pith. machine review for the scientific record. sign in

arxiv: 2605.09275 · v1 · submitted 2026-05-10 · 💻 cs.LG

Recognition: 2 theorem links

· Lean Theorem

DiffATS: Diffusion in Aligned Tensor Space

Jinhua Lyu , Tianmin Yu , Brian Kim , Lizhuo Zhou , Chanwook Park , Naichen Shi
Authors on Pith no claims yet

Pith reviewed 2026-05-12 04:34 UTC · model grok-4.3

classification 💻 cs.LG
keywords diffusion modelsTucker decompositionorthogonal Procrustestensor primitivesgenerative modelinglow-rank approximationspatiotemporal datahomeomorphism
0
0 comments X

The pith

Aligning Tucker factor matrices to a medoid anchor via orthogonal Procrustes creates compact, decodable primitive spaces where diffusion models can be trained directly on high-dimensional data.

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

The paper constructs data-dependent primitives by first applying Tucker decomposition to extract a core tensor and mode-wise factors from low-rank multilinear data such as images, videos, and PDE fields. It then selects a medoid anchor and uses orthogonal Procrustes to align all factor matrices, removing the rotational freedom that would otherwise make equivalent representations look different. This alignment produces matrix and tensor Grassmannian primitives that remain faithful to the original tensors. Diffusion models are trained in this smaller space and decoded back through explicit multilinear reconstruction, yielding unconditional and conditional generation results while compressing the data between 3.9 and 210 times without any separate autoencoder. A sympathetic reader would care because the approach removes the usual requirement for large pretrained compressors when scaling diffusion to spatiotemporal fields.

Core claim

We construct data-dependent tensor primitives without pretrained compression autoencoders by starting from Tucker decomposition and applying orthogonal Procrustes alignment of the factor matrices to medoid anchors. This resolves gauge ambiguity and yields matrix Grassmannian primitives and tensor Grassmannian primitives. We prove that the proposed primitive maps are homeomorphisms between low-rank tensors and their corresponding primitive spaces, certifying that the representations are non-degenerate and topologically faithful. Building on these primitives, DiffATS trains diffusion models directly on the aligned tensor primitives and achieves strong unconditional and conditional generation性能

What carries the argument

Orthogonal Procrustes alignment of Tucker factor matrices to a single medoid anchor, which resolves rotational gauge ambiguity and produces unique, reconstructible Grassmannian primitives for direct diffusion training and multilinear decoding.

If this is right

  • Diffusion training and sampling can occur in a space 3.9 to 210 times smaller than the original data while still supporting high-quality unconditional and conditional generation.
  • Generated primitives decode back to full tensors through a fixed multilinear product using the core tensor and aligned factors, without any learned decoder network.
  • The homeomorphism property ensures that topological features of the low-rank tensor manifold are preserved in the primitive space used by diffusion.
  • The same aligned primitive construction applies across images, videos, and solutions to partial differential equations.

Where Pith is reading between the lines

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

  • If the medoid selection remains stable across related datasets, the alignment step could be reused without recomputation when new data arrives.
  • Similar alignment procedures might be developed for other tensor formats such as CP decomposition to obtain analogous compact spaces for generative models.
  • The Grassmannian structure of the primitives invites direct use of geometry-aware sampling methods that respect the manifold metric during diffusion.

Load-bearing premise

The data must admit a sufficiently low multilinear rank structure that Tucker decomposition captures it well, and aligning every factor set to one medoid anchor must fully remove the rotational ambiguity in a way that keeps diffusion training stable and reconstruction accurate.

What would settle it

A dataset of low-rank tensors where samples generated by DiffATS, after multilinear reconstruction from the diffused primitives, show substantially worse fidelity, mode coverage, or distribution statistics than a standard diffusion model trained on the uncompressed original tensors.

Figures

Figures reproduced from arXiv: 2605.09275 by Brian Kim, Chanwook Park, Jinhua Lyu, Lizhuo Zhou, Naichen Shi, Tianmin Yu.

