During June 2026, contributed to the vllm-project/vllm-omni repository by delivering an enhancement to the VACE Image-to-Video workflow. This work involved refactoring the existing example to conform with standard task templates and integrating model_extras, which improved both the consistency and configurability of image-to-video generation. The approach emphasized maintainability and reusability, aligning the demo with project conventions and simplifying downstream integration for future experiments. Utilizing Python scripting, image processing, and machine learning techniques, the developer focused on reliable feature delivery and build/test discipline, supporting more robust demos and accelerating iteration for stakeholders without introducing standalone critical bug fixes.