LiteAttention: A Temporal Sparse Attention for Diffusion Transformers

The paper titled 'DiffPro: Joint Timestep and Layer-Wise Precision Optimization for Efficient Diffusion Inference' presents a new framework aimed at improving the efficiency of diffusion models, which are known for generating high-quality images but require extensive computational resources. DiffPro optimizes inference by tuning timesteps and layer precision without additional training, achieving significant reductions in latency and memory usage. The framework combines a sensitivity metric, dynamic activation quantization, and a timestep selector, resulting in up to 6.25x model compression and 2.8x faster inference.