Pith. sign in

REVIEW 3 major objections 5 minor 72 references

Forecasting subtle brain disease works better when models predict quantized latent change, not future scans.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 10:17 UTC pith:BEY4P3DB

load-bearing objection Solid residual+FSQ package for low-signal MRI forecasting; theory is clean under its margin, but that margin is never measured on the actual latents. the 3 major comments →

arxiv 2607.08270 v1 pith:BEY4P3DB submitted 2026-07-09 cs.CV

Progression as Latent Drift: Generative Forecasting of Slow-Evolving Pathologies

Yuxiang Feng , Juncheng Wang , Chao Xu , Wenlong Hou , Huihan Wang , Yijie Qian , Yang Liu , Baigui Sun
show 2 more authors
Yong Liu Shujun Wan
This is my paper
classification cs.CV
keywords latent driftneurodegenerative forecastingFinite Scalar Quantizationidentity collapsecontinuous interpolation traplongitudinal 3D MRIpatient-specific brain simulationgenerative sequence models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Neurodegeneration changes the brain so slowly that a future MRI looks almost like today’s scan. Direct generative models then either copy the current anatomy (identity collapse) or smear scanner noise across the volume (continuous interpolation trap). This paper argues both failures are structural: baseline anatomy dominates gradients, and smooth networks cannot isolate sparse biological drift from dense nuisance. Latent Drift instead compresses each scan, predicts only the residual change in that latent space, and quantizes the residual with Finite Scalar Quantization so small fluctuations become exactly zero while consistent structural drift survives. On longitudinal 3D brain MRI the approach improves structural agreement with true progression and downstream Alzheimer’s diagnostic accuracy over diffusion and autoregressive baselines. A sympathetic reader cares because earlier, patient-specific forecasts of who will deteriorate, where, and how fast could tighten clinical trials and widen the window for intervention before irreversible damage accumulates.

Core claim

In the low-signal regime of slow-evolving brain pathology, generative forecasting succeeds when the target is the quantized temporal residual (latent drift) rather than the absolute future volume; residual prediction removes stationary anatomy from the objective, and Finite Scalar Quantization acts as a non-Lipschitz dead-zone that annihilates dense nuisance while preserving sparse biological support.

What carries the argument

Latent Drift with Finite Scalar Quantization (FSQ): the model predicts discrete residual tokens Δz_q = Q_h(z_fut − z_cur) instead of z_fut, where the quantizer step is calibrated so noise-only coordinates map to zero and true pathology coordinates do not.

Load-bearing premise

True biological change in the latent space must be stronger than the dense imaging noise so the quantizer’s dead-zone can erase noise without also erasing real atrophy.

What would settle it

If, on held-out multi-site or fast-changing regions (e.g., ventricles), the recovered change ratio falls well below the ideal of 1 while baselines remain closer, or if clinical accuracy collapses once the separation margin is violated, the dead-zone claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Patient-specific MRI forecasts can be produced by adding predicted discrete drift tokens back to a baseline scan rather than synthesizing an entire future volume.
  • Clinical trial enrichment can use forecasted trajectories to prioritize participants likely to show measurable progression within a given horizon.
  • The same residual-plus-dead-zone pattern can be applied to other slow-evolving imaging modalities once a comparable latent encoder exists.
  • Downstream diagnostic models run on the forecasts retain higher accuracy when the generative target is quantized drift rather than absolute anatomy.

Where Pith is reading between the lines

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

  • The same identity-collapse and interpolation pathologies likely appear in any longitudinal medical sequence whose year-scale change is a tiny fraction of total variance, so residual quantization may transfer beyond MRI.
  • Region-aware or adaptive dead-zone grids would be a natural next test if ventricles or other high-dynamic regions systematically under-recover change under a global step size.
  • If the separation margin cannot be guaranteed at acquisition time, explicit noise modeling or multi-site harmonization must precede quantization rather than being left to the dead-zone.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper formalizes generative forecasting of slow-evolving neurodegeneration on longitudinal 3D MRI and argues that standard latent sequence models fail via two modes: identity collapse (optimization dominated by stationary anatomy when predicting absolute future states) and the continuous interpolation trap (Lipschitz predictors cannot isolate sparse biological drift from dense sample-specific noise). It proposes Latent Drift: a residual tokenizer that predicts quantized latent change Δz via Finite Scalar Quantization (FSQ) as a non-Lipschitz dead-zone, followed by an autoregressive transformer that forecasts discrete drift tokens conditioned on baseline anatomy and clinical metadata. Appendix Theorems 3–4 formalize irreducible signal error for Lipschitz interpolators and exact support recovery under a separation margin ε < γ with step h > 2ε. On ADNI/AIBL pairs, the method reports gains over diffusion and AR baselines on Diff-SSIM, NCC, and downstream AD classifier utility (Table 1: Acc 88.33, F1 87.51), with ablations on residual vs pixel targets, quantizers, and FSQ grids.

