pith. sign in

arxiv: 2605.26032 · v1 · pith:XBKU3GHPnew · submitted 2026-05-25 · 💻 cs.CV · cond-mat.stat-mech· cs.AI· cs.LG

Everything at Every Scale: Scale-Invariant Diffusion with Continuous Super-Resolution

Zixin Jessie Chen , Zhuo Chen , Archer Wang , Jeff Gore , William T. Freeman , Congyue Deng , Marin Solja\v{c}i\'c This is my paper

Pith reviewed 2026-06-29 22:55 UTC · model grok-4.3

classification 💻 cs.CV cond-mat.stat-mechcs.AIcs.LG
keywords diffusion modelssuper-resolutionscale invarianceimage generationk-spaceunconditional diffusioncontinuous resolution
0
0 comments X

The pith

A single unconditional diffusion model performs both image generation and continuous super-resolution by varying only the starting timestep.

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

The paper establishes that generation and super-resolution can be treated as two instances of reversing information loss across scales within one framework. It designs a forward process that attenuates fine-scale content while adding spectrum-matched noise in k-space, turning scale into an explicit coordinate of the dynamics. The trained reverse process then handles both tasks without any architecture changes, conditioning branches, guidance, or per-scale retraining. A reader would care because this removes the need for separate models or task-specific training when moving between generation and upsampling at arbitrary factors.

Core claim

SKILD is a scale-invariant k-space image learning diffusion model whose forward process attenuates image content from fine to coarse scales while injecting spectrum-matched Gaussian noise, making scale an explicit coordinate of the diffusion dynamics. The same trained reverse process performs generation and continuous super-resolution by varying only the starting timestep, with no task-specific architecture, no conditioning branch, no classifier-free guidance, and no retraining per scale factor.

What carries the argument

The scale-invariant k-space forward process that attenuates content from fine to coarse scales while adding spectrum-matched noise, turning scale into an explicit coordinate of the diffusion dynamics.

If this is right

  • The same checkpoint reaches FID 2.65 and Inception Score 9.63 on unconditional CIFAR-10 generation.
  • The checkpoint performs 2x to 8x super-resolution on ImageNet and outperforms conditional models on perceptual metrics.
  • The checkpoint reconstructs critical Ising models with connected four-point correlations that closely track ground truth.
  • No retraining or architecture changes are needed when switching between generation and arbitrary-scale super-resolution.

Where Pith is reading between the lines

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

  • The method could extend naturally to other domains where data exhibit scale invariance, such as physical simulations at multiple resolutions.
  • Continuous variation of the starting timestep suggests the model could support arbitrary non-integer scale factors without interpolation artifacts.
  • If the k-space attenuation truly encodes scale as a coordinate, the approach might allow zero-shot transfer to related multi-scale tasks like denoising at intermediate resolutions.

Load-bearing premise

The forward process that attenuates image content from fine to coarse scales while injecting spectrum-matched Gaussian noise makes scale an explicit coordinate that the reverse process can exploit without task-specific modifications or hidden conditioning.

What would settle it

A test showing that the same checkpoint, when started at different timesteps, fails to match the perceptual quality of scale-specific conditional models on 2x through 8x super-resolution of ImageNet images while also matching unconditional generation FID.

Figures

Figures reproduced from arXiv: 2605.26032 by Archer Wang, Congyue Deng, Jeff Gore, Marin Solja\v{c}i\'c, William T. Freeman, Zhuo Chen, Zixin Jessie Chen.

