ToDo: Token Downsampling for Efficient Generation of High-Resolution Images
This addresses the problem of slow and memory-intensive high-resolution image generation for users of diffusion models, offering a practical speedup with competitive fidelity.
The paper tackles the computational inefficiency of attention mechanisms in image diffusion models by proposing ToDo, a training-free token downsampling method that accelerates Stable Diffusion inference by up to 2x for common sizes and 4.5x for high resolutions like 2048x2048.
Attention mechanism has been crucial for image diffusion models, however, their quadratic computational complexity limits the sizes of images we can process within reasonable time and memory constraints. This paper investigates the importance of dense attention in generative image models, which often contain redundant features, making them suitable for sparser attention mechanisms. We propose a novel training-free method ToDo that relies on token downsampling of key and value tokens to accelerate Stable Diffusion inference by up to 2x for common sizes and up to 4.5x or more for high resolutions like 2048x2048. We demonstrate that our approach outperforms previous methods in balancing efficient throughput and fidelity.