Zhaolin Lu

Ziji Lu

Text-to-image diffusion models often struggle to achieve accurate semantic alignment between generated images and text prompts while maintaining efficiency for deployment on resource-constrained hardware. Existing approaches either incur substantial computational overhead through noise optimization or compromise semantic fidelity by aggressively pruning tokens. In this work, we propose OptiPrune, a unified framework that combines distribution-aware initial noise optimization with similarity-based token pruning to address both challenges simultaneously. Specifically, (1) we introduce a distribution-aware noise optimization module guided by attention scores to steer the initial latent noise toward semantically meaningful regions, mitigating issues such as subject neglect and feature entanglement; (2) we design a hardware-efficient token pruning strategy that selects representative base tokens via patch-wise similarity, injects randomness to enhance generalization, and recovers pruned tokens using maximum similarity copying before attention operations. Our method preserves the Gaussian prior during noise optimization and enables efficient inference without sacrificing alignment quality. Experiments on benchmark datasets, including Animal-Animal, demonstrate that OptiPrune achieves state-of-the-art prompt-image consistency with significantly reduced computational cost.

Zhanyi Lu, Yue Zhou

Facial stylization aims to transform facial images into appealing, high-quality stylized portraits, with the critical challenge of accurately learning the target style while maintaining content consistency with the original image. Although previous StyleGAN-based methods have made significant advancements, the generated results still suffer from artifacts or insufficient fidelity to the source image. We argue that these issues stem from neglecting semantic shift of the generator during stylization. Therefore, we propose a facial stylization method that integrates semantic preservation constraint and pseudo-paired supervision to enhance the content correspondence and improve the stylization effect. Additionally, we develop a methodology for creating multi-level pseudo-paired datasets to implement supervisory constraint. Furthermore, building upon our facial stylization framework, we achieve more flexible multimodal and reference-guided stylization without complex network architecture designs or additional training. Experimental results demonstrate that our approach produces high-fidelity, aesthetically pleasing facial style transfer that surpasses previous methods.