Paper deep dive
CONFLUX: A Latent Diusion Model for 3D Chest-CT Synthesis with RL Post-Training
Max Van Puyvelde, Halil Ibrahim Gulluk, Wim Van Criekinge, Olivier Gevaert
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 7/7/2026, 5:58:01 AM
Summary
The paper introduces CONFLUX, a controllable 3D latent diffusion model for synthesizing chest CT volumes. It integrates a 3D variational autoencoder for compression, a rectified-flow transformer for generation conditioned on structured radiological metadata, and an online reinforcement learning post-training stage using Group-Relative Policy Optimization (GRPO) to enhance conditioning faithfulness. CONFLUX achieves superior synthesis quality (FID 32.3) compared to baselines like MAISI and GenerateCT, and the authors release a ~200k synthetic chest-CT dataset.
Entities (8)
Relation Signals (6)
CONFLUX → employs → Group Relative Policy Optimization
confidence 95% · we add an online reinforcement-learning post-training stage (group-relative policy optimization) that rewards how reliably a classifier recovers the requested findings
Group Relative Policy Optimization → improves → Conditioning Faithfulness
confidence 90% · post-training removes 47% of the shortfall relative to real-scan reliability.
CONFLUX → trainedon → CT-RATE
confidence 90% · We use CT-RATE [7], a chest-CT corpus with paired structured metadata... the autoencoder trains on the QC-passing set
CONFLUX → uses → 3D Variational Autoencoder
confidence 90% · a 3D variational autoencoder compresses each volume, and a rectified-flow transformer generates in the latent space.
CONFLUX → outperforms → MAISI
confidence 85% · The model leads strong volumetric baselines on tri-planar Frechet distance (FID 32.3 vs. 74.6 for MAISI)
CONFLUX → outperforms → GenerateCT
confidence 85% · Competitors are MAISI [5]... and GenerateCT [6]... Our model leads both baselines
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract: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.
Tags
Links
- Source: https://arxiv.org/abs/2607.02998v1
- Canonical: https://arxiv.org/abs/2607.02998v1
Trouble viewing inline? Open PDF directly →
Full Text
36,192 characters extracted from source content.
Expand or collapse full text
11institutetext: Department of Biomedical Data Science, Stanford University School of Medicine 22institutetext: Department of Mathematical Modelling, Statistics & Bioinformatics, Ghent University 33institutetext: Department of Electrical Engineering, Stanford University 33email: maxvpuyv@stanford.edu, gulluk@stanford.edu, wim.vancriekinge@ugent.be, ogevaert@stanford.edu CONFLUX: A Latent Diffusion Model for 3D Chest-CT Synthesis with RL Post-Training Max Van Puyvelde M. Van Puyvelde and H. I. Gulluk are joint first authors; W. Van Criekinge and O. Gevaert are joint senior authors. H. Ibrahim Gulluk Wim Van Criekinge Olivier Gevaert Abstract 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 (1818 abnormality findings, sex, age, and reconstruction kernel) through adaptive layer normalization. The model leads strong volumetric baselines on tri-planar Fréchet distance (FID 32.332.3 vs. 74.674.6 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%47\% of the shortfall relative to real-scan reliability. We release the model and a ∼200 200k synthetic chest-CT dataset with conditioning metadata spanning a wide variety of clinical findings. 1 Introduction Generative models of 3D medical images are used to augment under-represented cohorts, share data under privacy constraints, and synthesize counterfactual volumes. Each use requires the model to be controllable: a sample requested to exhibit a set of clinical attributes must realize them. Chest CT, which is volumetric, high-resolution, and described by structured radiological metadata, is an especially hard case; a model that is both high-fidelity and controllable on it is directly useful for cohort augmentation and controlled study design. High-resolution synthesis has converged on a latent rectified-flow design: an autoencoder compresses the signal and a single-stream transformer trained under a flow-matching objective generates in the learned latent space [15, 13, 3, 2]. We present CONFLUX, a natively 3D instance of this design for chest CT, conditioned on structured radiological metadata (abnormality findings, sex, age, and reconstruction kernel) through adaptive layer normalization. We show it is competitive with a strong volumetric baseline on