Foundation VAEs for 3D CT Reconstruction, Augmentation, and Generation

Alan Yuille; Jiang Bian; Jingjing Fu; Nan Liu; Qi Chen; Shuhan Ding; Yu Gu; Zongwei Zhou

arxiv: 2605.30893 · v1 · pith:HVJE3UFRnew · submitted 2026-05-29 · 💻 cs.CV

Foundation VAEs for 3D CT Reconstruction, Augmentation, and Generation

Qi Chen , Shuhan Ding , Yu Gu , Nan Liu , Jiang Bian , Alan Yuille , Zongwei Zhou , Jingjing Fu This is my paper

Pith reviewed 2026-06-28 22:47 UTC · model grok-4.3

classification 💻 cs.CV

keywords foundation vaect reconstructionlatent diffusionmedical image generation3d ct volumestraining-free vaevariational autoencoderct augmentation

0 comments

The pith

A single foundation VAE pretrained on natural images and videos reconstructs 3D CT volumes, suppresses noise, and supports generation while keeping its encoder and decoder frozen.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that a variational autoencoder trained at scale on everyday images and videos can act as a ready-made compressor and generator for computed tomography scans of the body. Because the encoder and decoder stay unchanged, the model maps CT volumes into an existing latent space that already captures structure, allowing reconstruction that keeps anatomy intact and reduces scanner noise. Segmentation models trained on these outputs show higher surface accuracy on tumors. The same frozen latent space hosts a conditional diffusion model that produces CT images with better fidelity across many diseases.

Core claim

A single Foundation VAE, pretrained at scale on natural images and videos, can serve as a unified interface for CT Reconstruction, Augmentation, and Generation. With both encoder and decoder frozen, the Foundation VAE reconstructs CT volumes with preserved anatomy while suppressing acquisition noise; training segmentation models on these reconstructions improves surface accuracy by 3.9% NSD on average for pancreatic tumor and lung tumor. Within the same Foundation VAE latent space, a conditional latent diffusion model achieves 3.9% lower average FVD with 36.2% higher CT CLIP score, and improves multi-disease generation faithfulness across 18 types by 2.76% AUC.

What carries the argument

The frozen foundation VAE whose latent space, learned from natural images and videos, maps 3D CT volumes for reconstruction and hosts a conditional diffusion model for generation.

If this is right

Reconstruction from the frozen VAE preserves anatomy and reduces acquisition noise, raising average NSD by 3.9 percent on pancreatic and lung tumor segmentation.
A conditional latent diffusion model placed in the same space lowers average FVD by 3.9 percent and raises CT CLIP score by 36.2 percent.
Multi-disease generation across 18 conditions gains 2.76 percent AUC in faithfulness.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same frozen VAE could be tested on other 3D medical volumes such as MRI without retraining the core model.
Clinical pipelines might insert the reconstruction step before any downstream task to reduce the effect of scanner variability.
Generation quality could be further probed by measuring how well synthetic volumes support rare-disease classification.

Load-bearing premise

Latent representations learned from natural images and videos transfer to 3D CT volumes well enough to preserve clinically relevant anatomy without any changes to the VAE.

What would settle it

A test set of CT volumes from varied scanners and diseases where the frozen VAE reconstructions lose key anatomical landmarks or fail to improve downstream segmentation accuracy.

Figures

Figures reproduced from arXiv: 2605.30893 by Alan Yuille, Jiang Bian, Jingjing Fu, Nan Liu, Qi Chen, Shuhan Ding, Yu Gu, Zongwei Zhou.

**Figure 1.** Figure 1: Foundation VAE for CT Reconstruction, CT Augmentation, and CT Generation. (1) CT Reconstruction: A Foundation VAE pretrained at scale on natural images and videos reconstructs a 3D CT volume via its frozen encoder E and decoder D. (2) CT Augmentation: Through zero shot transfer to CT, the reconstructed volumes provide a boundary enhanced training view, improving downstream segmentation especially on sur… view at source ↗

**Figure 2.** Figure 2: CT reconstruction and segmentation comparison of off-the-shelf video VAEs without medical fine-tuning on MSD (Antonelli et al., 2022) Task06 Lung and Task07 Pancreas. Top row shows reconstructions with segmentation overlays, bottom row shows voxel-wise error maps. Red insets zoom in lesion regions and the blue inset in the Real column shows ground-truth labels. Green contours denote tumors and yellow cont… view at source ↗