Figure 1
Figure 1. Figure 1: Generation results of DiffATS on CelebA-HQ 1024, Moving MNIST, Kármán vortex street and 2-d Burgers’ equation. “GT” denotes the ground truth. ∗All authors are with Northwestern University. †Corresponding authors: naichen.shi@northwestern.edu. Preprint. arXiv:2605.09275v1 [cs.LG] 10 May 2026 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: Across all settings, DiffATS outperforms DCTDiff [ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline overview of DiffATS. Our main contributions are summarized as follows: • We introduce MGPs and TGPs as data-adaptive, parameter-efficient representations for low-rank matrices and tensors. These primitives retain the expressive power of Tucker decompositions while resolving the rotational ambiguity through anchored OP alignment. • Theoretically, we prove the continuity and bijectivity of the primi… view at source ↗
Figure 3
Figure 3. Figure 3: Effect of OP alignment. (a) Density and learned score fields ∇(U,V ) log pt(U, V ) at diffusion step t = 10 for four sampling distributions of V (arrows normalized within each panel). Pink shading shows the density p10(U, V ), with darker color indicating higher density. (b) Sample KDE of X generated by the trained sθ. (c) Samples generated by DMs trained on CelebA-HQ (1024 × 1024), with and without OP ali… view at source ↗
Figure 4
Figure 4. Figure 4: OP alignment of factor matrices makes diffusion models easier to train. (Left): Without alignment. (Right): With OP alignment. To visualize the dispersion reduction on realis￾tic data, we randomly select N = 1000 images from CelebA-HQ, patchify each into 32 × 32 patches, and compute the SVD of the patch ma￾trix on the first channel, yielding a right singular matrix Vi per image. We select the anchor as the… view at source ↗
Figure 5
Figure 5. Figure 5: Patchification. RMSE ↓ PSNR ↑ (dB) Rank r Patchify No patch Patchify No patch 16 7.55 12.35 31.03 26.50 32 5.52 8.57 33.99 29.80 64 3.47 5.54 38.49 33.83 128 1.74 2.89 45.19 40.01 [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Generated samples on CelebA-HQ at 1024 × 1024. Five samples per method. DiffATS DCTdiff SDIFT AvgPool [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Generated samples on Moving MNIST. Five video samples per method, with each sample shown as a sequence of 5 frames. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Generated samples on 1-d Burgers. Two randomly selected samples per method, visualized as 2-d spatiotemporal renderings. The top row shows the ground truth; subsequent rows show DiffATS and the baselines. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Generated samples on 1-d reaction-diffusion. Two randomly selected samples per method, visualized as 2-d spatiotemporal renderings. The top row shows the ground truth; subsequent rows show DiffATS and the baselines. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Generated samples on 2-d Burgers. Two randomly selected samples per method, visualized as 3-d spatiotemporal renderings. The top row shows the ground truth; subsequent rows show DiffATS and the baselines. DiffATS most faithfully reproduces the spatial structure and temporal evolution of the ground-truth trajectories, consistent with its leading relative error and rMSE in Tab. 12. 26 [PITH_FULL_IMAGE:figu… view at source ↗
Figure 11
Figure 11. Figure 11: Generated samples on 2-d Kármán vortex street. Two randomly selected samples per method, visualized as 5 frames. The top row shows the ground truth; subsequent rows show DiffATS and the baselines. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
read the original abstract

Direct diffusion modeling of high-resolution spatiotemporal fields is computationally challenging. Parameter-efficient primitives address this by representing high-dimensional data with a compact set of parameters. In this paper, we construct data-dependent tensor primitives without pretrained compression autoencoders. Our construction starts from Tucker decomposition, which captures low-rank multilinear structure through a core tensor and mode-wise factors. However, Tucker factors are non-unique: the same tensor can be represented by different rotated factors, which complicates generative modeling. We address this issue with orthogonal Procrustes (OP) alignment. Specifically, we select medoid anchor matrices from the data and align the factor matrices to resolve the gauge ambiguity. This yields matrix Grassmannian primitives and tensor Grassmannian primitives that are compact, data-adaptive, and directly decodable by explicit multilinear reconstruction. Theoretically, we prove that the proposed primitive maps are homeomorphisms between low-rank tensors and their corresponding primitive spaces, certifying that the representations are non-degenerate and topologically faithful. Building on these primitives, we propose *Diffusion in Aligned Tensor Space* (DiffATS), a generative framework that trains diffusion models directly on aligned tensor primitives. Across images, videos, and PDE solutions, DiffATS achieves strong unconditional and conditional generation performance while compressing original data by $3.9\times$ to $210\times$, without relying on any pretrained deep compression autoencoders.