distribution-level quality while exposing direct control over clinical attributes. Flow matching trains the model so its outputs reproduce the data distribution for each conditioning vector, an aggregate property that holds over many samples but is not guaranteed for any single one. The objective rewards overall realism and never checks whether an individual generated volume exhibits its requested findings, so a sample can look realistic yet under-express or omit one. The gap is observable post hoc: a classifier trained on real volumes reads requested attributes from generated samples less reliably than from real ones, but the likelihood objective cannot target it. Reinforcement learning (RL) optimizes it directly, maximizing an explicit reward on the model’s own samples. Group-relative policy optimization has recently been adapted from language models to flow-matching image generators [10, 20]; we adapt it to 3D structured-attribute medical synthesis as a post-training stage that improves conditioning faithfulness, verified by an independent judge model. Contributions. • A controllable 3D latent rectified-flow model for chest CT. A 3D convolutional VAE and a single-stream rectified-flow transformer conditioned directly on structured radiological metadata through adaLN modulation, with synthesis quality competitive with a strong 3D CT diffusion baseline (Sec. 3.1–3.2, Sec. 4.1). • RL post-training that improves conditioning faithfulness. A group-relative policy optimization stage that fine-tunes the flow model against a faithfulness reward, improving how reliably requested findings appear in generated volumes. The gain is verified by an independent held-out judge; to our knowledge it is the first GRPO post-training of a 3D medical flow model (Sec. 3.3, Sec. 4.3). • A released model and synthetic dataset. The trained model together with a ∼200,000 200,000-volume controllable, faithfulness-optimized synthetic chest-CT dataset whose conditioning metadata spans a wide variety of findings (Sec. 4.4). 2 Related Work Latent diffusion compresses volumes with an autoencoder and learns a diffusion or flow model over the latent representation [15]. The image-generation architecture has converged on the diffusion transformer [13], trained under rectified-flow / flow-matching objectives [11, 9, 3] with adaptive-normalization conditioning, as in recent open systems [2]. In 3D medical imaging, conditional latent diffusion has been applied to brain MRI [14, 17] and to whole-body and chest CT [5, 18, 6]; MAISI, a 3D CT latent diffusion model with mask- and metadata-based conditioning [5], is our reference competitor for synthesis quality (Sec. 4.1). Treating denoising as a multi-step decision process enables policy-gradient fine-tuning of diffusion models against non-differentiable rewards [1, 4]. Group-relative policy optimization (GRPO) [16], which replaces a learned value function with a group-relative advantage, was adapted to flow-matching generators by Flow-GRPO [10] via an ODE-to-SDE conversion, and to several visual generation settings by DanceGRPO [20]; subsequent work studies the stability of the importance ratio in this regime [19]. Faithfulness, the agreement between requested and realized attributes, is commonly measured by scoring generated samples with a classifier trained on real data [14]. 3 Method CONFLUX is a three-stage latent diffusion model (Fig. 1): a VAE compresses a CT volume to a low-resolution latent representation (Sec. 3.1); a single-stream rectified-flow transformer generates in that latent space (Sec. 3.2); and a third stage post-trains this model to improve its faithfulness to the requested conditioning (Sec. 3.3). Figure 1: CONFLUX architecture. Stage 1: a 3D convolutional VAE (f=8f=8, C=16C=16; 216×176×200→16×27×22×25216×176×200\!→\!16×27×22×25) encodes a CT volume to a latent representation and is frozen thereafter. Stage 2: a single-stream rectified-flow transformer (L=12L=12, d=768d=768, patch size p=2p=2) is trained from scratch over the normalized latent space by flow matching; patchified tokens carry 3D axial RoPE and are modulated block-wise by the structured metadata vector c (Eq. (1)) through adaLN-zero, and a sample is drawn by integrating the Euler ODE from t=1t=1 to t=0t=0 and decoding with the frozen Stage-1 decoder. Stage 3: GRPO post-training draws a group of rollouts per prompt with the stochastic sampler (Eq. (2)), scores each by the frozen findings-reward classifier (Eq. (3)), and turns the group-relative advantages into a clipped, KL-anchored policy update against the frozen pre-RL reference. 