Figure 1
Figure 1. Figure 1: Conceptual illustration of our SKILD on a self-similar fractal image. During the forward [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Variance power spectra of natural-image datasets. (a) Spectra of CIFAR-10, ImageNet-128, and ImageNet-256 exhibit similar power-law decay over their shared frequency range, indicating approximate scale invariance. (b) Variance of ImageNet-256 computed independently for each color channel (RGB), with a power-law fit recovering the k −2 frequency decay. efficiency, or inductive bias [29–36]. Recent works hav… view at source ↗
Figure 3
Figure 3. Figure 3: Uncurated samples of generated images on CIFAR-10; more in Appendix G (Figure G.17). [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: 4× super-resolution samples on ImageNet-256. The model is initialized from a 64 × 64 low-resolution forward state and reconstructs high-frequency details through the reverse process [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Continuous super-resolution on ImageNet-128. (a) Low-resolution inputs at effective [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: (a-b) Benchmark of four-point correlator accuracy. Our reconstruction closely tracks the [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
read the original abstract

Creating images from noise is image generation; reconstructing fine details from coarse inputs is super-resolution. Despite their practical differences, both can be understood as reversing information loss across scales. We introduce $\textbf{SKILD}$, a $\textbf{S}$cale-invariant $\textbf{K}$-Space $\textbf{I}$mage $\textbf{L}$earning $\textbf{D}$iffusion model that unifies generation and continuous super-resolution within a single unconditional framework. Both natural images and critical physical systems exhibit scale invariance, and we leverage it to design a forward process that attenuates image content from fine to coarse scales while injecting spectrum-matched Gaussian noise, making scale an explicit coordinate of the diffusion dynamics. The same trained reverse process performs generation and continuous super-resolution by varying only the starting timestep: $\textit{no task-specific architecture, no conditioning branch, no classifier-free guidance, no retraining per scale factor}$. Empirically, SKILD reaches FID $2.65$ and Inception Score $9.63$ on unconditional CIFAR-10, performs $2\times$--$8\times$ super-resolution on ImageNet from a single unconditional checkpoint while outperforming conditional models across perceptual metrics, and reconstructs critical Ising models whose connected four-point correlations closely track the ground truth.

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

3 major / 2 minor

Summary. The manuscript introduces SKILD, a scale-invariant k-space diffusion model that unifies unconditional image generation and continuous super-resolution. A forward process attenuates image content across scales while injecting spectrum-matched Gaussian noise, making scale an explicit coordinate; the identical trained reverse process then performs both tasks by varying only the starting timestep, with no task-specific architecture, conditioning, or per-scale retraining. Reported results include FID 2.65 and IS 9.63 on unconditional CIFAR-10, competitive 2×–8× super-resolution on ImageNet from a single checkpoint, and reconstruction of Ising-model four-point correlations.

Significance. If the distributional alignment between arbitrary low-resolution inputs and the forward-process states holds, the work would provide a substantive unification of two core inverse problems in diffusion models, removing the need for conditioning branches or task-specific training. The extension to critical physical systems adds cross-domain value, and the single-checkpoint SR results would be practically useful if robust. The absence of explicit verification of the alignment assumption in the abstract, however, leaves the central unification claim difficult to assess from the provided summary alone.

major comments (3)
  1. [Forward process definition (likely §3)] The unification claim rests on the premise that an arbitrary low-resolution input (e.g., bicubic-downsampled ImageNet at 4×) is statistically interchangeable with the state reached by applying the forward process (k-space attenuation + spectrum-matched noise) to its high-resolution counterpart at the matching t. No quantitative evidence—such as spectrum histograms, Wasserstein distance, or KL divergence between the two distributions—is referenced in the abstract or summary; without this, the reverse process is solving a different inverse problem than the one on which it was trained.
  2. [Experiments (§4, ImageNet SR and CIFAR-10)] Abstract reports strong SR metrics from a single unconditional checkpoint but provides no ablation on the spectrum-matching parameters, no error bars on FID/IS, and no direct comparison of starting-point distributions for real low-res inputs versus forward-process trajectories. These omissions make it impossible to isolate whether performance gains derive from the scale-invariant design or from other implementation choices.
  3. [Physical systems experiment (likely §4.3)] The Ising four-point correlation results are presented as closely tracking ground truth, yet the abstract supplies no details on how the lattice data are embedded into the image diffusion framework, what (if any) modifications to the k-space operator are required, or whether the same unconditional checkpoint is used without retraining.
minor comments (2)
  1. [Abstract] The abstract would benefit from a single-sentence statement of the forward-process equation or the precise form of spectrum matching to allow readers to assess the scale-invariance claim immediately.
  2. [Method] Notation for the timestep coordinate t and the scale parameter should be introduced consistently when the forward process is first defined.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive feedback. We address each major comment below with clarifications drawn from the full manuscript and indicate revisions that will strengthen the presentation of the unification claim and experimental details.