**Figure 3.** Figure 3: CT Generation with Foundation VAE. Building on Foundation VAE, healthy and abnormal CT volumes are generated in the fixed latent space of the frozen Foundation VAE. The architecture consists of two parts: (1) Conditioning on Anatomy and Radiology Reports (§ 3.1), where organ and disease masks are encoded by the same frozen Foundation VAE and concatenated with the noised latent at each denoising block for s… view at source ↗

**Figure 4.** Figure 4: Axial, sagittal, and coronal slices of 3D CT volumes synthesized by different methods under the same prompts. Our method generates spatially consistent and anatomically detailed volumes that better align with organ structure and disease-mask guidance. Red arrows indicate correctly synthesized atelectasis, while yellow arrows mark incorrect or spurious findings. disease masks. For the downstream multi-label… view at source ↗

**Figure 5.** Figure 5: Comparison of organ grounding between Real, MAISI, and Ours using pre-trained VISTA3D and TotalSegmentator segmentations. Lung, heart, vessels, and ribs are shown as semitransparent overlays on the center axial, coronal, and sagittal slices. Our method follows the target organ masks more faithfully, producing sharper boundaries and improved alignment for thin structures. 4.2. Quantitative Evaluation of … view at source ↗

**Figure 6.** Figure 6: Controllable CT generation across multiple disease types with targeted disease-mask overlays. The synthesized abnormalities match the specified regions and exhibit disease-consistent appearances, and remaining anatomically coherent [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

**Figure 7.** Figure 7: Qualitative comparison between real (left) and ours (right) across representative slices. Color text highlights key findings in the report. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗

**Figure 8.** Figure 8: Qualitative comparison between real (left) and ours (right) across representative slices. Color text highlights key findings in the report.(Cont.) Bilateral basal segments of the lower lobes with patchy ground-glass opacities and interstitial expansion Nodules in both lungs, the largest measuring approximately 10 mm in diameter in the lower lobe (a) Real (b) Ours [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗

**Figure 9.** Figure 9: Failure cases: qualitative comparisons of representative slices between real CT (left) and ours (right). Colored text highlights key findings in the report. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗

**Figure 10.** Figure 10: Three-view qualitative comparison between real CT and generative models. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗

read the original abstract

Variational autoencoders (VAEs) compress high resolution CT volumes into compact latents while preserving clinically relevant structure. However, training CT-specific VAEs from scratch or heavily fine-tuning them incurs substantial computational and engineering cost, and often degrades under heterogeneous scanners, protocols, and diseases. This paper makes a progressive stride toward training-free medical VAEs by leveraging a critical observation: a single Foundation VAE, pretrained at scale on natural images and videos, can serve as a unified interface for CT Reconstruction, Augmentation, and Generation. With both encoder and decoder frozen, the Foundation VAE reconstructs CT volumes with preserved anatomy while suppressing acquisition noise; training segmentation models on these reconstructions improves surface accuracy by 3.9% NSD on average for pancreatic tumor and lung tumor. Within the same Foundation VAE latent space, a conditional latent diffusion model achieves 3.9% lower average FVD with 36.2% higher CT CLIP score, and improves multi-disease generation faithfulness across 18 types by 2.76% AUC. These results demonstrate Foundation VAEs as a practical interface for scalable CT representation reuse and faithful CT generation. Our code and demo are available at https://github.com/qic999/Foundation-VAE.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Frozen natural-image VAE shows usable CT transfer in the reported numbers, but the no-adaptation premise still needs tighter controls.

read the letter

The paper's main contribution is demonstrating that a VAE pretrained on natural images and videos can be kept entirely frozen and still produce usable 3D CT reconstructions, augmentations, and conditional generations. They report a 3.9% NSD lift on tumor segmentation, plus gains on FVD, CT CLIP, and multi-disease AUC when a latent diffusion model is trained in that space. The practical angle—reusing an off-the-shelf foundation model to avoid training a CT-specific VAE from scratch—is the part worth paying attention to.

What works is the unified framing and the code release. If the numbers hold, it gives a concrete way to cut compute and data needs for medical volumes. The abstract is clear about the frozen encoder/decoder setup, which makes the claim falsifiable in principle.