Significance. If the residual-plus-dead-zone design is the main driver of the reported clinical and structural gains, the work offers a concrete architectural prescription for low-signal longitudinal medical forecasting and a useful failure-mode vocabulary (identity collapse, continuous interpolation trap). Strengths include formal appendix proofs under stated sparsity/margin assumptions, multi-axis evaluation (generative fidelity, Diff-SSIM/NCC, patient-disjoint frozen AD classifier), residual-vs-absolute and quantizer ablations, longitudinal trajectory plots, and region-wise recovered-to-ideal ratios. The contribution is incremental relative to prior residual and latent-progression work but is well-motivated for the sparse-change regime and would be of interest to medical generative modeling if the FSQ margin claim is better grounded.

major comments (3)
  1. The load-bearing claim that FSQ breaks the continuous interpolation trap (and thereby drives clinical gains) rests on Theorem 2 / Appendix A.2: exact support recovery requires min|δj| ≥ γ on the true support, dense nuisance ||η||∞ ≤ ε with ε < γ, and calibration h > 2ε. The manuscript never reports empirical distributions of |Δz_raw| on pathology-supported vs non-supported latent coordinates after Stage-1 encoding, nor the fraction of true-change coordinates that fall inside the dead-zone for the chosen grid [8,8,8,5,5,5]. Without that check, the dead-zone may erase subtle true drift (especially ventricles/fast regions noted in Limitations) or act mainly as ordinary compression; residual targeting alone (Table 2) could explain much of the lift. Please add latent residual histograms or support-recovery diagnostics on held-out pairs, or soften the causal attribution of gains to the non-Lip
  2. Table 1 and §5.2: CycleGAN achieves better FID (i3d/cls.) than Latent Drift while lagging on Diff-SSIM and clinical metrics. The narrative treats this as evidence that competitors reproduce static anatomy, but the paper does not quantify how much of the Diff-SSIM/clinical gap is closed by residual targeting alone versus FSQ versus the AR generator (Tables 2–5 are partial). A fuller factorial (absolute vs residual × continuous vs FSQ × generator family) on the same test split would make the central architectural claim more decisive; currently the clinical superiority is clear but the mechanism attribution remains partly confounded.
  3. §5.1 / clinical protocol: Downstream utility uses a frozen ViViT-style AD classifier (>91% on real scans) on generated futures, with patient-disjoint cohorts. This is a strong external metric, but the paper does not report calibration or failure modes of the classifier on synthetic images (e.g., whether high Acc/F1 can arise from non-pathological intensity shifts that the classifier happens to score as AD). A short sensitivity check—classifier confidence histograms on real vs generated futures, or a secondary volumetric biomarker (hippocampal/ventricular volume change)—would strengthen the claim that forecasts are clinically faithful rather than classifier-exploiting.
minor comments (5)
  1. Fig. 2 caption and surrounding text claim high statistical similarity between current and future states; a quantitative summary (e.g., mean |Δ| / volume variance or SSIM distribution) in the main text would make the low-signal premise more concrete.
  2. Notation: Δz_raw, Δz^q, and s / h for the FSQ step appear with slight inconsistency between Eq. (5) and Appendix A.2; unify the step-size symbol.
  3. Table numbering in the main text refers to “Table 6” for main results while the displayed table is Table 1; renumber consistently.
  4. Limitations correctly flag shared-grid under-representation of fast regions and multi-site deformation; a short quantitative note on how often ventricle change falls near the dead-zone would help readers gauge severity.
  5. Related Work could more explicitly position against Brain Latent Progression and NeuroAR on residual vs absolute targets rather than only listing them.

Circularity Check

0 steps flagged

No circularity: theorems are conditional mathematical statements under explicit assumptions; empirical gains are measured on held-out external metrics without reducing predictions to fitted inputs.

full rationale

The paper's derivation chain is self-contained and non-circular. Identity collapse is motivated by a magnitude argument (||z_cur|| ≫ ||δ_pathology||) plus gradient visualizations (Fig. 3), not by defining the residual target in terms of the claimed win. The Continuous Interpolation Trap (Theorem 1 / Appendix Theorem 3) is a standard Lipschitz-interpolation lower bound: any L-Lipschitz f that interpolates dense η cannot recover sparse support of δ, with R_δ(f) ≳ E||η||_2. FSQ recovery (Theorem 2 / Appendix Theorem 4) is likewise conditional: under the stated separation ε < γ and calibration h > 2ε the dead-zone annihilates noise-only coordinates and preserves supp(δ). These are ordinary mathematical implications, not self-definitional tautologies or fitted parameters renamed as predictions. Empirical claims (Table 1, Diff-SSIM/NCC/F1/Acc) compare generative models on held-out ADNI/AIBL pairs against external classifiers and FID extractors; ablations (Tables 2–5) are ordinary hyper-parameter selection, not circular forecasting of the same quantities used for fitting. No load-bearing self-citation uniqueness theorem, no ansatz smuggled via prior author work, and no renaming of a known empirical pattern. The ε < γ margin is an assumption (correctness risk, not circularity). Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 4 invented entities

The central claim rests on standard learning-theory structure (Lipschitz hypothesis classes, ERM interpolation), domain assumptions about sparse pathology vs dense imaging noise with a hard amplitude margin, and several hand-chosen quantization/training knobs. FSQ itself is prior art; the paper’s load-bearing novelty is applying a calibrated dead-zone to latent residuals. No new physical entity is postulated; the invented pieces are named failure modes and the Latent Drift pipeline.