read point-by-point responses

  1. Referee: The unification claim rests on the premise that an arbitrary low-resolution input (e.g., bicubic-downsampled ImageNet at 4×) is statistically interchangeable with the state reached by applying the forward process (k-space attenuation + spectrum-matched noise) to its high-resolution counterpart at the matching t. No quantitative evidence—such as spectrum histograms, Wasserstein distance, or KL divergence between the two distributions—is referenced in the abstract or summary; without this, the reverse process is solving a different inverse problem than the one on which it was trained.

    Authors: Section 3 of the full manuscript presents spectrum histograms and qualitative comparisons supporting the alignment by construction via the k-space attenuation and spectrum-matched noise. We agree that explicit quantitative metrics would make the central claim easier to assess from the abstract alone, and we will add KL divergence and Wasserstein distance calculations between the low-resolution input distributions and forward-process states in the revised manuscript. revision: yes

  2. Referee: Experiments (§4, ImageNet SR and CIFAR-10) Abstract reports strong SR metrics from a single unconditional checkpoint but provides no ablation on the spectrum-matching parameters, no error bars on FID/IS, and no direct comparison of starting-point distributions for real low-res inputs versus forward-process trajectories. These omissions make it impossible to isolate whether performance gains derive from the scale-invariant design or from other implementation choices.

    Authors: Ablations on spectrum-matching parameters appear in the supplementary material, and FID/IS values are computed over multiple random seeds. We will move key ablations to the main text, explicitly include error bars on the reported metrics, and add a figure directly comparing the starting-point distributions of real low-resolution inputs versus forward-process trajectories to better isolate the contribution of the scale-invariant design. revision: yes

  3. Referee: The Ising four-point correlation results are presented as closely tracking ground truth, yet the abstract supplies no details on how the lattice data are embedded into the image diffusion framework, what (if any) modifications to the k-space operator are required, or whether the same unconditional checkpoint is used without retraining.

    Authors: Section 4.3 describes embedding the Ising lattice configurations as grayscale images, with no modifications to the k-space operator and using the identical unconditional checkpoint without retraining. We will add a concise summary of this experimental setup to the abstract and introduction in the revision. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained

full rationale

The paper introduces a new forward process (k-space attenuation plus spectrum-matched noise) that explicitly encodes scale as timestep t, trains a single unconditional reverse network on it, and reports that varying only the starting t enables both generation and continuous SR. This chain rests on the explicit construction of the forward operator and on external empirical metrics (FID, IS, perceptual scores on CIFAR-10/ImageNet/Ising), not on any fitted parameter being renamed as a prediction, self-definitional equations, or load-bearing self-citations. No quoted step reduces the central claim to its own inputs by construction; the distributional-alignment prerequisite raised by the skeptic is a correctness question, not a circularity reduction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Abstract-only review; ledger entries are inferred from stated design choices. The central claim rests on domain assumptions about scale invariance rather than new mathematical axioms or fitted constants explicitly listed.

free parameters (1)
  • spectrum-matching parameters
    Parameters controlling noise spectrum matching at each scale are required to implement the forward process but are not quantified in the abstract.
axioms (1)
  • domain assumption Natural images and critical physical systems exhibit scale invariance.
    Invoked to justify designing the forward process around scale as an explicit coordinate.

pith-pipeline@v0.9.1-grok · 5784 in / 1351 out tokens · 44317 ms · 2026-06-29T22:55:10.579116+00:00 · methodology

