[NeurIPS’25] Fourier Token Merging: Understanding and Capitalizing Frequency Domain for Efficient Image Generation
Published in The 39th Annual Conference on Neural Information Processing Systems (NeurIPS), 2025
A lightweight framework that transforms intermediate tokens via the Discrete Fourier Transform, using low-frequency structural priors for robust token clustering and merging — up to 25% lower inference latency on Stable Diffusion v1.5 with comparable visual fidelity, with theoretical error bounds for frequency-domain merging.
Recommended citation: Jiesong Liu, Xipeng Shen. (2025). "Fourier Token Merging: Understanding and Capitalizing Frequency Domain for Efficient Image Generation." NeurIPS 2025
Download Paper