Worked on the huggingface/diffusers repository to enhance the reliability of classifier-free guidance in the ZImagePipeline. Addressed a critical bug by correcting the threshold logic to ensure the guidance scale operates according to the intended formula, directly impacting the predictability of generated outputs. Employed Python and leveraged expertise in image processing and machine learning to implement the fix, followed by comprehensive end-to-end validation and the addition of regression safeguards. This approach minimized the risk of similar issues recurring in production, improved the correctness of guidance behavior, and reduced support overhead, reflecting a methodical and quality-focused engineering process throughout the development cycle.