During April 2025, Tingchun Wang focused on stabilizing diffusion outputs in the nvidia-cosmos/cosmos-transfer1 repository by addressing a spatiotemporal control weight resize bug. Using Python and leveraging deep learning and computer vision expertise with PyTorch, Tingchun implemented an interpolation-based fix that ensures control weight maps are accurately resized to target dimensions. This targeted code change improved the reliability and consistency of diffusion-based inference across diverse inputs, reducing output artifacts and misalignment. The solution was delivered with minimal risk to the existing pipeline, enhancing system stability and predictability while lowering support overhead and increasing user trust in the model’s outputs.