The soft spots are in the supporting evidence. No error bars appear, baseline details are thin, and the abstract does not spell out how CT intensity ranges or slice spacing are mapped into the natural-image latent space. The central assumption—that directions learned on photos and video separate anatomy from acquisition noise in CT without any adaptation—rests on the downstream metric improvements. Those gains could partly reflect generic denoising rather than the specific transferred latents, and the stress-test note on this point is on target. Without ablations that swap in a randomly initialized or CT-fine-tuned VAE, it is hard to isolate how much the foundation pretraining is doing.

This is aimed at groups already working on latent diffusion or reconstruction for CT who want to test foundation-model reuse. It is coherent enough on its own terms to deserve a serious referee, even though the transfer claim will need more controls before it can be treated as settled.

Referee Report

2 major / 1 minor

Summary. The paper claims that a single Foundation VAE pretrained at scale on natural images and videos can serve as a unified frozen interface for 3D CT reconstruction (preserving anatomy while suppressing noise), augmentation, and generation. With encoder and decoder frozen, it reports average improvements of 3.9% NSD on pancreatic/lung tumor segmentation, 3.9% lower FVD, 36.2% higher CT CLIP score, and 2.76% AUC on multi-disease generation across 18 types; code and demo are released.

Significance. If the frozen transfer holds, the result would enable substantial reuse of large-scale natural-image/video VAEs for medical CT tasks, lowering the cost of domain-specific VAE training and supporting scalable representation learning. Explicit release of code and demo strengthens reproducibility.

major comments (2)

[Abstract] Abstract: the central claim that frozen natural-image/video VAE latents preserve clinically relevant anatomy in CT volumes (different intensity scale, noise statistics, 3D structure) is load-bearing, yet the abstract provides no error bars, baseline details, dataset descriptions, or verification that gains arise from the transferred latents rather than generic denoising/preprocessing.
[Methods/Experiments] The weakest assumption (transfer of latents without any adaptation or fine-tuning) is not secured by an ablation that isolates the VAE mapping from intensity normalization or simple denoising; downstream metric gains could be explained by preprocessing alone.

minor comments (1)

[Abstract] Abstract: the phrase 'progressive stride toward training-free medical VAEs' is slightly overstated given that a conditional latent diffusion model is still trained on the frozen latents.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript to strengthen the presentation and experimental validation.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that frozen natural-image/video VAE latents preserve clinically relevant anatomy in CT volumes (different intensity scale, noise statistics, 3D structure) is load-bearing, yet the abstract provides no error bars, baseline details, dataset descriptions, or verification that gains arise from the transferred latents rather than generic denoising/preprocessing.

Authors: We agree that the abstract is concise and would benefit from additional supporting details. In the revised version we will expand it to report error bars on the key metrics (NSD, FVD, CT CLIP score, AUC), name the primary datasets and baselines, and add a clause noting that the experiments section verifies the contribution of the transferred latents beyond standard preprocessing. revision: yes
Referee: [Methods/Experiments] The weakest assumption (transfer of latents without any adaptation or fine-tuning) is not secured by an ablation that isolates the VAE mapping from intensity normalization or simple denoising; downstream metric gains could be explained by preprocessing alone.

Authors: This is a fair observation. The manuscript already compares against standard intensity-normalization and denoising pipelines and shows additional gains when the frozen foundation VAE is used, but it does not contain a dedicated ablation that applies only those preprocessing steps without the VAE. We will add this targeted ablation in the revision to isolate the VAE mapping more clearly. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical transfer results with no self-referential derivations

full rationale

The paper's central claim rests on empirical evaluation of a frozen pretrained VAE transferred to CT volumes, reporting downstream metrics such as NSD, FVD, and AUC on segmentation and generation tasks. No equations, parameter fits, or uniqueness theorems are invoked that reduce the output to the input by construction. No self-citations serve as load-bearing premises for the transfer assumption, and the work is self-contained against external benchmarks via reported comparisons.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on empirical transfer of natural-image VAE latents to CT without theoretical justification or new entities; the diffusion model training introduces fitted parameters but these are downstream of the frozen VAE claim.

free parameters (1)

conditional latent diffusion model parameters
Parameters of the diffusion model trained in the VAE latent space are fitted to CT data for the generation task.

axioms (1)

domain assumption Pretrained VAE latent space from natural images and videos captures clinically relevant CT structure
Invoked when claiming frozen encoder/decoder suffices for reconstruction and generation without adaptation.

pith-pipeline@v0.9.1-grok · 5769 in / 1209 out tokens · 22751 ms · 2026-06-28T22:47:54.596239+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 13 canonical work pages · 3 internal anchors