3.1 Latent autoencoder Generating directly in voxel space is costly, so the first stage learns a compact latent representation in which the flow model operates. A 3D convolutional variational autoencoder encodes each preprocessed volume ∈ℝ1×D×H×W x ^1× D× H× W into a diagonal-Gaussian latent ∈ℝC×D/f×H/f×W/f z ^C× D/f× H/f× W/f (mean and per-element variance from an encoder E), downsampling by f=8f=8 per spatial axis to C=16C=16 channels; a deeper decoder G reconstructs ^=G() x=G( z). Both use group-normalized residual blocks with a 3D self-attention block at the lowest resolution, the encoder–decoder asymmetry following latent-diffusion autoencoders [15]. Training minimizes an ℓ1 _1 reconstruction loss, a Kullback–Leibler regularizer toward (,)N( 0, I) (weight 10−610^-6), and a tri-planar LPIPS perceptual loss [21] averaged over slices along the three canonical planes (LPIPS is 2D). After training E,GE,G are frozen. Rectified-flow training assumes near-unit-variance targets, so the flow model operates on the normalized latent ~=(−m)s z=( z-m)\,s with a single global scale s and shift m estimated from the training latents, inverted as =~/s+m z= z/s+m before decoding or scoring (values in Table 4). 3.2 Conditional rectified-flow transformer The second stage is a single-stream transformer that generates a latent embedding from a Gaussian noise sample. Under the rectified-flow formulation, a data latent ~0 z_0 and noise ϵ∼(,) ε ( 0, I) are joined by the straight-line path ~t=(1−t)~0+tϵ z_t=(1-t)\, z_0+t\, ε (t=0t=0 data, t=1t=1 noise), travelled at the constant velocity ϵ−~0 ε- z_0. The network (~t,t,) v_ θ( z_t,t, c) is trained to predict this velocity by minimizing its expected squared error against the target ϵ−~0 ε- z_0, and at sampling time the predicted velocity field is integrated from noise back to a data latent. The latent volume is patchified into tokens with 3D axial rotary position embeddings; conditioning enters through adaptive layer-normalization (adaLN-zero). The conditioning vector concatenates the available structured metadata, =[find∈0,118,csex∈0,1,age∈Δ6,ker∈Δ15]∈ℝ42, c= [\, c_find∈\0,1\^18,\;c_sex∈\0,1\,\; c_age∈ ^6,\; c_ker∈ ^15\, ] ^42, (1) i.e. 1818 binary findings, sex, a one-hot age decade (Δk ^k the k-simplex vertices), and a one-hot reconstruction kernel. For classifier-free guidance [8] c is dropped to a null embedding with probability 0.10.1 during training; sampling integrates the probability-flow ODE d~=dtd z= v_ θ\,dt from t=1t=1 to t=0t=0 on a time-shifted Euler grid. Training uses precomputed latent moments, so E is never evaluated in stage two. 3.3 Reinforcement-learning post-training Flow matching trains the generator to match the data distribution, but supplies no signal that a particular requested attribute is realized in a given sample. The post-training stage adds that signal: it samples volumes from the model, scores how well each matches its requested conditioning, and updates the model to increase that agreement, adapting Flow-GRPO [10] to this setting. Policy-gradient updates require a probability for each sampling step, which the deterministic ODE sampler does not provide, so we replace it with a stochastic sampler that preserves the model’s marginals while injecting noise at each step, ~t+Δt=~t+Δt(~t,t)+σ(t)|Δt|,=+σ(t)22t(~t+(1−t)), z_t+ t= z_t+ t\, d( z_t,t)+σ(t) t \; ξ, d= v_ θ+ σ(t)^22t ( z_t+(1-t) v_ θ ), (2) (Δt<0 t<0, action noise ∼(,) ξ ( 0, I)). Each step is now a Gaussian draw, giving a T-step rollout a tractable log-probability. Its per-step term is averaged over the d≈2.4×105d≈2.4×10^5 latent dimensions rather than summed: summation over so many dimensions inflates the ratio between the updated and sampling policies and destabilizes training, a known failure mode of flow-model RL. The reward measures how faithfully a generated volume realizes its requested findings, read by a frozen classifier fϕf_φ, r(~0,)=−∑g∈find,sex,age,kerωgℓg(fϕ(),g),=~0/s+m,r( z_0, c)=-\!\! _g∈\find,sex,age,ker\\!