CAT Pruning: Cluster-Aware Token Pruning For Text-to-Image Diffusion Models
Keywords: Generative Models, Efficient Machine Learning
Abstract: Diffusion models have transformed generative tasks, particularly in text-to-image synthesis, but their iterative denoising process is computationally intensive. We present a novel acceleration strategy that combines token-level pruning with cache mechanisms to address this challenge. By utilizing Noise Relative Magnitude, we identify significant token changes across iterations. Additionally, we incorporate spatial clustering and distributional balance to enhance token selection. Our experiments demonstrate 50\%-60\% reduction in computational cost while maintaining model performance, offering a substantial improvement in the efficiency of diffusion models.
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5910
Loading