[1]

arXiv preprint arXiv:2403.17834 (2024)

Ai, D., Lin, S., Wu, J., Zhu, J., Liu, T., Xu, L., Hu, X., Xie, Y ., Zhang, J., Zhou, Y ., Liu, Y ., and Li, X. Ct-rate: A large-scale dataset of chest ct volumes with paired radiol- ogy reports.arXiv preprint arXiv:2403.17834,

work page arXiv
[2]

E., Zhang, X., Zhu, M., Alabbad, M

Baharoon, M., Luo, L., Moritz, M., Kumar, A., Kim, S. E., Zhang, X., Zhu, M., Alabbad, M. H., Alhazmi, M. S., Mistry, N. P., et al. Rexgroundingct: A 3d chest ct dataset for segmentation of findings from free-text reports.arXiv preprint arXiv:2507.22030,

work page arXiv
[3]

MONAI: An open-source framework for deep learning in healthcare

Cardoso, M. J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y ., Murrey, B., Myronenko, A., Zhao, C., Yang, D., et al. Monai: An open-source framework for deep learning in healthcare.arXiv preprint arXiv:2211.02701,

work page internal anchor Pith review Pith/arXiv arXiv
[4]

Towards generalizable tumor synthesis

Chen, Q., Chen, X., Song, H., Xiong, Z., Yuille, A., Wei, C., and Zhou, Z. Towards generalizable tumor synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 11147–11158, 2024a. Chen, Q., Lai, Y ., Chen, X., Hu, Q., Yuille, A., and Zhou, Z. Analyzing tumors by synthesis.Generative Machine Learning Models in Me...

work page arXiv
[5]

arXiv:2409.11169v2

URL https://arXiv.org/abs/ 2409.11169. arXiv:2409.11169v2. Guo, P., Zhao, C., Yang, D., He, Y ., Nath, V ., Xu, Z., Bassi, P. R., Zhou, Z., Simon, B. D., Harmon, S. A., et al. Text2ct: Towards 3d ct volume generation from free- text descriptions using diffusion model.arXiv preprint arXiv:2505.04522, 2025a. Guo, P., Zhao, C., Yang, D., Xu, Z., Nath, V ., T...

work page arXiv
[6]

He, Y ., Guo, P., Tang, Y ., Myronenko, A., Nath, V ., Xu, Z., Yang, D., Zhao, C., Simon, B., Belue, M., et al

URL https: //arxiv.org/abs/2510.20639. He, Y ., Guo, P., Tang, Y ., Myronenko, A., Nath, V ., Xu, Z., Yang, D., Zhao, C., Simon, B., Belue, M., et al. Vista3d: A unified segmentation foundation model for 3d medical imaging. InProceedings of the Computer Vision and Pattern Recognition Conference, pp. 20863–20873,

work page arXiv
[7]

Kingma, D. P. and Welling, M. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

work page internal anchor Pith review Pith/arXiv arXiv
[8]

R., Xu, D., et al

Li, X., Shuai, Y ., Liu, C., Chen, Q., Wu, Q., Guo, P., Yang, D., Zhao, C., Bassi, P. R., Xu, D., et al. Text-driven tumor synthesis.arXiv preprint arXiv:2412.18589,

work page arXiv
[9]

Medvae: Effi- cient automated interpretation of medical images with large-scale generalizable autoencoders.arXiv preprint arXiv:2502.14753, 2025a

Varma, M., Kumar, A., van der Sluijs, R., Ostmeier, S., Blankemeier, L., Chambon, P., Bluethgen, C., Prince, J., Langlotz, C., and Chaudhari, A. Medvae: Effi- cient automated interpretation of medical images with large-scale generalizable autoencoders.arXiv preprint arXiv:2502.14753, 2025a. Varma, M., Kumar, A., van der Sluijs, R., Ostmeier, S., Blankemei...

work page arXiv 2013
[10]

Wan: Open and Advanced Large-Scale Video Generative Models

Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.-W., Chen, D., Yu, F., Zhao, H., Yang, J., et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314,

work page internal anchor Pith review Pith/arXiv arXiv
[11]

Freetumor: Ad- vance tumor segmentation via large-scale tumor synthesis

Wu, L., Zhuang, J., Ni, X., and Chen, H. Freetumor: Ad- vance tumor segmentation via large-scale tumor synthesis. arXiv preprint arXiv:2406.01264,

work page arXiv
[12]

