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

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Medical Imaging · Depth: Expert, long

Summary

CONFLUX is a novel latent diffusion model designed for controllable 3D chest computed tomography (CT) synthesis, integrating a 3D variational autoencoder for volume compression and a rectified-flow transformer for latent space generation. The model is conditioned on structured radiological metadata, including 18 abnormality findings, sex, age, and reconstruction kernel, using adaptive layer normalization. CONFLUX achieves a tri-planar Fréchet distance (FID) of 32.3, significantly outperforming baselines like MAISI (74.6) and GenerateCT (145.4), nearing the VAE reconstruction ceiling of 22.6. A crucial online reinforcement-learning post-training stage, employing group-relative policy optimization, enhances conditioning faithfulness, recovering 47% of the reliability gap compared to real scans, as validated by an independent classifier. The trained model and a ~200,000-volume synthetic chest-CT dataset, complete with diverse conditioning metadata, have been released.

Key takeaway

For AI Scientists and Machine Learning Engineers developing controllable 3D medical image generators, CONFLUX demonstrates a powerful approach. If you are augmenting under-represented cohorts or designing controlled studies, you should consider its latent diffusion architecture with RL post-training. This method significantly improves both synthesis quality and the faithfulness of requested clinical attributes, offering a reliable path to high-fidelity synthetic data. Explore the released ~200,000-volume dataset for your research.

Key insights

Latent diffusion with RL post-training significantly enhances controllable 3D medical image synthesis fidelity and faithfulness.

Principles

• Latent rectified-flow transformers are effective for high-resolution 3D medical image generation.
• Reinforcement learning directly optimizes conditioning faithfulness in generative models.
• Independent classifiers are vital for unbiased faithfulness evaluation.

Method

A three-stage process: VAE compresses CT to latent, rectified-flow transformer generates in latent space with adaLN-zero conditioning, then GRPO post-training improves faithfulness via a classifier reward.

In practice

• Employ 3D VAEs for compact medical volume latent representations.
• Apply GRPO to fine-tune generative models for improved attribute faithfulness.
• Share synthetic datasets with metadata to augment real cohorts.

Topics

• Latent Diffusion Models
• 3D Chest CT Synthesis
• Reinforcement Learning
• Medical Image Generation
• Group-Relative Policy Optimization
• Data Augmentation

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• NAMD: Virtual Follow-up Computed Tomography Synthesis via Nodule-Aligned Multimodal Diffusion Models for Early Lung Cancer Diagnosis — NAMD synthesizes future lung nodule CTs from baseline and EHR using aligned latent diffusion for early cancer diagnosis.
• Scaling Generative Foundation Models for Chest Radiography with Rectified Flow Transformers — A 1.3B-parameter generative foundation model for chest radiographs achieves expert-indistinguishable synthesis and controllable editing.
• Anatomy-Guided Residual Motion Diffusion for Controllable 4D Cardiac MRI Synthesis — A cascaded latent diffusion model synthesizes controllable, anatomically consistent 4D cardiac MRI, improving AI model robustness with limited data.
• WING: A Window-Prior-Based Generative Network with Gated Inception for Cross-Modality CT Synthesis — WING synthesizes CT from MRI/CBCT using windowed representations and a multi-component generative network for improved radiotherapy planning.
• LUCID: Learned Undersampling-Adaptive Consistency-Guided Inference with Deterministic Flow Matching for Sparse-View CT Reconstruction — LUCID employs Flow Matching and sparsity-adaptive consistency guidance for robust sparse-view CT reconstruction, reducing artifacts and hallucinations.
• When is 3D Worth It? A Resource-Performance Frontier for CNNs and Transformers in Lung CT — Input dimensionality significantly impacts model performance and stability in volumetric medical imaging.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.