SciGroveBeta
Medicine

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

Max Van Puyvelde, Halil Ibrahim Gulluk, Wim Van Criekinge, Olivier Gevaert

Featured July 14, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

A smart computer model creates realistic 3D chest X-rays with specific health conditions, and then uses a special reinforcement learning trick to make sure the generated images truly show what was asked for, like a specific lung problem.

In depth
The paper introduces CONFLUX, a novel 3D latent rectified-flow model for synthesizing high-fidelity chest CT scans. Its core innovation lies in combining a 3D variational autoencoder with a rectified-flow transformer for efficient generation in latent space, and crucially, integrating an online reinforcement learning post-training stage. This post-training, using group-relative policy optimization (GRPO), explicitly fine-tunes the model to ensure that generated volumes faithfully exhibit their requested clinical attributes, addressing a key limitation of standard flow-matching objectives.

Key Takeaways

  • 1
    CONFLUX is a natively 3D latent rectified-flow model that achieves state-of-the-art synthesis quality for chest CT volumes, outperforming strong volumetric baselines on metrics like tri-planar Fréchet distance (FID).
  • 2
    The model incorporates adaptive layer normalization (adaLN) to condition generation on structured radiological metadata, enabling direct control over 18 abnormality findings, sex, age, and reconstruction kernel.
  • 3
    A novel reinforcement learning (RL) post-training stage, based on group-relative policy optimization (GRPO), significantly improves the faithfulness of generated images to their requested clinical attributes, recovering 47% of the gap to real-scan reliability.

Conceptual Flow

HIGH LEVEL
1
Methodology: How CONFLUX Works

The system first squishes big X-ray pictures into tiny codes, then uses these codes to draw new X-rays, and finally teaches itself to make sure the new X-rays perfectly match what was requested.

Big X-ray Picture
Shrink to Code
Tiny X-ray Code
2
Results: Better, More Reliable X-rays

The new method makes X-rays that look much more real and reliably show the specific health problems they were asked to create, which is a big step forward.

Old Way: X-ray Request
Old Way: Generated X-ray
Often Misses Details
New Way: X-ray Request
New Way: Generated X-ray