\! _g\, _g (f_φ( z),\, c_g ), z= z_0/s+m, (3) the negative weighted cross-entropy between predicted and requested conditioning, with the latent un-normalized first because fϕf_φ is trained on raw autoencoder latents. This classifier is a compact 3D convolutional network (group-normalized convolutional blocks, global average pooling, and a linear head over the 1818 findings) trained on real latent representations; the independent judge used in evaluation (Sec. 4.3) shares this architecture but scores decoded volumes at the voxel level. We optimize the findings group only (ωfind=1 _find=1, the others 0); the unweighted groups serve as a specificity check (Sec. 4). For each step we draw G conditioning vectors and N rollouts per vector and form group-relative advantages in the standard way, subtracting each prompt’s mean reward and dividing by the batch-wide standard deviation (more stable than a per-group estimate at the small N affordable here). The policy is updated with the usual clipped PPO surrogate on the per-step importance ratio π/πold _ θ/ _ θ^old, penalized by a Kullback–Leibler divergence to the frozen pre-RL reference πref _ref. With one on-policy epoch the clip is inactive and this penalty is the sole brake on drift; the configuration is given in Table 4. 4 Experiments We use CT-RATE [7], a chest-CT corpus with paired structured metadata and 1818 NLP-extracted abnormality labels. Volumes are cropped to a lung bounding box and resized to 216×176×200216×176×200 with intensities in ∼[−1,1] [-1,1]. A quality-control pass removes scans failing thresholds on lung fraction, left–right balance, and voxel spacing, and removes feet-first and bone-kernel acquisitions; the autoencoder trains on the QC-passing set (∼40,800 40,800 volumes) and the flow model on a one-scan-per-patient subset of 18,41718,417, with 3,0393,039 patient-disjoint volumes held out for validation. Because the 1818 findings are NLP-extracted rather than expert-annotated, faithfulness is measured against these predicted labels. 4.1 Synthesis quality We measure distribution-level quality with a tri-planar 2D-FID (Inception-v3 pool3 features on evenly spaced axial/coronal/sagittal slices, averaged), complemented by density and coverage [12] on the same tri-planar features (k=5k=5; fidelity vs. manifold coverage) and diversity (mean pairwise MS-SSIM among generated volumes) against the real floor. Each method is scored on N=501N=501 generated volumes against the held-out validation reference, with identical lung windowing and seed. Competitors are MAISI [5], a 3D CT latent diffusion model, and GenerateCT [6], a text-conditional 3D chest-CT generator; VAE reconstruction upper-bounds any latent model on our autoencoder. Fig. 2 shows generated volumes across a range of conditioning profiles, with matching coronal and axial views. Figure 2: CONFLUX samples. Generated volumes conditioned on real CT-RATE metadata; coronal (top) and axial (bottom) mid-slices, lung window. (a) F, 4040–5050: cardiomegaly, lung nodule, mosaic attenuation; (b) M, 70+: pleural effusion, opacity, consolidation; (c) M, 2020–3030 and (d) M, 3030–4040: no abnormality; (e) M, 3030–4040: lung nodule, bronchiectasis. Not all listed findings appear in the shown mid-slice. More samples with per-volume conditioning are in Appendix 0.C. Our model leads both baselines (Table 1): tri-planar FID 32.332.3 against 74.674.6 (MAISI) and 145.4145.4 (GenerateCT), approaching the 22.622.6 VAE-reconstruction ceiling. Every FID and density/coverage gap to ours is significant (unpaired bootstrap, 95%95\% CIs exclude zero); our samples are 3.5×3.5× denser and cover 6×6× more of the real manifold than MAISI. Diversity sits near the real floor for all methods (0.510.51 ours vs. 0.410.41 real; collapse would approach 11). Competitors are resampled into our lung space, inflating their cross-plane FID, so the resize-robust axial plane is the cleanest comparison and still favors ours (24.724.7 vs. 55.655.6 and 70.270.2). Table 1: Synthesis quality vs. 3D CT baselines (N=501N=501; ±95%± 95\% bootstrap CI). Tri-planar and axial 2D-FID (↓ ), density and coverage [12] (↑ ), and diversity (pairwise MS-SSIM, real floor 0.410.41). VAE reconstruction is the autoencoder upper bound; best generative result in bold. Method FID↓tri_tri FID↓ax_ax Density ↑ Coverage ↑ Diversity GenerateCT [6] 145.4±2.3145.4±2.3 70.2±3.170.2±3.1 0.008±.0020.008±.002 0.009±.0020.009±.002 0.4860.486 MAISI [5] 74.6±3.374.6±3.3 55.6±2.955.6±2.9 0.016±.0030.016±.003 0.017±.0030.017±.003 0.4710.471 CONFLUX (ours) 32.3±1.932.3 1.9 24.7±1.324.7 1.3 0.056±.0060.056±.006 0.105±.0080.105±.008 0.5070.507 VAE recon (ceiling) 22.6±1.822.6±1.8 18.9±1.318.9±1.3 0.161±.0100.161±.010 0.466±.0150.466±.015 0.4530.453 4.2 Reward classifier The GRPO reward is a frozen