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.

Referee Report

2 major / 3 minor

Summary. The manuscript proposes DiffATS, a generative framework that constructs compact data-dependent tensor primitives via Tucker decomposition followed by orthogonal Procrustes alignment to medoid anchors, thereby resolving gauge ambiguity in the factors. It claims to prove that the resulting maps to matrix Grassmannian primitives and tensor Grassmannian primitives are homeomorphisms, certifying non-degenerate and topologically faithful representations, and reports strong unconditional and conditional generation performance with compression ratios ranging from 3.9× to 210× on images, videos, and PDE solutions without any pretrained deep autoencoders.

Significance. If the homeomorphism claims are rigorously established and the empirical results hold under the stated assumptions, the work offers a meaningful contribution to parameter-efficient diffusion modeling for high-dimensional spatiotemporal data. The explicit multilinear reconstruction, avoidance of pretrained compression models, and reported compression factors represent concrete strengths that could enable more interpretable and scalable generative approaches in scientific and multimedia domains.

major comments (2)
  1. [Theoretical section (homeomorphism proof)] Theoretical section on homeomorphism: The central claim that the primitive maps are homeomorphisms (stated in the abstract and developed in the theory portion) requires the full pipeline—Tucker decomposition plus orthogonal Procrustes alignment to a fixed medoid anchor—to be a continuous bijection with continuous inverse. The Procrustes step solves for the orthogonal matrix via SVD of the cross-product of factor matrices; when singular values are equal or nearly equal (common in video or PDE tensors), the solution is non-unique and the map can exhibit discontinuities under infinitesimal perturbations. The manuscript does not address this degeneracy or provide a proof that continuity holds globally, which directly undermines the certification of 'non-degenerate and topologically faithful' representations needed to justify direct diffusion training on the primitives.
  2. [Theoretical and experimental sections] § on data assumptions and medoid selection: The homeomorphism and stable diffusion training rest on the assumption that the data admits sufficiently low-rank multilinear structure and that alignment to a single dataset medoid fully resolves gauge ambiguity in a topologically faithful manner. No analysis is given of how the choice of medoid (a free parameter) or Tucker core ranks affects continuity or reconstruction quality across the entire space of low-rank tensors, nor is there discussion of whether the construction extends beyond the finite training set. This leaves the global bijectivity claim unsupported.
minor comments (3)
  1. [Abstract and §1] The abstract and introduction introduce 'matrix Grassmannian primitives' and 'tensor Grassmannian primitives' without a concise definition or diagram; adding a short formal definition early would improve accessibility.
  2. [Notation and method sections] Notation for the aligned factor matrices versus the original Tucker factors is used inconsistently in places; a single table summarizing all symbols and their relationships would reduce reader confusion.
  3. [Experiments] Experimental figures showing generated samples would benefit from accompanying quantitative metrics (e.g., FID, PSNR, or reconstruction error) rather than relying solely on visual inspection.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive critique of our theoretical claims. We address each major comment below and will revise the manuscript to strengthen the rigor of the homeomorphism results.

read point-by-point responses

  1. Referee: Theoretical section on homeomorphism: The central claim that the primitive maps are homeomorphisms (stated in the abstract and developed in the theory portion) requires the full pipeline—Tucker decomposition plus orthogonal Procrustes alignment to a fixed medoid anchor—to be a continuous bijection with continuous inverse. The Procrustes step solves for the orthogonal matrix via SVD of the cross-product of factor matrices; when singular values are equal or nearly equal (common in video or PDE tensors), the solution is non-unique and the map can exhibit discontinuities under infinitesimal perturbations. The manuscript does not address this degeneracy or provide a proof that continuity holds globally, which directly undermines the certification of 'non-degenerate and topologically faithful' representations needed to justify direct diffusion training on the primitives.

