HiLo-Token: Input-Adaptive High-Low Frequency Token Compression for Efficient Image Editing

The article introduces HiLo-Token, an input-adaptive token compression framework designed to enhance the efficiency of image editing tasks in generative AI models, particularly Diffusion Transformers (DiTs). By dynamically allocating token budgets based on spatial frequency—retaining more tokens in high-frequency areas and using downsampled tokens for low-frequency regions—HiLo-Token achieves significant latency reductions, yielding speedups of 3.13x, 2.59x, and 1.67x on A100-80GB GPUs across different mask ratios while maintaining generation quality. This development is crucial for practitioners aiming to optimize performance in real-time image editing applications.