Large motion video autoencoding with cross-modal video vae.arXiv preprint arXiv:2412.17805,

Xing, Y ., Fei, Y ., He, Y ., Chen, J., Xie, J., Chi, X., and Chen, Q. Large motion video autoencoding with cross-modal video vae.arXiv preprint arXiv:2412.17805,

work page arXiv
[13]

arXiv:2310.03559v? Zhao, S., Zhang, Y ., Cun, X., Yang, S., Niu, M., Li, X., Hu, W., and Shan, Y

URL https://arxiv.org/abs/ 2310.03559. arXiv:2310.03559v? Zhao, S., Zhang, Y ., Cun, X., Yang, S., Niu, M., Li, X., Hu, W., and Shan, Y . Cv-vae: A compatible video vae for latent generative video models.Advances in Neural Information Processing Systems, 37:12847–12871,

work page arXiv
[14]

Relative Latent Size

andW AN2.2(Wan et al., 2025): video autoencoders released with the Wan video model family, used as latent encoders and decoders for large scale video generation. •VideoV AE+(Xing et al., 2024): a cross modal video V AE designed for large motion video autoencoding. • IVV AE(Wu et al., 2025): an improved video V AE that strengthens spatiotemporal reconstruc...

2025
[15]

The diseased subset is defined by its overlap with ReXGround- ingCT (Baharoon et al., 2025)

is our primary dataset. The diseased subset is defined by its overlap with ReXGround- ingCT (Baharoon et al., 2025). The train/test split is re-defined because the original ReXGroundingCT split leaves some disease categories unrepresented in the test set; a strict patient-level split is enforced to prevent any leakage. In total, the generation task compri...

2025
[16]

The synthetic training data use the same text prompts, organ masks, and disease masks

For the downstream classification task, we further sample 500 training volumes from the generative model training set. The synthetic training data use the same text prompts, organ masks, and disease masks. We also sample 200 validation volumes from the original CT-RATE dataset, as disease masks are not required for classification. All models are implement...

2022

[1] [1]

arXiv preprint arXiv:2403.17834 (2024)

Ai, D., Lin, S., Wu, J., Zhu, J., Liu, T., Xu, L., Hu, X., Xie, Y ., Zhang, J., Zhou, Y ., Liu, Y ., and Li, X. Ct-rate: A large-scale dataset of chest ct volumes with paired radiol- ogy reports.arXiv preprint arXiv:2403.17834,

work page arXiv

[2] [2]

E., Zhang, X., Zhu, M., Alabbad, M

Baharoon, M., Luo, L., Moritz, M., Kumar, A., Kim, S. E., Zhang, X., Zhu, M., Alabbad, M. H., Alhazmi, M. S., Mistry, N. P., et al. Rexgroundingct: A 3d chest ct dataset for segmentation of findings from free-text reports.arXiv preprint arXiv:2507.22030,

work page arXiv

[3] [3]

MONAI: An open-source framework for deep learning in healthcare

Cardoso, M. J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y ., Murrey, B., Myronenko, A., Zhao, C., Yang, D., et al. Monai: An open-source framework for deep learning in healthcare.arXiv preprint arXiv:2211.02701,

work page internal anchor Pith review Pith/arXiv arXiv

[4] [4]

Towards generalizable tumor synthesis

Chen, Q., Chen, X., Song, H., Xiong, Z., Yuille, A., Wei, C., and Zhou, Z. Towards generalizable tumor synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 11147–11158, 2024a. Chen, Q., Lai, Y ., Chen, X., Hu, Q., Yuille, A., and Zhou, Z. Analyzing tumors by synthesis.Generative Machine Learning Models in Me...

work page arXiv

[5] [5]

arXiv:2409.11169v2

URL https://arXiv.org/abs/ 2409.11169. arXiv:2409.11169v2. Guo, P., Zhao, C., Yang, D., He, Y ., Nath, V ., Xu, Z., Bassi, P. R., Zhou, Z., Simon, B. D., Harmon, S. A., et al. Text2ct: Towards 3d ct volume generation from free- text descriptions using diffusion model.arXiv preprint arXiv:2505.04522, 2025a. Guo, P., Zhao, C., Yang, D., Xu, Z., Nath, V ., T...

work page arXiv

[6] [6]

