Frequency Decoupling Attention (FDA) is a mechanism within the Frequency-aware Skip Estimation (FaSE) module of DiffCR, designed to align the $\epsilon$-prediction prior from a latent diffusion model with compressed latents across different timesteps. It achieves this by decoupling frequency components, crucial for efficient and high-fidelity image compression.
Frequency Decoupling Attention (FDA) is a key component in a new image compression method called DiffCR, which uses diffusion models. It helps align different parts of the image information, specifically the predicted noise and compressed data, to make the compression process much faster and produce higher quality images.
FDA
Was this definition helpful?