designed a text-to-texture algorithm using diffusion models and image-to-image translation- Implemented the algorithm in PyTorch, as well as other utilities in C++ and CUDA- Achieved high-quality texture synthesis and consistency between material maps using text prompts- Integrated the prototype into ComfyUI to gather internal user feedback- Deployed the model on Azure to simplify users’ life- Collaborate with managers, artists and programmers to polish the final model

We achieved SOTA on 3D reconstruction, material estimation and relighting for shiny objects. We improved based recent advances in 3D reconstruction with NeRF and SDF. This project was extended to a paper accepted to 3DV 2025.- Developed a novel algorithm and network architecture in PyTorch- Generated a standard dataset for future work’s benchmarking- Trained the model on HPC cluster with multiple graphics cards- Wrote numerous bash scripts to facilitate training and post-processing- Leveraged many libraries in computer vision and image processing, such as openexr, OpenCV, colmap - Scientific paper writing in Latex

We analyzed mathematically and numerically the behavior of a dynamic system around its bifurcation point. A simulation algorithm was also implemented to test our conclusion.