classifier that reads the 1818 findings from the latent (Sec. 3.3), so its accuracy bounds the quality of the reward signal. On the real validation split it reaches macro AUROC 0.7930.793 over the 1818 labels, above every CT-CLIP variant reported on the same CT-RATE benchmark [7] (Table 2), outperforming CT-CLIP zero-shot on 17/1817/18 findings. It does so while scoring the 8×8×-compressed latent rather than full-resolution CT, indicating the tokenizer preserves the diagnostic signal. The comparison is not fully controlled (we score a one-scan-per-patient subset from the latent; the baselines, the full validation split at full resolution), but the dataset, labels, findings, and metric are identical. Table 2: Per-finding AUROC on CT-RATE (real validation): our latent classifier vs. CT-CLIP variants and CT-Net [7]. Top 88 findings by our AUROC; macro over all 1818 (full table in Table 5). Best per row in bold. Finding Ours CT-Net CT-CLIPzs_zs CT-CLIPft_ft Cardiomegaly 0.92 0.79 0.86 0.92 Pleural effusion 0.91 0.77 0.90 0.93 Arterial wall calcif. 0.91 0.63 0.85 0.87 Coronary artery calcif. 0.91 0.63 0.85 0.86 Interlobular septal thick. 0.88 0.73 0.77 0.80 Pericardial effusion 0.83 0.65 0.77 0.78 Mosaic attenuation 0.81 0.74 0.77 0.78 Bronchiectasis 0.78 0.57 0.65 0.61 Macro (all 18) 0.793 0.629 0.731 0.756 4.3 Conditioning faithfulness We measure faithfulness as agreement between requested and realized conditioning, read by a classifier on the generated samples. The reward classifier operates in the latent space and is the model optimized by GRPO; because using it for evaluation is circular, we judge instead with an independent image-space classifier trained on decoded real volumes and never used as the reward. We generate from 200200 fixed validation prompts with 44 samples each at 5050 ODE steps and report macro average precision (AP) and macro AUROC over the 1818 findings. Post-training raises both under the independent judge (Table 3): AP 0.330→0.3440.330→0.344 and AUROC 0.684→0.6990.684→0.699. The absolute gains are small, but so is the headroom: the judge reads real CT at only AP 0.3600.360, so the gain already recovers 47%47\% of the base-to-real gap. A paired-by-prompt bootstrap (B=2000B=2000), which cancels prompt-difficulty noise, confirms both gains significant (p=0.042p=0.042 and 0.0140.014). The unweighted groups (sex, age, kernel) stay unchanged within noise, so the gain is specific to the optimized findings. Table 3: Conditioning faithfulness under the independent image-space judge, findings macro over 1818 labels (n=800n=800 generated). The real-data ceiling is the judge scored on real volumes; Δ and its ±1± 1 s.e. come from a paired-by-prompt bootstrap, both gains significant. Headroom recovered is Δ as a fraction of the base-to-ceiling gap. AP ↑ AUROC ↑ Base (pre-RL) 0.3300.330 0.6840.684 CONFLUX 0.3440.344 0.6990.699 Real-data ceiling 0.3600.360 0.7460.746 Δ (CONFLUX −- base) +0.014±0.008+0.014±0.008 +0.015±0.006+0.015±0.006 Headroom recovered 47%47\% 24%24\% 4.4 Model and dataset release We release the trained model and, conditioning it on the real CT-RATE metadata proportions, a ∼200,000 200,000-volume synthetic chest-CT dataset. Each volume is paired with its conditioning vector, spanning a wide variety of findings, for cohort augmentation and conditional study design at a scale unavailable in real corpora. The dataset is available at https://huggingface.co/datasets/gevaertlab/conflux-chest-ct and the model checkpoints at https://huggingface.co/gevaertlab/conflux. 5 Conclusion We presented CONFLUX, a controllable, natively 3D latent rectified-flow model for chest CT, conditioned on structured radiological metadata, that leads strong 3D baselines on quality while giving direct control over clinical attributes. An RL post-training stage raises conditioning faithfulness, the first GRPO post-training of a 3D medical flow model to our knowledge. We release the model and a ∼200,000 200,000-volume synthetic chest-CT dataset. References [1] Black, K., Janner, M., Du, Y., Kostrikov, I., Levine, S.: Training diffusion models with reinforcement learning. In: ICLR (2024) [2] Black Forest Labs: FLUX. https://blackforestlabs.ai (2024), technical report [3] Esser, P., Kulal, S., Blattmann, A., et al.: Scaling rectified flow transformers for high-resolution image synthesis. In: ICML (2024) [4] Fan, Y., et al.: DPOK: Reinforcement learning for fine-tuning text-to-image diffusion models. In: NeurIPS (2023) [5] Guo, P., Zhao, C., Yang, D., et al.: MAISI: Medical AI for synthetic imaging. arXiv preprint arXiv:2409.11169 (2024) [6] Hamamci, I.E., Er, S., et al.: GenerateCT: Text-conditional generation of 3d chest CT volumes. In: ECCV (2024) [7] Hamamci, I.E., et al.: Developing generalist foundation models from a multimodal dataset for 3D computed tomography. arXiv preprint arXiv:2403.17834 (2024) [8] Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS Workshop (2021) [9] Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: ICLR (2023) [10] Liu, J., et al.: Flow-GRPO: Training flow matching models via online RL. arXiv preprint arXiv:2505.05470 (2025) [11] Liu, X., Gong, C., Liu, Q.: Flow straight and fast: Learning to generate and transfer data with rectified flow. In: ICLR (2023) [12] Naeem, M.F., Oh, S.J., Uh, Y., Choi, Y., Yoo, J.: Reliable fidelity and diversity metrics for generative models. In: ICML (2020) [13] Peebles, W., Xie, S.: Scalable diffusion models with transformers. In: ICCV (2023) [14] Pinaya, W.H.L., et al.: Brain imaging generation with latent diffusion models. In: MICCAI Workshop DGM4MICCAI (2022) [15] Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022) [16] Shao, Z., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024) [17] Van Puyvelde, M., Gulluk, I., Van Criekinge, W., Gevaert, O.: BrainG3N: A dual-purpose tokenizer for controllable 3D brain MRI generation. arXiv preprint arXiv:2606.19651 (2026) [18] Wang, H., et al.: 3D MedDiffusion: A 3D medical latent diffusion model for controllable and high-quality medical image generation. arXiv preprint arXiv:2412.13059 (2024) [19] Wang, J., Liang, J., Liu, J., et al.: GRPO-guard: Mitigating implicit over-optimization in flow matching via regulated clipping. arXiv preprint arXiv:2510.22319 (2025) [20] Xue, Z., et al.: DanceGRPO: Unleashing GRPO on visual generation. arXiv preprint arXiv:2505.07818 (2025) [21] Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR (2018) Appendix 0.A Training Parameters Table 4: Training configuration for the three stages. Latent moments are cached; the post-training SDE uses the FLUX time-shifted Euler grid; the reward classifier and reference policy are frozen. Per-channel latent std is 1.271.27–2.482.48. Stage Hyperparameter Value Autoencoder downsampling factor f 88 latent channels C 1616 Flow transformer hidden width d 768768 depth L 1212 attention heads 1212 patch size p 22 latent norm. scale s 0.55090.5509 latent norm. shift m 0.18690.1869 GRPO post-training prompts per step G 22 rollouts per prompt N 1616 SDE steps T 1010 action noise σ(t)σ(t) 0.5t0.5 t PPO clip ϵε 0.20.2 KL weight βKL _KL 0.030.03 inner epochs 11 learning rate 3×10−63×10^-6 Appendix 0.B Per-finding reward-classifier accuracy Table 5: Per-finding AUROC on CT-RATE (real validation), all 1818 findings: our latent classifier vs. CT-CLIP variants and CT-Net [7]. Sorted by our AUROC; best per row in bold. Finding Ours CT-Net CT-CLIPzs_zs CT-CLIPft_ft Cardiomegaly 0.92 0.79 0.86 0.92 Pleural effusion 0.91 0.77 0.90 0.93 Arterial wall calcif. 0.91 0.63 0.85 0.87 Coronary artery calcif. 0.91 0.63 0.85 0.86 Interlobular septal thick. 0.88 0.73 0.77 0.80 Pericardial effusion 0.83 0.65 0.77 0.78 Mosaic attenuation 0.81 0.74 0.77 0.78 Bronchiectasis 0.78 0.57 0.65 0.61 Atelectasis 0.76 0.61 0.68 0.69 Consolidation 0.76 0.66 0.72 0.75 Medical material 0.76 0.66 0.72 0.75 Lung opacity 0.76 0.60 0.63 0.69 Hiatal hernia 0.75 0.55 0.70 0.72 Emphysema 0.73 0.54 0.74 0.75 Lymphadenopathy 0.72 0.61 0.70 0.70 Peribronchial thick. 0.71 0.53 0.69 0.71 Pulmonary fibrotic seq. 0.70 0.52 0.57 0.65 Lung nodule 0.67 0.55 0.57 0.65 Macro (all 18) 0.793 0.629 0.731 0.756 Appendix 0.C Additional samples and conditioning Figure 3: CONFLUX samples. Forty synthetic chest CT volumes drawn at random from the model, each as a coronal (top) and axial (bottom) mid-slice in a lung window, numbered 0–3939; per-patient conditioning is given in Table LABEL:tab:appendix-meta. The range of body habitus, anatomy, and abnormality reflects the diversity and realism of generated volumes; matching coronal/axial views show cross-plane 3D coherence. Table 6: Conditioning metadata for each synthetic volume in Fig. 3. # Sex Age Kernel Conditioned abnormalities 0 F 0–20 Br40f none 1 F 70+ YA Arterial wall calcification, Cardiomegaly, Coronary artery wall calcification, Lymphadenopathy, Atelectasis, Lung nodule, Lung opacity 2 M 60–70 Bl56f Arterial wall calcification, Coronary