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

69 extracted references · 16 canonical work pages · 1 internal anchor

  1. [1]

    Field, D. J. Relations between the statistics of natural images and the response properties of cortical cells.J. Opt. Soc. Am. A4, 2379–2394 (1987). URL https://opg.optica.org/ josaa/abstract.cfm?URI=josaa-4-12-2379

    1987

  2. [2]

    Ruderman, D. L. The statistics of natural images.Network: computation in neural systems5, 517 (1994)

    1994

  3. [3]

    & van Hateren, J

    van der Schaaf, A. & van Hateren, J. Modelling the power spectra of natural images: Statistics and information.Vision Research36, 2759–2770 (1996). URL https://www. sciencedirect.com/science/article/pii/0042698996000028

    work page arXiv 1996

  4. [4]

    Simoncelli, E. P. & Olshausen, B. A. Natural image statistics and neural representation.Annual review of neuroscience24, 1193–1216 (2001)

    2001

  5. [5]

    Wilson, K. G. & Kogut, J. The renormalization group and the epsilon expansion.Physics Reports12, 75–199 (1974)

    1974

  6. [6]

    A., Polyakov, A

    Belavin, A. A., Polyakov, A. M. & Zamolodchikov, A. B. Infinite conformal symmetry in two-dimensional quantum field theory.Nuclear Physics B241, 333–380 (1984)

    1984

  7. [7]

    & Hinton, G

    Krizhevsky, A. & Hinton, G. Learning multiple layers of features from tiny images. Tech. Rep., University of Toronto (2009)

    2009

  8. [8]

    In2009 IEEE Conference on Computer Vision and Pattern Recognition, 248–255 (2009)

    Deng, J.et al.ImageNet: A large-scale hierarchical image database. In2009 IEEE Conference on Computer Vision and Pattern Recognition, 248–255 (2009)

    2009

  9. [9]

    Scale-space filtering: A new approach to multi-scale description

    Witkin, A. Scale-space filtering: A new approach to multi-scale description. InICASSP ’84. IEEE International Conference on Acoustics, Speech, and Signal Processing, vol. 9, 150–153 (1984)

    1984

  10. [10]

    Koenderink, J. J. The structure of images.Biological Cybernetics50, 363–370 (1984). URL https://doi.org/10.1007/BF00336961

    work page doi:10.1007/bf00336961 1984

  11. [11]

    Scale-space theory: A basic tool for analyzing structures at different scales

    Lindeberg, T. Scale-space theory: A basic tool for analyzing structures at different scales. Journal of applied statistics21, 225–270 (1994)

    1994

  12. [12]

    Feature detection with automatic scale selection.International journal of computer vision30, 79–116 (1998)

    Lindeberg, T. Feature detection with automatic scale selection.International journal of computer vision30, 79–116 (1998)

    1998

  13. [13]

    Lowe, D. G. Distinctive image features from scale-invariant keypoints.International Journal of Computer Vision60, 91–110 (2004). URL https://doi.org/10.1023/B: VISI.0000029664.99615.94

    work page doi:10.1023/b: 2004

  14. [14]

    Burt, P. J. & Adelson, E. H. The laplacian pyramid as a compact image code. InReadings in computer vision, 671–679 (Elsevier, 1987)

    1987

  15. [15]

    Olshausen, B. A. & Field, D. J. Natural image statistics and efficient coding.Network: Compu- tation in Neural Systems7, 333 (1996). URL https://doi.org/10.1088/0954-898X/7/ 2/014

    work page doi:10.1088/0954-898x/7/ 1996

  16. [16]

    & Ganguli, S

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

    2015

  17. [17]

    & Abbeel, P

    Ho, J., Jain, A. & Abbeel, P. Denoising diffusion probabilistic models.Advances in neural information processing systems33, 6840–6851 (2020). 10

    2020

  18. [18]

    Song, Y .et al.Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456(2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  19. [19]

