Developed and delivered FlowGRPO Image-based Rewards Support within the volcengine/verl repository, enabling both generative and rule-based reward models for image generation tasks. The work involved implementing an ImageRewardManager and extending existing reward loop components to process visual inputs, providing end-to-end training signals for image-based workflows. Integration was validated with the Qwen-VL OCR path and a rule-based reward based on jpeg compressibility, with dedicated unit tests ensuring reliability. Utilizing Python, image processing, and machine learning techniques, the contribution improved test coverage, enhanced CI readiness, and aligned the framework with future roadmap goals for image generation model experimentation and deployment.