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

Researchers at Stanford University and Ghent University present CONFLUX, a controllable, natively 3D chest-CT generator that combines a 3D variational autoencoder with a rectified-flow transformer operating in latent space. The model compresses CT volumes by a factor of 8 into 16-channel latents and uses adaptive layer normalization to condition generation on 18 abnormality findings, sex, age range, and reconstruction kernel. Its design relates to FLUX from Black Forest Labs, while its group-relative policy optimization, or GRPO, adapts reinforcement-learning ideas popularized by DeepSeekMath. Trained on 18,417 CT-RATE scans, CONFLUX achieves a tri-planar FID of 32.3, compared with 74.6 for MAISI and 145.4 for GenerateCT. The paper’s main novelty is RL post-training: a frozen classifier rewards whether generated volumes realize their requested findings, and an independent image-space judge shows macro average precision increasing from 0.330 to 0.344 and macro AUROC from 0.684 to 0.699. This recovers 47% of the reliability gap between pre-RL synthesis and real scans, without improving unrelated sex, age, or kernel conditioning. The authors release the model and approximately 200K metadata-linked synthetic chest-CT volumes spanning diverse clinical findings.