    Nichol, A. Q. & Dhariwal, P. Improved denoising diffusion probabilistic models. InInterna- tional conference on machine learning, 8162–8171 (PMLR, 2021)

    2021

  20. [20]

    & Nichol, A

    Dhariwal, P. & Nichol, A. Diffusion models beat gans on image synthesis.Advances in neural information processing systems34, 8780–8794 (2021)

    2021

  21. [21]

    & Laine, S

    Karras, T., Aittala, M., Aila, T. & Laine, S. Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems35, 26565–26577 (2022)

    2022

  22. [22]

    On the importance of noise scheduling for diffusion models.arXiv preprint arXiv:2301.10972(2023)

    Chen, T. On the importance of noise scheduling for diffusion models.arXiv preprint arXiv:2301.10972(2023)

    work page arXiv 2023

  23. [23]

    Ho, J.et al.Cascaded diffusion models for high fidelity image generation.Journal of Machine Learning Research23, 1–33 (2022)

    2022

  24. [24]

    arXiv preprint arXiv:2309.03350(2023)

    Teng, J.et al.Relay diffusion: Unifying diffusion process across resolutions for image synthesis. arXiv preprint arXiv:2309.03350(2023)

    work page arXiv 2023

  25. [25]

    & Rezchikov, S

    Cotler, J. & Rezchikov, S. Renormalizing diffusion models.arXiv preprint arXiv:2308.12355 (2023)

    work page arXiv 2023

  26. [26]

    & Solin, A

    Rissanen, S., Heinonen, M. & Solin, A. Generative modelling with inverse heat dissipation. arXiv preprint arXiv:2206.13397(2022)

    work page arXiv 2022

  27. [27]

    & Ashida, Y

    Masuki, K. & Ashida, Y . Generative diffusion model with inverse renormalization group flows. arXiv preprint arXiv:2501.09064(2025)

    work page arXiv 2025

  28. [28]

    & Avestimehr, S

    Sheshmani, A., You, Y .-Z., Buyukates, B., Ziashahabi, A. & Avestimehr, S. Renormalization group flow, optimal transport, and diffusion-based generative model.Physical Review E111, 015304 (2025)

    2025

  29. [29]

    & Mallat, S

    Guth, F., Coste, S., De Bortoli, V . & Mallat, S. Wavelet score-based generative modeling. Advances in neural information processing systems35, 478–491 (2022)

    2022

  30. [30]

    & Tran, A

    Phung, H., Dao, Q. & Tran, A. Wavelet diffusion models are fast and scalable image generators. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10199–10208 (2023)

    2023

  31. [31]

    doi:10.48550/arXiv.2412.15032 , abstract =

    Ning, M.et al.Dctdiff: Intriguing properties of image generative modeling in the dct space. arXiv preprint arXiv:2412.15032(2024)

    work page arXiv 2024

  32. [32]

    Falck, F.et al.A fourier space perspective on diffusion models.arXiv preprint arXiv:2505.11278 (2025)

    work page arXiv 2025

  33. [33]

    Frequency

    Yu, H., Luo, H., Yuan, H., Rong, Y . & Zhao, F. Frequency autoregressive image generation with continuous tokens.arXiv preprint arXiv:2503.05305(2025)

    work page arXiv 2025

  34. [34]

    B., Frolov, S., Raue, F., Palacio, S

    Moser, B. B., Frolov, S., Raue, F., Palacio, S. & Dengel, A. Waving goodbye to low-res: A diffusion-wavelet approach for image super-resolution. In2024 International Joint Conference on Neural Networks (IJCNN), 1–8 (IEEE, 2024)

    2024

  35. [35]

    & Liu, J

    Gao, X., Xu, Z., Zhao, J. & Liu, J. Frequency-controlled diffusion model for versatile text-guided image-to-image translation. InProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, 1824–1832 (2024)

    2024

  36. [36]

    & Cattin, P

    Friedrich, P., Wolleb, J., Bieder, F., Durrer, A. & Cattin, P. C. Wdm: 3d wavelet diffusion models for high-resolution medical image synthesis. InMICCAI workshop on deep generative models, 11–21 (Springer, 2024)