    Authors: We acknowledge the potential for non-uniqueness in the SVD-based Procrustes solution when singular values coincide. Our existing proof establishes the homeomorphism under the generic assumption that singular values are distinct (a standard technical condition ensuring uniqueness of the orthogonal factor and continuity of the overall map). This assumption holds on an open dense set and is satisfied for the low-rank structures in our image, video, and PDE datasets. We will revise the theoretical section to explicitly state this genericity condition, include a detailed continuity argument under the assumption, and add a brief discussion of practical handling for near-degenerate cases (e.g., via small random perturbations or selection of the closest valid rotation). These changes preserve the core claims while addressing the referee's concern about global continuity. revision: partial

  2. Referee: § on data assumptions and medoid selection: The homeomorphism and stable diffusion training rest on the assumption that the data admits sufficiently low-rank multilinear structure and that alignment to a single dataset medoid fully resolves gauge ambiguity in a topologically faithful manner. No analysis is given of how the choice of medoid (a free parameter) or Tucker core ranks affects continuity or reconstruction quality across the entire space of low-rank tensors, nor is there discussion of whether the construction extends beyond the finite training set. This leaves the global bijectivity claim unsupported.

    Authors: The medoid is defined as the training tensor minimizing the sum of distances to all others after initial Tucker decomposition, serving as a fixed, data-driven anchor. We will add a new subsection analyzing sensitivity to medoid choice and core-rank selection, showing that the aligned primitive map remains a homeomorphism on the manifold of fixed multilinear rank tensors for any choice of medoid (the alignment is always to the same fixed reference). For extension beyond the training set, new tensors undergo Tucker decomposition followed by alignment to the precomputed training medoid; this yields a well-defined map on the entire low-rank tensor space. We will include both theoretical justification of bijectivity under the low-rank assumption and empirical checks on held-out samples to confirm reconstruction fidelity and continuity. These additions directly support the global claims. revision: yes

Circularity Check

0 steps flagged

No circularity; derivation relies on independent multilinear algebra

full rationale

The paper's central chain—Tucker decomposition to obtain factors, followed by explicit orthogonal Procrustes alignment to a fixed medoid anchor, followed by a claimed homeomorphism proof—does not reduce any result to its own inputs by construction. The homeomorphism is asserted as a theorem proven from the maps' definition using standard properties of Tucker decompositions and orthogonal matrices, without fitting parameters to the target diffusion loss or renaming fitted quantities as predictions. No load-bearing self-citation, ansatz smuggling, or uniqueness theorem imported from the authors' prior work appears in the provided derivation; the alignment resolves gauge freedom by a deterministic procedure whose continuity and bijectivity are treated as separately provable rather than tautological.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 1 invented entities

The approach rests on the standard assumption that real-world spatiotemporal data possesses low-rank multilinear structure amenable to Tucker decomposition, plus the paper-specific step of medoid-based Procrustes alignment to remove rotational freedom. No large set of fitted constants is introduced beyond implicit rank choices.

free parameters (2)
  • Tucker core ranks
    The dimensions of the core tensor and factor matrices must be chosen per dataset; these act as hyperparameters that control compression and fidelity.
  • Medoid anchor selection
    Choice of which factor matrices serve as alignment targets is data-dependent and requires a selection procedure.
axioms (2)
  • domain assumption Input data tensors admit useful low-rank multilinear approximations via Tucker decomposition
    Invoked to justify the initial decomposition step before alignment.
  • ad hoc to paper Orthogonal Procrustes alignment to a medoid anchor produces unique, topologically faithful primitives
    This is the key paper-specific assumption that enables direct diffusion on the aligned space.
invented entities (1)
  • Matrix Grassmannian primitives and tensor Grassmannian primitives no independent evidence
    purpose: Compact, data-adaptive, explicitly decodable representations that remove gauge ambiguity for generative modeling
    Newly defined objects obtained after Procrustes alignment; no independent external evidence is supplied in the abstract.

