Browse Teach Curate

Cached Adaptive Token Merging for Diffusion Model Acceleration

A training-free method to accelerate diffusion-model image generation by reducing the tokens fed to self-attention. It merges token pairs whose similarity exceeds a threshold, and improves plain token merging with an adaptive threshold tuned to the observed similarity distribution and a cache that stores merged pairs and reuses them across adjacent, highly similar denoising steps, cutting redundant computation while preserving image quality.

Data Text

Cached Adaptive Token Merging: Dynamic Token Reduction and Redundant Computation Elimination in

Cached Adaptive Token Merging (CA-ToMe) is a training-free way to speed up image generation with diffusion models. Because self-attention costs grow quadratically with the number of tokens, the metho…