    2024

  37. [37]

    Bansal, A.et al.Cold diffusion: Inverting arbitrary image transforms without noise.Advances in Neural Information Processing Systems36, 41259–41282 (2023). 11

    2023

  38. [38]

    & Wang, L

    Tian, K., Jiang, Y ., Yuan, Z., Peng, B. & Wang, L. Visual autoregressive modeling: Scalable image generation via next-scale prediction.Advances in neural information processing systems 37, 84839–84865 (2024)

    2024

  39. [39]

    & Pasztor, E

    Freeman, W. & Pasztor, E. Learning low-level vision. InProceedings of the Seventh IEEE International Conference on Computer Vision, vol. 2, 1182–1189 vol.2 (1999)

    1999

  40. [40]

    C., He, K

    Dong, C., Loy, C. C., He, K. & Tang, X. Image super-resolution using deep convolutional networks.IEEE transactions on pattern analysis and machine intelligence38, 295–307 (2015)

    2015

  41. [41]

    & Mu Lee, K

    Lim, B., Son, S., Kim, H., Nah, S. & Mu Lee, K. Enhanced deep residual networks for single image super-resolution. InProceedings of the IEEE conference on computer vision and pattern recognition workshops, 136–144 (2017)

    2017

  42. [42]

    In Proceedings of the European conference on computer vision (ECCV) workshops, 0–0 (2018)

    Wang, X.et al.Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, 0–0 (2018)

    2018

  43. [43]

    & Timofte, R

    Zhang, K., Liang, J., Van Gool, L. & Timofte, R. Designing a practical degradation model for deep blind image super-resolution. InProceedings of the IEEE/CVF international conference on computer vision, 4791–4800 (2021)

    2021

  44. [44]

    InProceedings of the IEEE/CVF international conference on computer vision, 1833–1844 (2021)

    Liang, J.et al.Swinir: Image restoration using swin transformer. InProceedings of the IEEE/CVF international conference on computer vision, 1833–1844 (2021)

    2021

  45. [45]

    & Shan, Y

    Wang, X., Xie, L., Dong, C. & Shan, Y . Real-esrgan: Training real-world blind super-resolution with pure synthetic data. InProceedings of the IEEE/CVF international conference on computer vision, 1905–1914 (2021)

    1905

  46. [46]

    Saharia, C.et al.Image super-resolution via iterative refinement.IEEE transactions on pattern analysis and machine intelligence45, 4713–4726 (2022)

    2022

  47. [47]

    Li, H.et al.Srdiff: Single image super-resolution with diffusion probabilistic models.Neuro- computing479, 47–59 (2022)

    2022

  48. [48]

    & Loy, C

    Yue, Z., Wang, J. & Loy, C. C. Resshift: Efficient diffusion model for image super-resolution by residual shifting.Advances in neural information processing systems36, 13294–13307 (2023)

    2023

  49. [49]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 25796–25805 (2024)

    Wang, Y .et al.Sinsr: diffusion-based image super-resolution in a single step. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 25796–25805 (2024)

    2024

  50. [50]

    & Song, J

    Kawar, B., Elad, M., Ermon, S. & Song, J. Denoising diffusion restoration models.Advances in neural information processing systems35, 23593–23606 (2022)

    2022

  51. [51]

    K., Zhao, Z., Sjölund, J

    Luo, Z., Gustafsson, F. K., Zhao, Z., Sjölund, J. & Schön, T. B. Image restoration with mean-reverting stochastic differential equations.arXiv preprint arXiv:2301.11699(2023)

    work page arXiv 2023

  52. [52]

    Liu, G.-H.et al.I 2SB: Image-to-image schrödinger bridge.arXiv preprint arXiv:2302.05872 (2023)

    work page arXiv 2023

  53. [53]

    Crystal statistics

    Onsager, L. Crystal statistics. i. a two-dimensional model with an order-disorder transition. Physical Review65, 117–149 (1944)

