Controllable generative models of 3D medical images can synthesize volumes with specified clinical attributes, but this demands samples that are simultaneously high-fidelity, natively 3D, and faithful to the requested conditioning. We present CONFLUX, a latent diffusion model for chest computed tomography (CT): a 3D variational autoencoder compresses each volume, and a rectified-flow transformer generates in the latent space. Generation is conditioned on structured radiological metadata (18 abnormality findings, sex, age, and reconstruction kernel) through adaptive layer normalization. The model leads strong volumetric baselines on tri-planar Frechet distance (FID 32.3 vs. 74.6 for MAISI) while exposing direct control over clinical attributes. To strengthen that control we add an online reinforcement-learning post-training stage (group-relative policy optimization) that rewards how reliably a classifier recovers the requested findings from each generated volume. Judged by a separate, independent classifier, post-training removes 47% of the shortfall relative to real-scan reliability. We release the model and a ~200k synthetic chest-CT dataset with conditioning metadata spanning a wide variety of clinical findings.
Alex: Welcome to another episode of ResearchPod. Today, we're looking at a new model called CONFLUX, designed to generate realistic 3D medical chest scans.
Sam: So the core challenge here is that current AI can make something that looks like a chest scan, but it often misses the specific medical details a doctor would actually need?
Alex: Exactly. Think of the difference between a painting that looks like a hospital room and one that accurately shows a specific piece of equipment in the right place. The first might fool a casual observer, but it wouldn't help a surgeon. That's the problem CONFLUX is trying to solve—generating scans that don't just look right, but contain the precise clinical findings a researcher is looking for. Things like a small nodule in a specific location, or fluid collecting around the lungs.
Sam: And getting thousands of real, labeled scans to train on isn't easy, because of patient privacy laws.
Alex: Right. Real medical data is tightly protected, which makes it very hard to build large training sets. That's part of why synthetic data—AI-generated scans—is so appealing. If you can generate realistic, medically accurate images on demand, you sidestep the privacy problem entirely.
Sam: So how does CONFLUX actually learn to put the right findings in the right places? If you just train a model to look realistic, there's no guarantee it'll include, say, fluid around the lungs where it's supposed to be.
Alex: That's the key insight. Think of it like a student learning to paint. The first round of training teaches them the general style—the broad strokes of what a chest scan looks like. But that doesn't mean they'll include the specific detail you asked for. So the authors added a second stage, and this is where the approach gets interesting.
Sam: How does the second stage work?
Alex: They bring in a separate AI—call it the judge. This judge has been trained on real scans and knows how to spot genuine medical findings. Every time the main model generates a new image, the judge checks it: is the requested finding actually there? If yes, the model gets a reward. If not, it gets nothing. Over thousands of attempts, the model learns to prioritize diagnostic accuracy, not just visual realism.
Sam: So it's less like an art class and more like a driving test. You don't pass just for looking confident—you have to actually stop at the right signs.
Alex: That's a good way to put it. Rather than grading each image in isolation, the system compares a whole batch of generated images against each other and identifies which ones best matched the requested findings. The model learns from that comparison—figuring out, over time, what separates a good attempt from a poor one.
Sam: But doesn't this whole system depend on how good the judge is? If the judge misses a subtle finding, the model never learns it's doing something wrong.
Alex: That is the central limitation, and the authors are direct about it. The model's accuracy is strictly bounded by the judge's accuracy. If the judge can't reliably detect a subtle feature, the reward signal becomes unreliable, and the model won't learn that detail. It's only as capable as its teacher.
Sam: So did this second stage actually make a measurable difference?
Alex: It did. By adding it, they were able to close nearly half the gap between their model's initial performance and the reliability of real-world medical scans. In a field where the margin for error is very small, that's a meaningful improvement.
Sam: And they're releasing the model alongside a large dataset of synthetic scans—which could be a real resource for researchers who can't access real patient data.
Alex: That's one of the more practically significant parts of the work. The hope is that other researchers can use these synthetic scans to train their own diagnostic tools, without needing to navigate the legal and ethical complexities of working with real patient records. The authors also suggest that future versions might pair these images with automatically generated diagnostic reports—so the system could produce both the scan and a written description of what it shows.
Sam: It sounds like a careful, incremental step in a field where accuracy is non-negotiable.
Alex: That's exactly the right framing. The value isn't in making images that look impressive—it's in making images that are genuinely useful. And this work suggests that reinforcement learning, with the right judge, is a credible path toward that goal. Thanks for listening to ResearchPod.