free parameters (5)
  • FSQ grid configuration = [8,8,8,5,5,5]
    Chosen by ablation; [8,8,8,5,5,5] selected for best clinical Acc/F1 trade-off among several grids (Table 4).
  • FSQ step size / dead-zone half-width h (or s)
    Must satisfy h > 2ε while preserving γ-scale signal; set as part of FSQ design and not independently measured from noise statistics in the main text.
  • Inference sampling temperature = 1.2
    AR decoding temperature set to 1.2 in implementation details; affects generated drift tokens.
  • Tokenizer/generator training hyperparameters = lr 1e-4 / 3e-4; 920 and 70 epochs
    Learning rates, epochs (920 / 70), batch sizes, Adam/AdamW betas, and loss mix (MSE+LPIPS+GAN) are design choices that shape the reported metrics.
  • Spatial resolution and pair horizon filter = 93×112×96; Δt≤48 months
    Volumes downsampled to 93×112×96; pairs restricted to Δt ≤ 48 months—choices that define the evaluation distribution.
axioms (6)
  • domain assumption Future latent decomposes as z_fut = z_cur + δ_pathology + η_nuisance with ||z_cur|| ≫ ||δ|| and η dense, sample-specific.
    Section 3 decomposition and low-signal regime claim (~1% volumetric variance); underpins identity-collapse analysis.
  • standard math Hypothesis class of continuous predictors is L-Lipschitz; ERM interpolators of dense noise cannot recover sparse support of δ.
    Theorem 1 / Appendix Theorem 3; standard Lipschitz + nearest-neighbor generalization argument.
  • domain assumption Biological drift is s-sparse in latent coordinates with min support magnitude γ; nuisance satisfies ||η||_∞ ≤ ε with ε < γ.
    Theorem 2 / Appendix A.2 separation condition; required for exact dead-zone support recovery.
  • standard math A scalar quantizer with h > 2ε annihilates noise-only coordinates and preserves supp(δ).
    Appendix Theorem 4 proof via rounding bins; mathematical once ε,γ,h are granted.
  • ad hoc to paper Pretrained/shared tokenizer latents place anatomically heterogeneous regions on a comparable scale so one global FSQ grid is adequate.
    Invoked when defending uniform FSQ across hippocampus/cerebellum/ventricle (Fig. 6); limitations admit region-aware grids may be needed.
  • domain assumption Downstream frozen AD classifier accuracy is a valid proxy for clinical utility of forecasts.
    Section 5.1 clinical evaluation protocol; patient-disjoint but still one diagnostic model as utility metric.
invented entities (4)
  • Identity collapse (named failure mode) no independent evidence
    purpose: Label optimization dominance of stationary anatomy when predicting absolute future states.
    Framing device for residual targets; not an external physical object. Supported by gradient visualizations (Fig. 3) inside the paper.
  • Continuous interpolation trap (named failure mode) no independent evidence
    purpose: Argue Lipschitz nets smear dense nuisance into dense predictions of Δz.
    Theoretical construct tied to Theorem 1; independent of any new particle/field. Evidence is the proof under stated noise model.
  • Latent Drift pipeline (residual FSQ tokenizer + AR drift generator) no independent evidence
    purpose: Operational architecture that predicts quantized temporal residuals instead of full future anatomy.
    Main methodological object of the paper; evaluated empirically against baselines. Not independently measured outside this work.
  • FSQ as topological dead-zone filter no independent evidence
    purpose: Interpret Finite Scalar Quantization as a non-Lipschitz projection that restores sparse support of pathology.
    FSQ is prior art; the dead-zone filter interpretation and residual application are paper-specific. Independent evidence would require external noise-calibration studies not provided.

pith-pipeline@v1.1.0-grok45 · 25918 in / 4060 out tokens · 46929 ms · 2026-07-10T10:17:50.777836+00:00 · methodology

0 comments
read the original abstract

Forecasting the future anatomy of slow-evolving neurodegenerative diseases could enable earlier, more targeted intervention and improve clinical trial design, but it remains challenging because true progression signals are subtle in longitudinal MRI. In this low-signal regime, transferring modern generative sequence models directly is unreliable: training is dominated by stable baseline anatomy and confounded by dense, sample-specific nuisance variation. We first provide a theoretical analysis that explains these failures through two modes. Identity collapse occurs when optimization is driven toward reproducing the current anatomy, which prevents the model from learning faint temporal change. The continuous interpolation trap arises when standard smooth networks cannot separate localized biological drift from pervasive noise, which leads to spurious changes that diffuse across the volume. To address both issues, we propose Latent Drift, a progressive generative framework that learns change in a compressed semantic representation rather than synthesizing full-resolution anatomy. This design removes pixel-level identity from the prediction target and concentrates model capacity on progression-relevant dynamics. We further apply Finite Scalar Quantization to the learned change representation, which suppresses small, high-frequency nuisance fluctuations while preserving consistent structural drift. Experiments on longitudinal 3D brain MRI show that Latent Drift improves patient-specific neuro-forecasting over diffusion and autoregressive transformer baselines across generative fidelity and clinically relevant evaluation metrics. Project page: \href{https://cutepkq.github.io/latent-drift}{https://cutepkq.github.io/latent-drift}.

