An Enhanced Semantic Binding Method Based on Balanced Attention Maps

Attention mechanisms in text-to-image models improve interpretability but suffer from cross-prompt bias causing semantic misalignment and quality degradation. We propose Balanced Attention Maps (BAM), a training-free method that dynamically adjusts attention maps during inference using linguistic structures in prompts, enhancing semantic alignment and visual coherence. Validated through experiments with three leading models across major benchmarks, BAM outperforms existing methods in both quantitative metrics (Inception Score, CLIP Score) and human evaluations (semantic consistency, visual naturalness). Results confirm BAM's superiority in maintaining generation fidelity while addressing cross-prompt bias through attention recalibration, without requiring retraining.