    1944

  54. [54]

    & Hoyer, P

    Hyvrinen, A., Hurri, J. & Hoyer, P. O.Natural Image Statistics: A Probabilistic Approach to Early Computational Vision.(Springer Publishing Company, Incorporated, 2009), 1st edn

    2009

  55. [55]

    & Rao, K

    Ahmed, N., Natarajan, T. & Rao, K. R. Discrete cosine transform.IEEE Transactions on ComputersC-23, 90–93 (1974)

    1974

  56. [56]

    Collective monte carlo updating for spin systems.Physical Review Letters62, 361–364 (1989)

    Wolff, U. Collective monte carlo updating for spin systems.Physical Review Letters62, 361–364 (1989)

    1989

  57. [57]

    & Hochreiter, S

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B. & Hochreiter, S. GANs trained by a two time-scale update rule converge to a local nash equilibrium. InAdvances in Neural Information Processing Systems, vol. 30 (2017). 12

    2017

  58. [58]

    InAdvances in Neural Information Processing Systems, vol

    Salimans, T.et al.Improved techniques for training GANs. InAdvances in Neural Information Processing Systems, vol. 29 (2016)

    2016

  59. [59]

    C., Sheikh, H

    Wang, Z., Bovik, A. C., Sheikh, H. R. & Simoncelli, E. P. Image quality assessment: From error visibility to structural similarity.IEEE Transactions on Image Processing13, 600–612 (2004)

    2004

  60. [60]

    A., Shechtman, E

    Zhang, R., Isola, P., Efros, A. A., Shechtman, E. & Wang, O. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 586–595 (2018)

    2018

  61. [61]

    & Yang, F

    Ke, J., Wang, Q., Wang, Y ., Milanfar, P. & Yang, F. MUSIQ: Multi-scale image quality transformer. InProceedings of the IEEE/CVF International Conference on Computer Vision, 5148–5157 (2021)

    2021

  62. [62]

    Wang, J., Chan, K. C. K. & Loy, C. C. Exploring CLIP for assessing the look and feel of images. InProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, 2555–2563 (2023)

    2023

  63. [63]

    & Ommer, B

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

    2022

  64. [64]

    Efthymiou, S., Beach, M. J. S. & Melko, R. G. Super-resolving the ising model with convolu- tional neural networks.Physical Review B99, 075113 (2019)

    2019

  65. [65]

    Shiina, K., Mori, H., Tomita, Y ., Lee, H. K. & Okabe, Y . Inverse renormalization group based on image super-resolution using deep convolutional networks.Scientific Reports11, 9617 (2021). 13 A DDPM formulation This appendix derives the DDPM posterior used in the main text and lists the training targets supported by our implementation. Figure A.8 shows a...

    work page arXiv 2021

  66. [66]

    Initialize the cluster and the active frontier to contain only this seed site

    Choose a seed site uniformly at random and let s⋆ be its spin. Initialize the cluster and the active frontier to contain only this seed site

  67. [67]

    If a candidate site touches k∈ {1,2,3,4} frontier sites, add it to the cluster with probability 1−(1−p) k

    Given the current frontier, inspect nearest-neighbor sites with spin s⋆ that are not already in the cluster. If a candidate site touches k∈ {1,2,3,4} frontier sites, add it to the cluster with probability 1−(1−p) k. This is equivalent to independently activating each bond from a frontier site with probabilityp

  68. [68]

    Set the newly added sites as the next frontier and repeat the previous step until the frontier is empty

  69. [69]

    We run several independent chains in parallel, each initialized with independent random spins in {−1,+1} 128×128

    Flip every spin in the completed cluster:s i ← −s i. We run several independent chains in parallel, each initialized with independent random spins in {−1,+1} 128×128. Every chain is thermalized for 2,000 Wolff transitions. After thermalization, we save the current configuration from each chain whenever all of them have accumulated at least2L2 flipped spin...