Figures

Figures reproduced from arXiv: 2607.08270 by Baigui Sun, Chao Xu, Huihan Wang, Juncheng Wang, Shujun Wan, Wenlong Hou, Yang Liu, Yijie Qian, Yong Liu, Yuxiang Feng.

Figure 1
Figure 1. Figure 1: Progression as Latent Drift. (a) Task Formulation: [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2 [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of gradient-based sig￾nal differences between current and pre￾dicted future states. gradients, trapping the model in an identity function (f(x) ≈ x). Empirical Validation: To empirically validate this source dominance, we trained a baseline continuous sequence model to directly predict zfut conditioned on patient metadata and time horizon. To trace the optimization focus, we visual￾ized the s… view at source ↗
Figure 4
Figure 4. Figure 4: The Progression Latent Drift Framework. Our two-stage architecture ex￾plicitly disentangles sparse biological progression from stationary anatomy and dense imaging noise. First-Stage (Latent Drift Tokenizer): Longitudinal MRI scans are patchified and compressed into continuous latent representations (z). To escape Iden￾tity Collapse, a Drifting Module computes the temporal residual (∆z1 = z2 − z1). Finite … view at source ↗
Figure 5
Figure 5. Figure 5: Personalized Longitudinal Progres￾sion Trajectory. This suggests the model avoids the Identity Collapse seen in competing methods. The high SSIM of the Palette and I2I-DiT trajectories indicates that these mod￾els settle into a near-identity mapping, mostly reproducing stationary anatomy instead of the sparse pathological signal δ. In contrast, the VQGAN [15] and RQ-Transformer [31] baselines drift and flu… view at source ↗
Figure 6
Figure 6. Figure 6: reports the recovered-to-ideal change ratio on three regions of interest that span roughly a 23× range of dynamic change (hippocampus, cerebellum, and ventricle). Our method stays within [0.93, 1.18] of the ideal ratio across all three regions, with a mean absolute deviation of 0.12, which is 2.6 to 5.8× closer to the ideal than the competing baselines. This indicates that the shared grid does not systemat… view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of gradient-based signal differences D.1 Qualitative Comparison Because progression over a typical clinical interval is extremely subtle, predic￾tions from competing methods often look almost identical to the input scan, and differences are difficult to perceive in the raw volumes [PITH_FULL_IMAGE:figures/full_fig_p030_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison of forecasted future scans across methods. Our method captures localized progression consistent with the ground truth, while competing meth￾ods stay close to the input anatomy [PITH_FULL_IMAGE:figures/full_fig_p031_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

72 extracted references · 72 canonical work pages · 3 internal anchors

  1. [1]

    Alzheimer Disease & Associated Disorders26, 17–27 (2012)

    Apostolova, L.G., Green, A.E., Babakchanian, S., Hwang, K.S., Chou, Y.Y., Toga, A.W., Thompson, P.M.: Hippocampal atrophy and ventricular enlargement in nor- mal aging, mild cognitive impairment (MCI), and Alzheimer disease. Alzheimer Disease & Associated Disorders26, 17–27 (2012)

    work page 2012

  2. [2]

    In: Proceedings of the IEEE/CVF international confer- ence on computer vision

    Arnab, A., Dehghani, M., Heigold, G., Sun, C., Lučić, M., Schmid, C.: Vivit: A video vision transformer. In: Proceedings of the IEEE/CVF international confer- ence on computer vision. pp. 6836–6846 (2021)

    work page 2021

  3. [3]

    Alzheimer’s Research & Therapy14(2022)

    Assunção, S.S.M., Sperling, R.A., Ritchie, C.W., Kerwin, D.R., Aisen, P.S., Lansdall, C.J., Atri, A., Cummings, J.L.: Meaningful benefits: a framework to assess disease-modifying therapies in preclinical and early Alzheimer’s disease. Alzheimer’s Research & Therapy14(2022)

    work page 2022

  4. [4]

    The New England journal of medicine367 9, 795–804 (2012)

    Bateman, R.J., Xiong, C., Benzinger, T.L., Fagan, A.M., Goate, A.M., Fox, N.C., Marcus, D.S., Cairns, N.J., Xie, X., Blazey, T.M., Holtzman, D.M., Santacruz, A.M., Buckles, V., Oliver, A., Moulder, K.L., Aisen, P.S., Ghetti, B.F., Klunk, W.E., McDade, E.M., Martins, R.N., Masters, C.L., Mayeux, R.P., Ringman, J.M., Rossor, M.N., Schofield, P.R., Sperling,...

    work page 2012

  5. [5]

    Mechanisms of ageing and de- velopment200, 111575 – 111575 (2021)

    Blinkouskaya, Y., Caçoilo, A., Gollamudi, T., Jalalian, S., Weickenmeier, J.: Brain aging mechanisms with mechanical manifestations. Mechanisms of ageing and de- velopment200, 111575 – 111575 (2021)

    work page 2021

  6. [6]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Brokman, J., Rachmil, O., Hofman, O., Betser, R., Giloni, A., Vainshtein, R., Kojima, H.: Training-free detection of text-to-video generations via over-coherence. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 3993–4003 (2026)

    work page 2026

  7. [7]

    Proceedings of the National Academy of Sciences100(26), 16053–16058 (2003)