artery wall calcification, Hiatal hernia, Pulmonary fibrotic sequela, Bronchiectasis 3 M 30–40 YA Lung opacity, Consolidation, Interlobular septal thickening 4 M 30–40 EA Hiatal hernia, Atelectasis 5 F 20–30 Br40f none 6 M 30–40 Br40f Lymphadenopathy, Peribronchial thickening, Consolidation, Bronchiectasis 7 F 20–30 A Atelectasis, Lung nodule 8 M 60–70 YA Arterial wall calcification, Lung opacity 9 F 40–50 YA Arterial wall calcification, Hiatal hernia, Lymphadenopathy, Atelectasis, Lung nodule, Pulmonary fibrotic sequela, Consolidation 10 F 30–40 EA Lymphadenopathy, Lung nodule 11 M 40–50 YA Lymphadenopathy, Lung nodule, Lung opacity 12 M 50–60 YB Lymphadenopathy, Lung nodule 13 F 50–60 L Coronary artery wall calcification, Lung nodule, Lung opacity, Pulmonary fibrotic sequela 14 F 70+ Bl56f Arterial wall calcification, Coronary artery wall calcification, Emphysema, Atelectasis, Lung nodule, Pulmonary fibrotic sequela, Mosaic attenuation pattern 15 F 70+ YA Arterial wall calcification, Coronary artery wall calcification, Atelectasis 16 M 50–60 Bl57d Lymphadenopathy, Lung opacity 17 M 20–30 Br36d none 18 M 20–30 EA none 19 F 40–50 YA Arterial wall calcification, Lymphadenopathy, Emphysema, Lung nodule, Lung opacity, Pulmonary fibrotic sequela 20 M 50–60 YA Lung opacity 21 F 40–50 YA none 22 M 40–50 EA Lymphadenopathy, Lung nodule 23 F 50–60 Br40f Cardiomegaly, Lymphadenopathy, Consolidation 24 F 60–70 YA Arterial wall calcification, Lymphadenopathy, Lung nodule, Pulmonary fibrotic sequela, Mosaic attenuation pattern 25 F 70+ YA Coronary artery wall calcification 26 M 50–60 L Lung opacity, Consolidation 27 M 70+ Bl56f Arterial wall calcification, Coronary artery wall calcification, Atelectasis, Lung opacity, Pulmonary fibrotic sequela 28 F 20–30 YA none 29 M 30–40 EA Lung nodule, Pulmonary fibrotic sequela 30 F 50–60 Br40f Arterial wall calcification, Coronary artery wall calcification, Hiatal hernia, Lung nodule 31 F 20–30 B none 32 F 50–60 YA Lymphadenopathy, Atelectasis, Lung opacity, Pulmonary fibrotic sequela 33 F 60–70 Br36d Lung opacity, Pulmonary fibrotic sequela, Peribronchial thickening, Consolidation, Bronchiectasis, Interlobular septal thickening 34 M 20–30 EA none 35 M 30–40 B Lung nodule 36 F 70+ YA Arterial wall calcification, Coronary artery wall calcification, Emphysema, Consolidation 37 F 40–50 Br40f none 38 M 20–30 Br40f none 39 F 30–40 YA none Figure 4: CONFLUX samples (continued). A further 4040 synthetic chest CT volumes drawn at random, each a coronal (top) and axial (bottom) mid-slice in a lung window, numbered 0–3939; per-patient conditioning is given in Table LABEL:tab:appendix-meta2. Table 7: Conditioning metadata for each synthetic volume in Fig. 4. # Sex Age Kernel Conditioned abnormalities 0 M 40–50 B Medical material, Atelectasis, Lung nodule, Lung opacity 1 F 20–30 YB Lung nodule 2 M 50–60 L Arterial wall calcification, Atelectasis 3 M 20–30 YA Lung opacity, Interlobular septal thickening 4 M 50–60 EA Hiatal hernia, Atelectasis, Lung nodule, Pulmonary fibrotic sequela 5 M 50–60 EA Arterial wall calcification, Lung opacity 6 F 20–30 Bl56f Arterial wall calcification, Cardiomegaly, Pericardial effusion, Coronary artery wall calcification, Hiatal hernia, Lymphadenopathy, Lung nodule 7 M 30–40 Bl57d Emphysema, Atelectasis 8 F 50–60 Br40f none 9 F 30–40 Br40f none 10 M 40–50 Bl56f Hiatal hernia, Lymphadenopathy, Atelectasis, Lung nodule, Bronchiectasis 11 F 30–40 Bl57d none 12 M 50–60 L Lymphadenopathy, Lung opacity 13 M 50–60 EA Lung nodule 14 F 30–40 Br60f Medical material, Emphysema, Pulmonary fibrotic sequela 15 F 30–40 Br40f none 16 M 40–50 B Emphysema, Lung opacity, Consolidation 17 F 70+ YA Emphysema, Atelectasis, Lung nodule, Peribronchial thickening 18 F 60–70 YA Arterial wall calcification, Hiatal hernia, Atelectasis, Lung opacity 19 F 40–50 YA Emphysema 20 F 70+ Br40f Arterial wall calcification, Coronary artery wall calcification, Lymphadenopathy, Lung nodule, Pulmonary fibrotic sequela 21 M 20–30 YA Lung opacity, Peribronchial thickening, Bronchiectasis 22 F 40–50 Br40f Coronary artery wall calcification, Lung nodule 23 M 50–60 EA Medical material, Coronary artery wall calcification, Lymphadenopathy 24 F 50–60 YA Arterial wall calcification, Lymphadenopathy, Emphysema, Atelectasis 25 M 20–30 Br40f Lymphadenopathy, Lung nodule, Lung opacity 26 M 60–70 B Arterial wall calcification, Cardiomegaly, Pericardial effusion, Lymphadenopathy, Emphysema, Atelectasis, Lung