Core Concepts
Efficient diffusion-based method for test-time adaptation using corruption editing.
Abstract
The article introduces Decorruptor, a novel test-time adaptation method that leverages diffusion models for efficient editing of corrupted images. By fine-tuning the model with a corruption modeling scheme, Decorruptor-DPM enhances robustness against distribution shifts. Additionally, Decorruptor-CM accelerates the model through consistency distillation, achieving faster inference times. Extensive experiments demonstrate superior performance and generalization capabilities across various architectures and datasets.
Stats
Our model achieves the best performance with a 100 times faster runtime than that of a diffusion-based baseline.
Decorruptor-CM enables 46 times faster input updates than DDA owing to latent-level computation and fewer generation steps.
Decorruptor-CM achieves similar corruption editing effects to Decorruptor-DPM’s 20 network function evaluations (NFEs) with only 4 NFEs.
Quotes
"Our model achieves the best performance with a 100 times faster runtime than that of a diffusion-based baseline."
"Decorruptor-CM enables 46 times faster input updates than DDA owing to latent-level computation and fewer generation steps."