    Caldwell, J.C., Miller, M.M., Wing, S., Soll, D.R., Eberl, D.F.: Dynamic analysis of larval locomotion in drosophila chordotonal organ mutants. Proceedings of the National Academy of Sciences100(26), 16053–16058 (2003)

    work page 2003

  8. [8]

    IEEE Transactions on Pattern Analysis and Machine Intelligence45, 10850–10869 (2022)

    Croitoru, F.A., Hondru, V., Ionescu, R.T., Shah, M.: Diffusion models in vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence45, 10850–10869 (2022)

    work page 2022

  9. [9]

    arXiv preprint arXiv:2601.16914 (2026)

    Cui, J., Wu, J., Li, M., Yang, T., Li, X., Wang, R., Bai, A., Ban, Y., Hsieh, C.J.: LoL: Longer than longer, scaling video generation to hour. arXiv preprint arXiv:2601.16914 (2026)

    work page arXiv 2026

  10. [10]

    DenOtter, T.D., Schubert, J.: Hounsfield unit (2019)

    work page 2019

  11. [11]

    2024 46th Annual International Conference of the IEEE En- gineering in Medicine and Biology Society (EMBC) pp

    Dhinagar, N.J., Thomopoulos, S.I., Laltoo, E., Thompson, P.M.: Counterfactual MRI generation with denoising diffusion models for interpretable Alzheimer’s dis- ease effect detection. 2024 46th Annual International Conference of the IEEE En- gineering in Medicine and Biology Society (EMBC) pp. 1–6 (2024)

    work page 2024

  12. [12]

    Academic Press (2015)

    DiStefano III, J.: Dynamic systems biology modeling and simulation. Academic Press (2015)

    work page 2015

  13. [13]

    Alzheimer’s & Dementia12(2016) Progression as Latent Drift 17

    Dubois, B., Hampel, H., Feldman, H., Scheltens, P., Aisen, P.S., Andrieu, S., Bakardjian, H., Benali, H., Bertram, L., Blennow, K., Broich, K., Cavedo, E., Crutch, S.J., Dartigues, J.F., Duyckaerts, C., Epelbaum, S., Frisoni, G.B., Gau- thier, S., Genthon, R., Gouw, A.A., Habert, M.O., Holtzman, D.M., Kivipelto, M., Lista, S., Molinuevo, J.L., O’Bryant, S...

    work page 2016

  14. [14]

    International psychogeriatrics21(4), 672–687 (2009)

    Ellis, K.A., Bush, A.I., Darby, D., De Fazio, D., Foster, J., Hudson, P., Lauten- schlager, N.T., Lenzo, N., Martins, R.N., Maruff, P., et al.: The australian imaging, biomarkers and lifestyle (aibl) study of aging: methodology and baseline character- istics of 1112 individuals recruited for a longitudinal study of alzheimer’s disease. International psych...

    work page 2009

  15. [15]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Esser, P., Rombach, R., Ommer, B.: Taming transformers for high-resolution image synthesis. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12873–12883 (2021)

    work page 2021

  16. [16]

    ArXiv (2025)

    Farki, A., Moradi, E., Koundal, D., Tohka, J.: Forecasting future anatomies: Lon- gitudianl brain MRI-to-MRI prediction. ArXiv (2025)

    work page 2025

  17. [17]

    In: Proceedings of the 33rd International Conference on Neural Information Processing Systems (2019)

    Fazlyab, M., Robey, A., Hassani, H., Morari, M., Pappas, G.J.: Efficient and ac- curate estimation of lipschitz constants for deep neural networks. In: Proceedings of the 33rd International Conference on Neural Information Processing Systems (2019)

    work page 2019

  18. [18]

    Neuroimage54(1), 313–327 (2011)

    Fonov, V., Evans, A.C., Botteron, K., Almli, C.R., McKinstry, R.C., Collins, D.L., Group, B.D.C., et al.: Unbiased average age-appropriate atlases for pediatric stud- ies. Neuroimage54(1), 313–327 (2011)

    work page 2011

  19. [19]

    Medical image analysis105, 103669 (2025)

    Fu, J., Zheng, Y., Dey, N., Ferreira, D., Moreno, R.: Synthesizing individualized aging brains in health and disease with generative models and parallel transport. Medical image analysis105, 103669 (2025)

    work page 2025

  20. [20]

    In: Symposium on Medical Information Process- ing and Analysis (2023)

    Gadewar, S.P., Ramesh, A., Liu, M., Gari, I.B., Nir, T.M., Thompson, P.M., Ja- hanshad, N.: Predicting individual brain MRIs at any age using style encoding generative adversarial networks. In: Symposium on Medical Information Process- ing and Analysis (2023)

    work page 2023

  21. [21]

    International Journal of Computer Vision134(3), 87 (2026)

    Haji-Ali, M., Menapace, W., Siarohin, A., Balakrishnan, G., Ordonez, V.: Taming data and transformers for audio generation. International Journal of Computer Vision134(3), 87 (2026)

    work page 2026

  22. [22]

    Neurology72, 1007 – 999 (2009)

    Henneman, W.J.P., Sluimer, J.D., Barnes, J., van der Flier, W.M., Sluimer, I.C., Fox, N.C., Scheltens, P., Vrenken, H., Barkhof, F.: Hippocampal atrophy rates in Alzheimer disease. Neurology72, 1007 – 999 (2009)

    work page 2009

  23. [23]

