Rethinking attention modules of SDXL in a flow perspective

We reinterpreted the attention mechanism in the training-free methods of diffusion models using flow matching theory. Based on this, we proposed the FlowAttnXL training-free framework for text-to-image tasks. It introduces two modules: Flow Matching Cross-Attention (FMXA) and Flow Arithmetic Perturbation Guidance (FAPG). These modules operate plug-and-play, requiring no retraining and compatible with pipelines like Stable Diffusion XL (SDXL). Experiments on MS COCO show significant improvements over state-of-the-art methods. Ablation studies validate the flow-based design, with FMXA improving text-image alignment and FAPG enhancing visual fidelity. This work establishes a unified framework for semantic alignment and feature propagation in diffusion models.