Worked on enhancing the ModelTC/LightX2V repository by addressing reproducibility issues in the WanStepDistillScheduler’s noise generation process. Focused on deep learning and model development using Python, the approach involved implementing a deterministic noise pipeline through a specified generator and fixed shape, ensuring consistent noise output across experimental runs. This solution improved the reliability of benchmarking and training by making results traceable and repeatable, supporting ongoing reproducibility initiatives. No new user-facing features were introduced during this period, as the primary goal was to stabilize core functionality while minimizing regression risk and maintaining high codebase quality throughout the development cycle.