    ArXivabs/2511.21114(2025)

    Honga, X., Lin, J., Wang, M.: Deformation-aware temporal generation for early prediction of Alzheimers disease. ArXivabs/2511.21114(2025)

    work page arXiv 2025

  24. [24]

    Whitwell, J., Ward, C., et al.: The alzheimer’s disease neuroimaging initiative (adni): Mri methods

    Jack Jr, C.R., Bernstein, M.A., Fox, N.C., Thompson, P., Alexander, G., Harvey, D., Borowski, B., Britson, P.J., L. Whitwell, J., Ward, C., et al.: The alzheimer’s disease neuroimaging initiative (adni): Mri methods. Journal of Magnetic Reso- nance Imaging: An Official Journal of the International Society for Magnetic Res- onance in Medicine27(4), 685–691 (2008)

    work page 2008

  25. [25]

    Jian,B.,Pan,J.,Li,Y.,Bongratz,F.,Li,R.,Rueckert,D.,Wiestler,B.,Wachinger, C.: TimeFlow: Temporal conditioning for longitudinal brain MRI registration and aging analysis (2025)

    work page 2025

  26. [26]

    Pattern Recognit.133, 109061 (2022)

    Jung, E., Luna, M., Park, S.H.: Conditional GAN with 3D discriminator for MRI generation of Alzheimer’s disease progression. Pattern Recognit.133, 109061 (2022)

    work page 2022

  27. [27]

    Medical image analysis88, 102846 (2022)

    Kazerouni, A., Aghdam, E.K., Heidari, M., Azad, R., Fayyaz, M., Hacihaliloglu, I., Merhof, D.: Diffusion models in medical imaging: A comprehensive survey. Medical image analysis88, 102846 (2022)

    work page 2022

  28. [28]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    KM, M.M., Arun, A., Laskar, Z., Jawahar, C.: PhyEduVideo: A benchmark for evaluating text-to-video models for physics education. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 8690– 8699 (2026) 18 Y. Feng et al

    work page 2026

  29. [29]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2019)

    Kwon, G., Han, C., Kim, D.S.: Generation of 3d brain MRI using auto-encoding generative adversarial networks. In: International Conference on Medical Image Computing and Computer-Assisted Intervention (2019)

    work page 2019

  30. [30]

    American Journal of Roentgenology203(5), 1013–1020 (2014)

    Lamba, R., McGahan, J.P., Corwin, M.T., Li, C.S., Tran, T., Seibert, J.A., Boone, J.M.: Ct hounsfield numbers of soft tissues on unenhanced abdominal ct scans: variability between two different manufacturers’ mdct scanners. American Journal of Roentgenology203(5), 1013–1020 (2014)

    work page 2014

  31. [31]

    In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition

    Lee, D., Kim, C., Kim, S., Cho, M., Han, W.S.: Autoregressive image generation using residual quantization. In: Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition. pp. 11523–11532 (2022)

    work page 2022

  32. [32]

    Neurology 80, 648 – 654 (2013)

    Leung, K.K., Bartlett, J.W., Barnes, J., Manning, E.N., Ourselin, S., Fox, N.C.: Cerebral atrophy in mild cognitive impairment and Alzheimer disease. Neurology 80, 648 – 654 (2013)

    work page 2013

  33. [33]

    arXiv preprint arXiv:2507.16869 (2025) 3

    Ma, Y., Feng, K., Hu, Z., Wang, X., Wang, Y., Zheng, M., Wang, B., Wang, Q., He, X., Wang, H., et al.: Controllable video generation: A survey. arXiv preprint arXiv:2507.16869 (2025)

    work page arXiv 2025

  34. [34]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Ma, Y., He, Y., Cun, X., Wang, X., Chen, S., Li, X., Chen, Q.: Follow your pose: Pose-guided text-to-video generation using pose-free videos. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 4117–4125 (2024)

    work page 2024

  35. [35]

    arXiv preprint arXiv:2506.05207 (2025)

    Ma, Y., Liu, Y., Zhu, Q., Yang, A., Feng, K., Zhang, X., Yan, Z., Li, Z., Han, S., Qi, C., et al.: Follow-your-motion: Video motion transfer via efficient spatial-temporal decoupled finetuning. arXiv preprint arXiv:2506.05207 (2025)

    work page arXiv 2025

  36. [36]

    Alzheimer’s & Dementia21(2025)

    Nakashima, S., Sato, K., Niimi, Y., Toda, T., Iwatsubo, T.: Therapeutic time window of disease-modifying therapy for early Alzheimer’s disease. Alzheimer’s & Dementia21(2025)

    work page 2025

  37. [37]

    NeuroImage p

    Narayanan, S., Nakamura, K., Nakamura, K., Fonov, V.S., Maranzano, J., Cara- manos, Z., Giacomini, P.S., Collins, D.L., Arnold, D.L.: Brain volume loss in in- dividuals over time: Source of variance and limits of detectability. NeuroImage p. 116737 (2020)

    work page 2020

  38. [38]

    Peebles,W.,Xie,S.:Scalablediffusionmodelswithtransformers.In:Proceedingsof the IEEE/CVF international conference on computer vision. pp. 4195–4205 (2023)

    work page 2023

  39. [39]