pith-pipeline@v0.9.0 · 5555 in / 1555 out tokens · 68716 ms · 2026-05-12T04:34:03.119396+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

66 extracted references · 66 canonical work pages · 10 internal anchors

  1. [1]

    J., Bambrick, J., et al

    Abramson, J., Adler, J., Dunger, J., Evans, R., Green, T., Pritzel, A., Ronneberger, O., Willmore, L., Ballard, A. J., Bambrick, J., et al. (2024). Accurate structure prediction of biomolecular interactions with alphafold 3.Nature, 630(8016):493–500

    work page 2024

  2. [2]

    Anderson, B. D. (1982). Reverse-time diffusion equation models.Stochastic Processes and their Applications, 12(3):313–326

    work page 1982

  3. [3]

    W., Fidler, S., and Kreis, K

    Blattmann, A., Rombach, R., Ling, H., Dockhorn, T., Kim, S. W., Fidler, S., and Kreis, K. (2023). Align your latents: High-resolution video synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22563–22575

    work page 2023

  4. [4]

    Chen, P., Sun, Y ., Cheng, L., Yang, Y ., Li, W., Liu, Y ., Liu, W., Bian, J., and Fang, S. (2025). Generating full-field evolution of physical dynamics from irregular sparse observations.arXiv preprint arXiv:2505.09284

    work page arXiv 2025

  5. [5]

    and Kahan, W

    Davis, C. and Kahan, W. M. (1970). The rotation of eigenvectors by a perturbation. iii.SIAM Journal on Numerical Analysis, 7(1):1–46

    work page 1970

  6. [6]

    W., and Doucet, A

    De Bortoli, V ., Mathieu, E., Hutchinson, M., Thornton, J., Teh, Y . W., and Doucet, A. (2022). Riemannian score-based generative modelling.Advances in neural information processing systems, 35:2406–2422

    work page 2022

  7. [7]

    De Lathauwer, L., De Moor, B., and Vandewalle, J. (2000a). A multilinear singular value decomposition.SIAM journal on Matrix Analysis and Applications, 21(4):1253–1278

    work page

  8. [8]

    De Lathauwer, L., De Moor, B., and Vandewalle, J. (2000b). On the best rank-1 and rank-(r 1, r 2,..., rn) approximation of higher-order tensors.SIAM journal on Matrix Analysis and Applications, 21(4):1324–1342

    work page

  9. [9]

    and Young, G

    Eckart, C. and Young, G. (1936). The approximation of one matrix by another of lower rank. Psychometrika, 1(3):211–218

    work page 1936

  10. [10]

    A., and Smith, S

    Edelman, A., Arias, T. A., and Smith, S. T. (1998). The geometry of algorithms with orthogo- nality constraints.SIAM journal on Matrix Analysis and Applications, 20(2):303–353

    work page 1998

  11. [11]

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

    work page 2021

  12. [12]

    Gu, S., Chen, D., Bao, J., Wen, F., Zhang, B., Chen, D., Yuan, L., and Guo, B. (2022). Vector quantized diffusion model for text-to-image synthesis. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696–10706

    work page 2022

  13. [13]

    C., Lu, Y ., Wagner, G

    Guo, J., Domel, G., Park, C., Zhang, H., Gumus, O. C., Lu, Y ., Wagner, G. J., Qian, D., Cao, J., Hughes, T. J., et al. (2025a). Tensor-decomposition-based a priori surrogate (taps) modeling for ultra large-scale simulations.Computer Methods in Applied Mechanics and Engineering, 444:118101

    work page

  14. [14]

    Guo, J., Kong, X., Li, Z., and Mao, J. (2026). Tucker diffusion model for high-dimensional tensor generation.arXiv preprint arXiv:2604.00481

    work page arXiv 2026

  15. [15]

