Music-to-dance generation requires precisely aligning movement dynamics with musical rhythm, yet existing methods rely on shallow conditioning or auxiliary beat-alignment objectives that fail to establish stable beat–motion correspondences. We present ChoreDiffusion, a diffusion-based framework that integrates explicit beat guidance directly into the denoising process. Central to our approach is a beat-enhanced cross-modal attention mechanism that injects beat-salience cues at every refinement step, promoting fine-grained synchronization beyond the reach of conventional conditioning pipelines. To support multiple dance styles within a unified model, we incorporate lightweight low-rank adaptation (LoRA) modules that encode style-specific motion signatures with only a small set of additional parameters per style, and a three-stage progressive curriculum stabilizes the joint learning of rhythmic alignment and stylistic expressivity. Experiments on two public multi-style dance benchmarks (AIST++ and FineDance) show that ChoreDiffusion achieves the lowest FID values among the compared generation methods on both benchmarks, while maintaining competitive rhythm alignment and multi-style controllability. These results indicate that embedding beat-aware guidance during generation, rather than applying it afterwards, is an effective route toward human-like musicality in music-driven choreography.
Motivated by the limitation that most contrastive knowledge distillation methods optimize only one perspective, this paper proposes a simple and efficient contrastive knowledge distillation method that formulates distillation as a dual-perspective objective over the class-wise and the sample-wise. Unlike prior contrastive knowledge distillation, we jointly align teacher and student logit distributions at both perspectives. Concretely, same input pairs serve as anchors whose similarities are optimized against negatives drawn from other classes, enforcing class-wise alignment and maximizing separability between classes. To counterbalance the potential distortion of sample geometry caused by pure class-wise alignment, we simultaneously perform sample-wise alignment, matching each student logit only to its teacher counterpart for the same input, irrespective of class relations. This dual alignment supplies complementary supervision, stabilizes sample-wise structure, and sharpens decision boundaries. Moreover, we add a label-guided same class contrast that promotes compactness within classes while further enhancing separability between classes. We further propose a delayed learning strategy that orchestrates multiple contrastive losses, enabling a more effective integration that preserves their respective advantages. Across CIFAR-100, ImageNet, Tiny-ImageNet, and STL-10, our approach surpasses most logit-based and feature-based baselines in image classification, transfer learning, and ViT distillation, while maintaining superior training efficiency.