    International Conference on Medical Image Computing and Computer-Assisted Intervention14227, 14–24 (2022)

    Peng, W., Adeli, E., Zhao, Q., Pohl, K.M.: Generating realistic brain MRIs via a conditional diffusion probabilistic model. International Conference on Medical Image Computing and Computer-Assisted Intervention14227, 14–24 (2022)

    work page 2022

  40. [40]

    Neurobiology of disease p

    Pereira, J.B., Hall, S., Jalakas, M., Grothe, M.J., Strandberg, O., Stomrud, E., Westman, E., van Westen, D., Hansson, O.: Longitudinal degeneration of the basal forebrain predicts subsequent dementia in Parkinson’s disease. Neurobiology of disease p. 104831 (2020)

    work page 2020

  41. [41]

    Brain Imaging Generation with Latent Diffusion Models

    Pinaya, W.H.L., Tudosiu, P.D.D., Dafflon, J., Costa, P.F.D., Fernandez, V., Nachev, P., Ourselin, S., Cardoso, M.J.: Brain imaging generation with latent dif- fusion models. ArXivabs/2209.07162(2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  42. [42]

    Medical Image Analysis84(2021)

    Pombo, G., Gray, R.J., Cardoso, M.J., Ourselin, S., Rees, G.E., Ashburner, J., Nachev, P.: Equitable modelling of brain imaging by counterfactual augmenta- tion with morphologically constrained 3D deep generative models. Medical Image Analysis84(2021)

    work page 2021

  43. [43]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Puglisi, L., Alexander, D.C., Ravì, D.: Enhancing spatiotemporal disease progres- sion models via latent diffusion and prior knowledge. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 173–183. Springer (2024) Progression as Latent Drift 19

    work page 2024

  44. [44]

    Medical image analysis106, 103734 – 103734 (2025)

    Puglisi, L., Alexander, D.C., Ravì, D.: Brain latent progression: Individual-based spatiotemporal disease progression on 3D brain mris via latent diffusion. Medical image analysis106, 103734 – 103734 (2025)

    work page 2025

  45. [45]

    arXiv preprint arXiv:2512.24100 (2025)

    Qian, Y., Wang, J., Feng, Y., Xu, C., Lu, W., Liu, Y., Sun, B., Chen, Y., Liu, Y., Wang, S.: Think before you move: Latent motion reasoning for text-to-motion generation. arXiv preprint arXiv:2512.24100 (2025)

    work page arXiv 2025

  46. [46]

    Biochimica et biophysica acta1822 3, 325–32 (2012)

    Rohrer, J.D.: Structural brain imaging in frontotemporal dementia. Biochimica et biophysica acta1822 3, 325–32 (2012)

    work page 2012

  47. [47]

    2022 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) pp

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. 2022 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) pp. 10674–10685 (2021)

    work page 2022

  48. [48]

    In: ACM SIGGRAPH 2022 confer- ence proceedings

    Saharia, C., Chan, W., Chang, H., Lee, C., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Palette: Image-to-image diffusion models. In: ACM SIGGRAPH 2022 confer- ence proceedings. pp. 1–10 (2022)

    work page 2022

  49. [49]

    In: Proceedings of the 32nd International Conference on Neural Information Processing Systems

    Scaman, K., Virmaux, A.: Lipschitz regularity of deep neural networks: analysis and efficient estimation. In: Proceedings of the 32nd International Conference on Neural Information Processing Systems. p. 3839–3848 (2018)

    work page 2018

  50. [50]

    OUP Oxford (2007)

    Schmidt-Rhaesa, A.: The evolution of organ systems. OUP Oxford (2007)

    work page 2007

  51. [51]

    IEEE Transactions on Medical Imaging39, 1170–1183 (2019)

    Sharma, A., Hamarneh, G.: Missing MRI pulse sequence synthesis using multi- modal generative adversarial network. IEEE Transactions on Medical Imaging39, 1170–1183 (2019)

    work page 2019

  52. [52]

    Progress in neurobiology110, 54–62 (2013)

    de Vugt, M.E., Verhey, F.R.J.: The impact of early dementia diagnosis and inter- vention on informal caregivers. Progress in neurobiology110, 54–62 (2013)

    work page 2013

  53. [53]

    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 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  54. [54]

    arXiv preprint arXiv:2601.14304 (2026)

    Wang, J., Hu, Z., Xu, C., Ren, S., Feng, Y., Liu, Y., Sun, B., Wang, S.: Guided by the plan: Enhancing faithful autoregressive text-to-audio generation with guided decoding. arXiv preprint arXiv:2601.14304 (2026)

    work page arXiv 2026

  55. [55]

    In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing

    Wang, J., Xu, C., Yu, C., Hu, Z., Xie, H., Yu, G., Shang, L., Wang, S.: Language model based text-to-audio generation: Anti-causally aligned collaborative residual transformers. In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. pp. 26036–26054 (2025)

    work page 2025

  56. [56]