    J., Domel, G., and Liu, W

    Guo, J., Xie, X., Park, C., Zhang, H., Politis, M. J., Domel, G., and Liu, W. K. (2025b). Interpolating neural network-tensor decomposition (inn-td): a scalable and interpretable approach for large-scale physics-based problems. InInternational Conference on Machine Learning, pages 21138–21162. PMLR

    work page

  16. [16]

    He, Y ., Yang, T., Zhang, Y ., Shan, Y ., and Chen, Q. (2022). Latent video diffusion models for high-fidelity long video generation.arXiv preprint arXiv:2211.13221

    work page internal anchor Pith review arXiv 2022

  17. [17]

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S. (2017). Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems, 30. 10

    work page 2017

  18. [18]

    Ho, J., Jain, A., and Abbeel, P. (2020). Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851

    work page 2020

  19. [19]

    J., Norouzi, M., and Salimans, T

    Ho, J., Saharia, C., Chan, W., Fleet, D. J., Norouzi, M., and Salimans, T. (2022a). Cascaded diffusion models for high fidelity image generation.Journal of Machine Learning Research, 23(47):1–33

    work page

  20. [20]

    Ho, J., Salimans, T., Gritsenko, A., Chan, W., Norouzi, M., and Fleet, D. J. (2022b). Video diffusion models.Advances in neural information processing systems, 35:8633–8646

    work page

  21. [21]

    G., Vignac, C., and Welling, M

    Hoogeboom, E., Satorras, V . G., Vignac, C., and Welling, M. (2022). Equivariant diffusion for molecule generation in 3d. InInternational conference on machine learning, pages 8867–8887. PMLR

    work page 2022

  22. [22]

    G., Frossard, P., Welling, M., Bronstein, M., and Correia, B

    Igashov, I., Stärk, H., Vignac, C., Schneuing, A., Satorras, V . G., Frossard, P., Welling, M., Bronstein, M., and Correia, B. (2024). Equivariant 3d-conditional diffusion model for molecular linker design.Nature Machine Intelligence, 6(4):417–427

    work page 2024

  23. [23]

    Karras, T., Aila, T., Laine, S., and Lehtinen, J. (2017). Progressive growing of gans for improved quality, stability, and variation.arXiv preprint arXiv:1710.10196

    work page internal anchor Pith review arXiv 2017

  24. [24]

    Karras, T., Aittala, M., Aila, T., and Laine, S. (2022). Elucidating the design space of diffusion- based generative models.Advances in neural information processing systems, 35:26565–26577

    work page 2022

  25. [25]

    Kingma, D. P. and Welling, M. (2013). Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  26. [26]

    Koehler, F., Niedermayr, S., Westermann, R., and Thuerey, N. (2024). Apebench: A benchmark for autoregressive neural emulators of pdes.Advances in Neural Information Processing Systems, 37:120252–120310

    work page 2024

  27. [27]

    Kolda, T. G. and Bader, B. W. (2009). Tensor decompositions and applications.SIAM review, 51(3):455–500

    work page 2009

  28. [28]

    Kong, Z., Ping, W., Huang, J., Zhao, K., and Catanzaro, B. (2020). Diffwave: A versatile diffusion model for audio synthesis.arXiv preprint arXiv:2009.09761

    work page arXiv 2020

  29. [29]

    Kruskal, J. B. and Wish, M. (1978).Multidimensional scaling. Number 11. Sage

    work page 1978

  30. [30]

    Kunisky, D. (2023). Generic manova limit theorems for products of projections.arXiv preprint arXiv:2301.09543

    work page arXiv 2023

  31. [31]

    Kynkäänniemi, T., Karras, T., Laine, S., Lehtinen, J., and Aila, T. (2019). Improved precision and recall metric for assessing generative models.Advances in neural information processing systems, 32

    work page 2019

  32. [32]

    Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., and Anandku- mar, A. (2020). Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  33. [33]

    Lienen, M., Lüdke, D., Hansen-Palmus, J., and Günnemann, S. (2023). From zero to turbulence: Generative modeling for 3d flow simulation.arXiv preprint arXiv:2306.01776

    work page arXiv 2023

  34. [34]