opacity, Pulmonary fibrotic sequela, Pleural effusion, Peribronchial thickening, Bronchiectasis 27 F 30–40 Br40f none 28 M 60–70 B Arterial wall calcification, Coronary artery wall calcification, Hiatal hernia, Lung opacity 29 M 30–40 Bl56f Lymphadenopathy, Lung opacity, Bronchiectasis 30 M 40–50 Br40f Medical material, Emphysema, Lung nodule, Pulmonary fibrotic sequela 31 M 20–30 YB none 32 M 20–30 YA Lung nodule 33 M 50–60 YA Hiatal hernia, Lung nodule, Pulmonary fibrotic sequela, Mosaic attenuation pattern, Peribronchial thickening 34 M 30–40 A Lung opacity, Interlobular septal thickening 35 M 40–50 YA Hiatal hernia, Lung opacity 36 M 30–40 Br60f Bronchiectasis 37 F 30–40 Br40f Medical material, Pericardial effusion, Atelectasis, Lung nodule, Pulmonary fibrotic sequela 38 M 30–40 Bl56f none 39 M 50–60 B Lung opacity Figure 5: CONFLUX samples (continued). A further 4040 synthetic chest CT volumes drawn at random, each a coronal (top) and axial (bottom) mid-slice in a lung window, numbered 0–3939; per-patient conditioning is given in Table LABEL:tab:appendix-meta3. Table 8: Conditioning metadata for each synthetic volume in Fig. 5. # Sex Age Kernel Conditioned abnormalities 0 M 30–40 YA Lung opacity 1 F 70+ YA Arterial wall calcification, Coronary artery wall calcification, Atelectasis, Lung nodule, Pulmonary fibrotic sequela 2 M 60–70 YA Arterial wall calcification, Coronary artery wall calcification, Emphysema, Pulmonary fibrotic sequela 3 F 50–60 YA Lung nodule, Lung opacity, Pulmonary fibrotic sequela 4 F 40–50 YA Medical material, Lung nodule 5 F 20–30 YA Pericardial effusion, Pleural effusion, Peribronchial thickening, Consolidation 6 F 70+ Bl56f Arterial wall calcification, Coronary artery wall calcification, Hiatal hernia, Lymphadenopathy, Lung nodule, Lung opacity, Pulmonary fibrotic sequela 7 M 50–60 YA Arterial wall calcification, Coronary artery wall calcification, Lymphadenopathy, Lung nodule, Lung opacity 8 M 40–50 EA Lung nodule, Pulmonary fibrotic sequela 9 M 30–40 EA Emphysema, Lung nodule, Pulmonary fibrotic sequela 10 M 40–50 Br40f Hiatal hernia, Atelectasis 11 F 60–70 YA Lymphadenopathy 12 M 60–70 YA Medical material, Arterial wall calcification, Cardiomegaly, Coronary artery wall calcification, Lymphadenopathy, Emphysema, Lung nodule, Pleural effusion, Mosaic attenuation pattern 13 M 70+ Br40f Arterial wall calcification, Cardiomegaly, Coronary artery wall calcification, Lymphadenopathy, Emphysema, Lung opacity, Pleural effusion, Peribronchial thickening, Consolidation 14 F 60–70 Br60f Cardiomegaly, Lymphadenopathy, Atelectasis 15 M 30–40 EA Lung nodule 16 F 40–50 YA Medical material, Atelectasis, Lung opacity, Consolidation 17 M 40–50 B Coronary artery wall calcification, Pulmonary fibrotic sequela 18 F 60–70 Bl56f Atelectasis, Lung opacity 19 M 30–40 Br40f Lymphadenopathy, Emphysema 20 M 20–30 other none 21 F 70+ Br60f Arterial wall calcification, Coronary artery wall calcification, Hiatal hernia, Lymphadenopathy, Emphysema, Lung nodule, Lung opacity, Consolidation 22 F 50–60 Bl56f Lung nodule 23 F 70+ YA Coronary artery wall calcification, Hiatal hernia, Atelectasis, Pleural effusion 24 F 70+ Br60f Medical material, Arterial wall calcification, Coronary artery wall calcification, Lung nodule, Lung opacity 25 M 40–50 Bl56f Lung nodule, Lung opacity 26 M 50–60 B Arterial wall calcification, Coronary artery wall calcification, Lymphadenopathy, Emphysema, Lung nodule, Pulmonary fibrotic sequela 27 F 30–40 Bl56f Lymphadenopathy, Lung nodule, Pulmonary fibrotic sequela 28 M 70+ YA Arterial wall calcification, Coronary artery wall calcification, Hiatal hernia, Lymphadenopathy, Lung nodule, Lung opacity, Pulmonary fibrotic sequela 29 M 50–60 Bl56f Medical material, Arterial wall calcification, Coronary artery wall calcification, Hiatal hernia, Emphysema, Lung nodule, Pulmonary fibrotic sequela, Bronchiectasis 30 M 50–60 B Medical material, Consolidation, Bronchiectasis 31 F 40–50 YA Lung nodule, Lung opacity 32 M 20–30 Br40f Lung opacity, Consolidation 33 M 30–40 B Lymphadenopathy, Consolidation 34 M 50–60 EA Lung nodule 35 M 30–40 Bl56f Lymphadenopathy, Emphysema, Atelectasis, Lung nodule 36 M 60–70 B Arterial wall calcification, Coronary artery wall calcification, Atelectasis, Lung nodule, Mosaic attenuation pattern 37 M 30–40 Br40f none 38 M 60–70 B Lung nodule, Pulmonary fibrotic sequela 39 F 40–50 YA Emphysema, Lung nodule, Lung opacity, Pulmonary fibrotic sequela