<p>Virtual reality (VR) technology has emerged as a transformative tool for innovation and entrepreneurship education, yet existing systems face critical limitations in scene generation realism and cognitive load management. This research proposes an integrated framework combining attention mechanism-enhanced scene generation with adaptive cognitive load balancing strategies to address these challenges. The framework incorporates a multi-scale attention architecture operating across spatial, channel, and temporal dimensions to selectively emphasize pedagogically critical entrepreneurial elements while maintaining contextual coherence. A real-time cognitive load assessment system monitors physiological signals and behavioral patterns to implement dynamic scene complexity adjustments, maintaining learners within optimal challenge zones. Experimental validation across 45 entrepreneurship scenarios with 120 participants demonstrates substantial improvements: scene quality scores increased by 26–45% compared to baseline methods, error rates decreased by 60.9%, and task completion times reduced by 23.7%. The system maintained stable cognitive load within target ranges for 87.3% of session duration during extended 60-minute training sessions. Ablation studies confirm the essential contributions of individual attention components, with quality degradation of 12–18% upon removal. These findings advance both computational innovation in generative VR systems and pedagogical methodologies for entrepreneurship education, providing practical solutions for scalable, personalized training applications across diverse educational and professional development contexts.</p>