    Liu, H., Chen, Z., Yuan, Y ., Mei, X., Liu, X., Mandic, D., Wang, W., and Plumbley, M. D. (2023). Audioldm: Text-to-audio generation with latent diffusion models.arXiv preprint arXiv:2301.12503

    work page arXiv 2023

  35. [35]

    Liu, Y ., Zhang, K., Li, Y ., Yan, Z., Gao, C., Chen, R., Yuan, Z., Huang, Y ., Sun, H., Gao, J., et al. (2024). Sora: A review on background, technology, limitations, and opportunities of large vision models.arXiv preprint arXiv:2402.17177

    work page internal anchor Pith review arXiv 2024

  36. [36]

    and Fattahi, S

    Ma, J. and Fattahi, S. (2024). Convergence of gradient descent with small initialization for unregularized matrix completion. InThe Thirty Seventh Annual Conference on Learning Theory, pages 3683–3742. PMLR. 11

    work page 2024

  37. [37]

    Ma, X., Wang, Y ., Chen, X., Jia, G., Liu, Z., Li, Y .-F., Chen, C., and Qiao, Y . (2024). Latte: Latent diffusion transformer for video generation.arXiv preprint arXiv:2401.03048

    work page internal anchor Pith review arXiv 2024

  38. [38]

    W., Maggioni, M., and Drineas, P

    Mahoney, M. W., Maggioni, M., and Drineas, P. (2006). Tensor-cur decompositions for tensor- based data. InProceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 327–336

    work page 2006

  39. [39]

    arXiv: 2412.15032[cs.CV]

    Ning, M., Li, M., Su, J., Jia, H., Liu, L., Beneš, M., Chen, W., Salah, A. A., and Ertugrul, I. O. (2024). Dctdiff: Intriguing properties of image generative modeling in the dct space.arXiv preprint arXiv:2412.15032

    work page arXiv 2024

  40. [40]

    Oquab, M., Darcet, T., Moutakanni, T., V o, H., Szafraniec, M., Khalidov, V ., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al. (2023). Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  41. [41]

    and Xie, S

    Peebles, W. and Xie, S. (2023). Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205

    work page 2023

  42. [42]

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., and Rombach, R. (2023). Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  43. [43]

    arXiv:2312.15796 [physics]

    Price, I., Sanchez-Gonzalez, A., Alet, F., Andersson, T. R., El-Kadi, A., Masters, D., Ewalds, T., Stott, J., Mohamed, S., Battaglia, P., et al. (2023). Gencast: Diffusion-based ensemble forecasting for medium-range weather.arXiv preprint arXiv:2312.15796

    work page arXiv 2023

  44. [44]

    J., Mohamed, S., and Wierstra, D

    Rezende, D. J., Mohamed, S., and Wierstra, D. (2014). Stochastic backpropagation and approximate inference in deep generative models. InInternational conference on machine learning, pages 1278–1286. PMLR

    work page 2014

  45. [45]

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. (2022). High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695

    work page 2022

  46. [46]

    and Štefankoviˇc, D

    Schaefer, M. and Štefankoviˇc, D. (2018). The complexity of tensor rank.Theory of Computing Systems, 62(5):1161–1174

    work page 2018

  47. [47]

    Schönemann, P. H. (1966). A generalized solution of the orthogonal procrustes problem. Psychometrika, 31(1):1–10

    work page 1966

  48. [48]

    Shu, D., Li, Z., and Farimani, A. B. (2023). A physics-informed diffusion model for high-fidelity flow field reconstruction.Journal of Computational Physics, 478:111972

    work page 2023

  49. [49]

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. (2015). Deep unsupervised learning using nonequilibrium thermodynamics. InInternational conference on machine learning, pages 2256–2265. pmlr

    work page 2015

  50. [50]

