AI RESEARCH

Domain-Specific Latent Representations Improve the Fidelity of Diffusion-Based Medical Image Super-Resolution

arXiv CS.AI

ArXi:2604.12152v1 Announce Type: cross Latent diffusion models for medical image super-resolution universally inherit variational autoencoders designed for natural photographs. We show that this default choice, not the diffusion architecture, is the dominant constraint on reconstruction quality.