DecQ: Detail-Condensing Queries for Enhanced Reconstruction and Generation in Representation Autoencoders

Researchers introduced DecQ, a query-based method for representation autoencoders that adds lightweight detail-condensing queries from intermediate vision features.

The approach improves reconstruction quality and image generation, boosting PSNR while preserving pretrained semantic representations.

DecQ also helps latent diffusion models converge faster and achieve better generative quality with minimal extra compute.