Worked on the nvidia-cosmos/cosmos-transfer1 repository to improve the stability of diffusion-based outputs by addressing a spatiotemporal control weight resize issue. Applied expertise in computer vision and deep learning to implement an interpolation-based fix in Python using PyTorch, ensuring that control weight maps are accurately resized to match target dimensions. This targeted code change reduced misalignment and output artifacts, resulting in more reliable and consistent inference across diverse inputs. The solution enhanced the robustness of the existing pipeline without introducing risk, ultimately lowering support needs and increasing user trust in the system’s diffusion results for a variety of applications.