CorGi: Accelerating Diffusion Transformers with Caching

This paper addresses the computational cost of Diffusion Transformers (DiT) in visual generation, a significant bottleneck. By introducing CorGi, a training-free method that caches and reuses transformer block outputs, the authors offer a practical solution to speed up inference without sacrificing quality. The focus on redundant computation and the use of contribution-guided caching are key innovations.

• •Proposes CorGi, a training-free method to accelerate DiT inference.
• •Utilizes block-wise interval caching to reduce redundant computation.
• •Introduces CorGi+ for text-to-image tasks, leveraging cross-attention maps.
• •Achieves up to 2.0x speedup while maintaining generation quality.