In current generative design tools, users are required to provide detailed prompts and repeatedly adjust them through trial and error to guide the model in generating satisfactory outcomes, leading to inefficiency and difficulty in cotrolling the results. This paper proposes a dynamic interactive product generation design method based on affective knowledge embedding and collaborative user feedback, aiming to achieve precise generation and adaptive optimization through user feedback. The core framework consists of three components: design element combination prediction, Stable Diffusion-based generation, and reinforcement learning optimization. First, a design element combination prediction model is constructed to mine frequently co-occurring element combinations from a library of common design elements. This module can predict possible design combinations based on the user-selected design elements. Then, the predicted combinations are converted into Stable Diffusion prompts to generate initial design schemes, which are presented to the user. On this basis, a dynamic evaluation-optimization mechanism is constructed using a deep Q-network, which calculates a reward function based on the user’s real-time selection of generated results. This guides the model to continuously optimize the element prediction and recommendation strategy, thereby achieving dynamic adjustment and adaptive optimization, and forming a closed-loop interactive process of generation–feedback–optimization.