NV-Generate-Body: 3D Latent Diffusion Model forControllable Anatomy, Tumor, and Demographic Body Mask Synthesis

Can Zhao; Pengfei Guo; Dong Yang; Yufan He; Yucheng Tang; Ahmed Harouni; Benjamin Simon; Mason Belue

cs.CVOct 16, 2026

NV-Generate-Body: 3D Latent Diffusion Model forControllable Anatomy, Tumor, and Demographic Body Mask Synthesis

Authors: Can Zhao, Pengfei Guo, Dong Yang, Yufan He, Yucheng Tang, Ahmed Harouni, Benjamin Simon, Mason Belue,

Organizations: NVIDIA · National Institutes of Health · University of Arkansas for Medical Sciences

Abstract

Curating large labeled CT datasets is expensive and constrained by patient privacy. We present NV-Generate-BodyMask, a 3D latent diffusion model that synthesizes whole-body CT segmentation masks with precise, joint control over organ anatomy, tumor burden, and patient demographics. The model operates in the compressed latent space of a frozen mask autoencoder, generating $256^3$ volumetric label maps conditioned on a 19-dimensional vector via Adaptive Group Normalization and cross-attention. A size-adaptive ROI loss with spatial halo weighting and a paired tumor injection pipeline enable accurate tumor size conditioning despite the scarcity of annotated tumor cases. Trained on 8,571 masks across 26 cohorts spanning 132 labels, the model achieves Spearman $ρ= 0.70$ -- $0.92$ for organ size, 85--100% tumor recall for lung, pancreatic, colon, and bone lesions, and strong demographic alignment (BMI $ρ= 0.78$ , sex-conditioned perfect prostate control).

Figures & tables

Slot	Field	Type	Range / Encoding
14	Age	continuous	$[0,\,120]$ yr, min-max $\to[0,1]$
15	Sex	categorical	F = 0, M = 1, unknown = $-1$
16	Weight	continuous	$[0,\,200]$ kg, min-max $\to[0,1]$
17	BMI	continuous	$[0,\,75]$ kg/m 2 , min-max $\to[0,1]$
18	Height	continuous	$[0,\,200]$ cm, min-max $\to[0,1]$

Table 1: Demographic conditioning slots [14–18]: field, encoding, and normalization range.

Component / Parameter	v1	v2
Channel multipliers [64, 128, 256, 512]	✓	✓
Residual blocks per level (2)	✓	✓
Self-attention (levels 3 & 4, 32 heads)	✓	✓
Cross-attention	✓	✓
Flash attention	✓	✓
Organ control: liver, pancreas, stomach,

Table 2: Diffusion UNet architecture and conditioning: v1 vs. v2. ✓= present.

Cohort	Cases
NLST	2,447
TCIA_Colon	1,141
StonyBrook-CT	1,103
AbdomenCT-1K	558
NSCLC_Radiogenomics	531
TotalSegmentatorV2	500

Table 3: Training cohorts (top 10 by size).

Tumor type	Crops	$p_{10}$ (mL)	$p_{50}$ (mL)	$p_{90}$ (mL)
Lung tumor	63	1.3	5.2	55.6
Pancreatic tumor	281	1.7	5.8	18.9
Hepatic tumor	615	1.8	18.3	234.6
Colon cancer	126	6.1	21.6	71.9
Bone lesion	112	19.2	64.1	262.7

Table 4: Tumor library statistics.

Variable	$N$	Coverage	Mean	Std	Range
BMI (kg/m 2 )	1,020	11.9%	29.3	6.3	18–63
Weight (kg)	666	7.8%	73.4	17.1	22–144
Age (yr)	930	10.9%	64.1	12.0	19–87
Sex	1,988	23.2%	1,197 M (60.2%) / 791 F (39.8%)

Table 5: Demographic coverage in the training set (total 8,571 cases).

Figure 1: Body envelope construction on a case where three scanner table components account for 41% of the initial body volume. (a) Original CT axial slice. (b) Body mask before table removal: all three components incorrectly included. (c) Body mask after removal of all residual air-density regions inside the body.

Region	$w_{r}^{\text{base}}$
Lung tumor	150
Pancreatic tumor	30
Colon cancer	30
Bone lesion	1.8
Hepatic tumor	2
Pancreas (organ)	3

Table 6: ROI loss base weights

w_{r}^{\text{base}}

Figure 2: Organ size calibration curves (CFG $\in\{0,1,2\}$ ). Each subplot shows generated volume (mL) vs. requested conditioning quantile, with the real-data reference overlaid. Spearman $\rho$ is reported per curve.