    Song, J., Meng, C., and Ermon, S. (2020a). Denoising diffusion implicit models. InInternational Conference on Learning Representations

    work page

  51. [51]

    and Ermon, S

    Song, Y . and Ermon, S. (2019). Generative modeling by estimating gradients of the data distribution.Advances in neural information processing systems, 32

    work page 2019

  52. [52]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Song, Y ., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. (2020b). Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  53. [53]

    Srivastava, N., Mansimov, E., and Salakhudinov, R. (2015). Unsupervised learning of video representations using lstms. InInternational conference on machine learning, pages 843–852. PMLR

    work page 2015

  54. [54]

    L., Taylor, E., and Loaiza-Ganem, G

    Stein, G., Cresswell, J., Hosseinzadeh, R., Sui, Y ., Ross, B., Villecroze, V ., Liu, Z., Caterini, A. L., Taylor, E., and Loaiza-Ganem, G. (2023). Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models.Advances in Neural Information Processing Systems, 36:3732–3784. 12

    work page 2023

  55. [55]

    Szegedy, C., Vanhoucke, V ., Ioffe, S., Shlens, J., and Wojna, Z. (2016). Rethinking the inception architecture for computer vision. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2818–2826

    work page 2016

  56. [56]

    Takamoto, M., Praditia, T., Leiteritz, R., MacKinlay, D., Alesiani, F., Pflüger, D., and Niepert, M. (2022). Pdebench: An extensive benchmark for scientific machine learning.Advances in neural information processing systems, 35:1596–1611

    work page 2022

  57. [57]

    Tu, S., Boczar, R., Simchowitz, M., Soltanolkotabi, M., and Recht, B. (2016). Low-rank solutions of linear matrix equations via procrustes flow. InInternational conference on machine learning, pages 964–973. PMLR

    work page 2016

  58. [58]

    Tucker, L. R. (1966). Some mathematical notes on three-mode factor analysis.Psychometrika, 31(3):279–311

    work page 1966

  59. [59]

    Unterthiner, T., Van Steenkiste, S., Kurach, K., Marinier, R., Michalski, M., and Gelly, S. (2018). Towards accurate generative models of video: A new metric & challenges.arXiv preprint arXiv:1812.01717

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  60. [60]

    Van Den Oord, A., Vinyals, O., et al. (2017). Neural discrete representation learning.Advances in neural information processing systems, 30

    work page 2017

  61. [61]

    Xu, M., Yu, L., Song, Y ., Shi, C., Ermon, S., and Tang, J. (2022). Geodiff: A geometric diffusion model for molecular conformation generation.arXiv preprint arXiv:2203.02923

    work page arXiv 2022

  62. [62]

    Xu, Y ., Wang, Y ., Luo, S., Gao, K., He, T., He, D., and Liu, C. (2026). Quotient-space diffusion models. InThe Fourteenth International Conference on Learning Representations

    work page 2026

  63. [63]

    Zeni, C., Pinsler, R., Zügner, D., Fowler, A., Horton, M., Fu, X., Wang, Z., Shysheya, A., Crabbé, J., Ueda, S., et al. (2025). A generative model for inorganic materials design.Nature, 639(8055):624–632

    work page 2025

  64. [64]

    and Xia, D

    Zhang, A. and Xia, D. (2018). Tensor svd: Statistical and computational limits.IEEE Transac- tions on Information Theory, 64(11):7311–7338

    work page 2018

  65. [65]

    Zhang, X., Du, S., and Gu, Q. (2018). Fast and sample efficient inductive matrix completion via multi-phase procrustes flow. InInternational Conference on Machine Learning, pages 5756–5765. PMLR

    work page 2018

  66. [66]

    fDOLwOcAA+yedyTcERK5GoVcJuo=

    Zhou, C., Chen, Z., Li, Z., Wang, J., Jiang, K., Li, P., Yu, R., Zhang, M., Bates, S., and Jaakkola, T. (2026). Rethinking diffusion models with symmetries through canonicalization with applications to molecular graph generation.arXiv preprint arXiv:2602.15022. 13 A Full Related Works Autoencoder-free Diffusion Models.Among autoencoder-free approaches, DC...

    work page arXiv 2026