    In: Proceed- ings of the Computer Vision and Pattern Recognition Conference

    Wang, J., Xu, C., Yu, C., Shang, L., Hu, Z., Wang, S., Bo, L.: Synchronized video- to-audio generation via mel quantization-continuum decomposition. In: Proceed- ings of the Computer Vision and Pattern Recognition Conference. pp. 3111–3120 (2025)

    work page 2025

  57. [57]

    Frontiers in Aging Neuroscience13(2021)

    Wittens, M.M.J., Allemeersch, G.J., Sima, D.M., Naeyaert, M., Vanderhasselt, T., Vanbinst, A.M., Buls, N., Brucker, Y.D., Raeymaekers, H., Fransen, E.A., Smeets, D., van Hecke, W., Nagels, G., Bjerke, M., de Mey, J., Engelborghs, S.: Inter- and intra-scanner variability of automated brain volumetry on three magnetic res- onance imaging systems in Alzheime...

    work page 2021

  58. [58]

    Scientific Reports10(2020)

    Yang, Q., Li, N., Zhao, Z., Fan, X., Chang, E.I.C., Xu, Y.: MRI cross-modality image-to-image translation. Scientific Reports10(2020)

    work page 2020

  59. [59]

    Neural Autoregressive Modeling of Brain Aging

    Yesiloglu, R., Peng, W., Islam, M.T., Adeli, E.: Neural autoregressive modeling of brain aging. ArXivabs/2507.22954(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  60. [60]

    102595 (2024) 20 Y

    Zhang, Y., Chen, J., Li, Y., Jiao, B., Luo, S.: Disease-modifying therapies for Alzheimer’sdisease:clinicaltrialprogressandopportunity.Ageingresearchreviews p. 102595 (2024) 20 Y. Feng et al

    work page 2024

  61. [61]

    Zhao,P.,Yu,B.:OnmodelselectionconsistencyofLASSO.TheJournalofMachine Learning Research7, 2541–2563 (2006)

    work page 2006

  62. [62]

    In: Proceedings of the IEEE interna- tional conference on computer vision

    Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE interna- tional conference on computer vision. pp. 2223–2232 (2017)

    work page 2017

  63. [63]

    Zoghby,M.,Erickson,B.,Conte,G.:GenerativeadversarialnetworksforbrainMRI synthesis: Impact of training set size on clinical application. Journal of Imaging Informatics in Medicine37, 1228 – 1238 (2024) Progression as Latent Drift 1 A Proof to Theorems A.1 Limits of Lipschitz Generative Models Under Dense Nuisance Noise Problem SetupLet(X,∥ · ∥ 2)be a compa...

    work page 2024

  64. [64]

    (Dense support)P(η j ̸= 0) = 1for allj,

    work page

  65. [65]

    (Bounded continuous amplitude)∥η∥∞ ≤ε,

    work page

  66. [66]

    (Zero mean)E[η] = 0,

    work page

  67. [67]

    LetD n ={(x i, yi)}n i=1 be an i.i.d

    (Independence) Independent ofxand across samples. LetD n ={(x i, yi)}n i=1 be an i.i.d. dataset whereyi =δ(x i) +η i. Weconsiderahypothesisclassofhigh-capacitycontinuousgenerativemodels: FL := f:X →R d ∥f(x)−f(x ′)∥2 ≤L∥x−x ′∥2 . (16) The modelf θ ∈ F L is trained via empirical risk minimization (ERM) onDn. We analyze the expected population signal error:...

    work page

  68. [68]

    (18) In particular, asn→ ∞with a fixed Lipschitz constantL,E[ρ(x)]→0, forcing the error on the true signal to be strictly lower-bounded by the noise amplitude

    (Generalization Lower Bound) Rδ(fθ)≥E∥η∥ 2 −(L+L δ)E[ρ(x)]. (18) In particular, asn→ ∞with a fixed Lipschitz constantL,E[ρ(x)]→0, forcing the error on the true signal to be strictly lower-bounded by the noise amplitude. 2 Y. Feng et al

    work page

  69. [69]

    (19) Consequently,f θ(xtest)is a dense vector on any bounded neighborhood of the training manifold, destroying thes-sparsity ofδ

    (Spurious Continuous Variance) For any test pointxtest satisfyingρ(x test)< |ηi,j | L , the model perfectly hallucinates the noise: fθ,j (xtest)̸= 0. (19) Consequently,f θ(xtest)is a dense vector on any bounded neighborhood of the training manifold, destroying thes-sparsity ofδ. Proof.Part I: Lower Bound on Signal Error. Letx∼ P X be an unseen test point,...

    work page

  70. [70]

    (Annihilation outside support) ∆zq,j (x) = 0∀j /∈S(x). (38)

    work page

  71. [71]

    (Support preservation) supp(∆zq(x)) =S(x). (39)

    work page

  72. [72]

    (40) — ProofFixx∈ X

    (Sparsity restoration) ∥∆zq(x)∥0 ≤s. (40) — ProofFixx∈ X. Progression as Latent Drift 5 Step 1: Coordinates outside the true support.For anyj /∈S(x), we haveδj(x) = 0, hence ∆zj(x) =η j. (41) By boundedness of noise, |∆zj(x)|=|η j| ≤ε. (42) Becauseh >2ε, we have |ηj|< h 2 . (43) Therefore rounding maps the coordinate to zero: Qh(∆zj(x)) = 0. (44) This pro...

    work page arXiv