He, Y ., Guo, P., Tang, Y ., Myronenko, A., Nath, V ., Xu, Z., Yang, D., Zhao, C., Simon, B., Belue, M., et al

URL https: //arxiv.org/abs/2510.20639. He, Y ., Guo, P., Tang, Y ., Myronenko, A., Nath, V ., Xu, Z., Yang, D., Zhao, C., Simon, B., Belue, M., et al. Vista3d: A unified segmentation foundation model for 3d medical imaging. InProceedings of the Computer Vision and Pattern Recognition Conference, pp. 20863–20873,

work page arXiv

[7] [7]

Kingma, D. P. and Welling, M. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

work page internal anchor Pith review Pith/arXiv arXiv

[8] [8]

R., Xu, D., et al

Li, X., Shuai, Y ., Liu, C., Chen, Q., Wu, Q., Guo, P., Yang, D., Zhao, C., Bassi, P. R., Xu, D., et al. Text-driven tumor synthesis.arXiv preprint arXiv:2412.18589,

work page arXiv

[9] [9]

Medvae: Effi- cient automated interpretation of medical images with large-scale generalizable autoencoders.arXiv preprint arXiv:2502.14753, 2025a

Varma, M., Kumar, A., van der Sluijs, R., Ostmeier, S., Blankemeier, L., Chambon, P., Bluethgen, C., Prince, J., Langlotz, C., and Chaudhari, A. Medvae: Effi- cient automated interpretation of medical images with large-scale generalizable autoencoders.arXiv preprint arXiv:2502.14753, 2025a. Varma, M., Kumar, A., van der Sluijs, R., Ostmeier, S., Blankemei...

work page arXiv 2013

[10] [10]

Wan: Open and Advanced Large-Scale Video Generative Models

Wan, T., Wang, A., Ai, B., Wen, B., Mao, C., Xie, C.-W., Chen, D., Yu, F., Zhao, H., Yang, J., et al. Wan: Open and advanced large-scale video generative models.arXiv preprint arXiv:2503.20314,

work page internal anchor Pith review Pith/arXiv arXiv

[11] [11]

Freetumor: Ad- vance tumor segmentation via large-scale tumor synthesis

Wu, L., Zhuang, J., Ni, X., and Chen, H. Freetumor: Ad- vance tumor segmentation via large-scale tumor synthesis. arXiv preprint arXiv:2406.01264,

work page arXiv

[12] [12]

Large motion video autoencoding with cross-modal video vae.arXiv preprint arXiv:2412.17805,

Xing, Y ., Fei, Y ., He, Y ., Chen, J., Xie, J., Chi, X., and Chen, Q. Large motion video autoencoding with cross-modal video vae.arXiv preprint arXiv:2412.17805,

work page arXiv

[13] [13]

arXiv:2310.03559v? Zhao, S., Zhang, Y ., Cun, X., Yang, S., Niu, M., Li, X., Hu, W., and Shan, Y

URL https://arxiv.org/abs/ 2310.03559. arXiv:2310.03559v? Zhao, S., Zhang, Y ., Cun, X., Yang, S., Niu, M., Li, X., Hu, W., and Shan, Y . Cv-vae: A compatible video vae for latent generative video models.Advances in Neural Information Processing Systems, 37:12847–12871,

work page arXiv

[14] [14]

Relative Latent Size

andW AN2.2(Wan et al., 2025): video autoencoders released with the Wan video model family, used as latent encoders and decoders for large scale video generation. •VideoV AE+(Xing et al., 2024): a cross modal video V AE designed for large motion video autoencoding. • IVV AE(Wu et al., 2025): an improved video V AE that strengthens spatiotemporal reconstruc...

2025

[15] [15]

The diseased subset is defined by its overlap with ReXGround- ingCT (Baharoon et al., 2025)

is our primary dataset. The diseased subset is defined by its overlap with ReXGround- ingCT (Baharoon et al., 2025). The train/test split is re-defined because the original ReXGroundingCT split leaves some disease categories unrepresented in the test set; a strict patient-level split is enforced to prevent any leakage. In total, the generation task compri...

2025

[16] [16]

The synthetic training data use the same text prompts, organ masks, and disease masks

For the downstream classification task, we further sample 500 training volumes from the generative model training set. The synthetic training data use the same text prompts, organ masks, and disease masks. We also sample 200 validation volumes from the original CT-RATE dataset, as disease masks are not required for classification. All models are implement...

2022