CONFLUX: A Latent Diffusion Model for 3D Chest-CT Synthesis with RL Post-Training

CONFLUX is a latent generative model for native 3D chest CT. A 3D variational autoencoder compresses volumes, while a rectified-flow Transformer generates samples in latent space. Conditioning covers 18 abnormality findings, sex, age, and reconstruction kernel through adaptive layer normalization. The paper reports a tri-planar FID of 32.3, compared with 74.6 for MAISI. An online group-relative policy optimization stage rewards agreement between requested findings and an independent classifier's predictions, reportedly removing 47% of the reliability gap to real scans. The authors release the model and about 200,000 synthetic CT volumes with metadata.