Organ	v1	v2
Liver	0.68	0.75
Spleen	0.49	0.80
Stomach	0.17	0.63
Pancreas	0.21	0.47
Left kidney	—	0.69
Right kidney	—	0.72

Table 7: Organ size calibration (Spearman

\rho

, CFG = 2).

\uparrow

higher is better.

	Recall ( $q{=}0.1$ )		Recall ( $q{=}0.9$ )		Size $\rho$
Tumor type	v1	v2	v1	v2	v1	v2
Lung tumor	54%	82%	86%	100%	0.23	0.76
Pancreatic tumor	54%	86%	44%	100%	0.01	0.56
Hepatic tumor	46%	42%	72%	80%	0.13	0.56
Colon cancer	58%	94%	76%	100%	$-$ 0.03	0.56
Bone lesion	94%	100%	100%	100%	0.33	0.70

Table 8: Tumor recall and size calibration.

\uparrow

higher is better. v1 uses CFG = 0 (no CFG); v2 uses CFG = 2.

Figure 3: Tumor recall (left) and size calibration (right) curves (CFG $\in\{0,1,2\}$ ). Left column: recall rate vs. requested conditioning quantile; GT is 100% at all quantiles. Right column: mean generated tumor size (mL) vs. quantile among present cases, with the GT reference overlaid. Spearman $\rho$ is reported per curve.

Metric	CFG = 0	CFG = 2
BMI $\to$ body volume ( $\rho$ )	0.73	0.78
Weight $\to$ body volume ( $\rho$ )	0.85	0.81
Male prostate recall	100%	100%
Female prostate recall	0%	0%

Table 9: Demographic condition alignment.

\uparrow

higher is better.

Figure 4: Organ and tumor size conditioning. Each panel shows coronal, sagittal, and axial views resampled to isotropic spacing; organs are color-coded by label. Within each column, all three samples share the same random seed — only the requested size quantile differs. Left column: Liver at $q{=}0.2$ (small), $0.5$ (median), $0.8$ (large); white contour highlights the liver. Right column: Bone lesion at $q{=}0.2$ , $0.5$ , $0.8$ ; white dashed circle marks the lesion centroid.

Figure 5: Demographic conditioning. Each panel shows coronal, sagittal, and axial views. Row 1: Sex conditioning — Male (prostate highlighted with white contour) vs. Female (no prostate generated). Row 2: BMI conditioning — lean (BMI = 18) vs. obese (BMI = 40), showing visibly different body cross-sections.

Medical Image Synthesis

Medical image synthesis creates, translates, reconstructs, and edits clinical images to fill missing data and support analysis.

See all 310 papers

CONFLUX: A Latent Diffusion Model for 3D Chest-CT Synthesis with RL Post-Training
Jul 3, 2026Max Van Puyvelde, Halil Ibrahim Gulluk, Wim Van Criekinge +1Latent Diffusion Models Medical Image Synthesis
Department of Biomedical Data Science, Stanford University School of Medicine · Department of Mathematical Modelling, Statistics & Bioinformatics, Ghent University · Department of Electrical Engineering, Stanford University
SALIENT: Frequency-Aware Paired Diffusion for Controllable Long-Tail CT Detection
Feb 26, 2026Yifan Li, Mehrdad Salimitari, Taiyu Zhang +2Medical Image Analysis Synthetic Data Augmentation
Trauma Radiology AI Laboratory (TRAIL), Department of Radiology and Nuclear Medicine, University of Maryland School of Medicine · Department of Radiology and Nuclear Medicine, University of Maryland School of Medicine
3D-LLDM: Label-Guided 3D Latent Diffusion Model for Improving High-Resolution Synthetic MR Imaging in Hepatic Structure Segmentation
Mar 25, 2026Kyeonghun Kim, Jaehyeok Bae, Youngung Han +10Medical Image Synthesis 3D Diffusion Models
OUTTA, Republic of Korea · Stanford University, USA · Seoul National University, Republic of Korea +6

Jul 3, 2026J/K move · Enter open · S save

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

Department of Biomedical Data Science, Stanford University School of Medicine · Department of Mathematical Modelling, Statistics & Bioinformatics, Ghent University · Department of Electrical Engineering, Stanford University

Latent Diffusion Models Medical Image Synthesis Conditional Diffusion Models

Open paper AskPDF arXiv

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.

Abstract

Medical Image Synthesis

Explore similar work

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

SALIENT: Frequency-Aware Paired Diffusion for Controllable Long-Tail CT Detection

3D-LLDM: Label-Guided 3D Latent Diffusion Model for Improving High-Resolution Synthetic MR Imaging in Hepatic Structure Segmentation