Pith. sign in

REVIEW 2 major objections 2 minor 79 references

Efficient and Training-Free Single-Image Diffusion Models

T0 review · 2 major / 2 minor · reviewed 2026-06-28 · grok-4.3

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

Pith's one-line read A finite collection of multi-scale patches from one image supplies a closed-form optimal denoiser that replaces neural training inside diffusion sampling.

desk verdict The paper replaces trained score networks with a closed-form denoiser computed from an image's own finite multi-scale patches, avoiding per-image training while claiming SOTA quality. read the letter →

arxiv 2606.04299 v1 pith:WM2KU2XF submitted 2026-06-03 cs.CV cs.LG

Haojun Qiu , Kiriakos N. Kutulakos , David B. Lindell This is my paper · ORCID
classification cs.CVcs.LG
keywords single-imagediffusionpatch-baseddenoisingtraining-freegenerationmulti-scalepatchesmodelsimageclosed-formdenoiserunconditional
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper establishes that a single reference image can be represented by the empirical distribution of its patches extracted at several fixed scales. Because this collection is finite and low-dimensional, an optimal denoiser for any noisy patch can be written in closed form and used directly as the score function during diffusion. The resulting sampler generates new images whose patch statistics match the original across scales, without any per-image neural-network optimization. A reader would care because the method converts an hours-long training step into a fast computation while reporting higher quality and diversity than earlier trained single-image diffusion models. The same machinery supports text-guided stylization, symmetrization, retargeting, and latent-space acceleration to megapixel outputs in one second.

What carries the argument

The multi-scale patch dataset together with its closed-form optimal denoiser, which is inserted as the score function inside the diffusion sampling loop.

What would settle it

Generated samples whose multi-scale patch histograms differ measurably from the reference image, or whose perceptual quality falls below that of trained single-image baselines on standard metrics.

Watch

Extended reading notes

Core claim

By treating a single image as a finite dataset of its patches at multiple scales, the score function for diffusion can be replaced by a tractable closed-form optimal denoiser. This substitution removes the requirement for neural-network training while still allowing the diffusion process to produce new images whose internal patch distributions match those of the reference image.

Load-bearing premise

The internal structure of an image is fully captured by the empirical distribution of its patches across a small number of discrete scales.

Editorial extensions

If this is right

  • Unconditional generation of new images from one reference example without any training.
  • Text-guided stylization achieved by combining the patch denoiser with external conditioning signals.
  • Image symmetrization and retargeting performed by the same patch-based diffusion process.
  • Megapixel outputs reachable in one second and gigapixel outputs in minutes through latent-space and acceleration techniques.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Classical patch-based restoration algorithms can be lifted directly into modern generative diffusion frameworks.
  • The method may extend to video or 3D data by replacing 2-D patches with corresponding spatio-temporal volumes.
  • Because the denoiser is analytic, further speed-ups could come from analytic approximations rather than learned networks.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

2 major / 2 minor

Summary. The paper proposes a training-free single-image diffusion model that extracts patches at multiple discrete scales from a reference image to form a finite dataset, then substitutes the closed-form optimal denoiser (posterior mean under the empirical patch measure) for the learned score network inside the diffusion ODE/SDE. It claims this yields SOTA generation quality and diversity versus trained single-image models, with applications in unconditional generation, text-guided stylization, symmetrization, retargeting, latent-space compatibility, and accelerations enabling megapixel output in one second.

Significance. If the patch-based closed-form score approximation holds, the result would be significant for enabling rapid, optimization-free single-image synthesis and for explicitly linking diffusion models to classical patch-based restoration. The training-free construction and reported acceleration techniques are concrete strengths that could reduce barriers to high-resolution single-image tasks.

major comments (2)
  1. [§3] §3 (method description): the substitution of the closed-form patch denoiser into the diffusion SDE/ODE is asserted without a derivation or error bound showing that the patch-wise posterior mean approximates the true score of the full-image distribution at each noise level.
  2. [Experiments] Experimental results section: the SOTA claim on quality and diversity is unsupported by any quantitative tables, FID/LPIPS scores, or ablation studies comparing against trained single-image baselines, leaving the performance assertion unverified.
minor comments (2)
  1. The description of how patch extraction at discrete scales connects to classical non-local means or patch-based priors could include explicit equation references to prior work for clarity.
  2. Notation for the number of scales and patch sizes (listed as free parameters) should be introduced with a clear table or diagram early in the method section.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed review and recommendation. We address each major comment below with clarifications based on the manuscript's modeling assumptions and commit to revisions that strengthen the presentation without altering the core contributions.

read point-by-point responses

  1. Referee: [§3] §3 (method description): the substitution of the closed-form patch denoiser into the diffusion SDE/ODE is asserted without a derivation or error bound showing that the patch-wise posterior mean approximates the true score of the full-image distribution at each noise level.

    Authors: The manuscript explicitly defines the generative model via the empirical distribution of multi-scale patches extracted from the reference image, so the closed-form denoiser is the exact posterior mean (and thus the exact score) under this finite patch measure rather than an approximation to an external full-image distribution. The diffusion process is then applied consistently to overlapping patches to synthesize the full image, with the connection to classical patch-based restoration noted in the text. We agree that a more explicit derivation would improve clarity and will revise §3 to include a short derivation of the score substitution under the patch-stationarity assumption, along with a brief discussion of the approximation error when patches are treated as locally independent. revision: yes

  2. Referee: [Experiments] Experimental results section: the SOTA claim on quality and diversity is unsupported by any quantitative tables, FID/LPIPS scores, or ablation studies comparing against trained single-image baselines, leaving the performance assertion unverified.

    Authors: The current experiments emphasize visual results to illustrate diversity, quality, and applications such as stylization and retargeting. We acknowledge that quantitative metrics would provide stronger verification of the SOTA claims relative to trained single-image baselines. We will add a table of FID and LPIPS scores, plus ablations on the number of scales and patch sizes, in the revised experimental section. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; derivation grounded in external classical result

full rationale

The paper derives its training-free diffusion model by extracting patches at discrete scales from a single image to form a finite dataset, then substituting the closed-form optimal denoiser (posterior mean under the empirical measure) for the score function. This step invokes a standard result from finite-data denoising rather than any fitted parameter, self-citation chain, or redefinition of the target output. No equation or claim reduces the generation process to the inputs by construction, and the method explicitly connects to classical patch-based restoration without load-bearing self-references. The central claims therefore remain independent of the paper's own outputs.

Assumptions & free parameters 1 free parameters · 2 assumptions · 0 invented entities

The central claim rests on the premise that patch statistics at a few discrete scales suffice to define image structure and that the finite dataset permits an exact closed-form score; no new entities are postulated.

free parameters (1)
  • number of scales and patch sizes
    Choice of discrete scales and patch dimensions is required to build the dataset; these are not derived from first principles.
assumptions (2)
  • domain assumption The internal structure of an image is defined by the distribution of its patches across multiple scales.
    Explicitly stated as the problem definition in the first sentence of the abstract.
  • domain assumption A finite collection of patches admits an optimal closed-form denoiser whose score can be used directly inside a diffusion sampler.
    Invoked to eliminate neural-network training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient and Training-Free Single-Image Diffusion Models." pith.science (2026). https://pith.science/paper/WM2KU2XF

@misc{pith2026260604299,
  author       = {Pith},
  title        = {Pith review of: Efficient and Training-Free Single-Image Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WM2KU2XF}},
  note         = {Machine review of arXiv:2606.04299}
}
read the original abstract

We consider the problem of generating images whose internal structure -- defined by the distribution of patches across multiple scales -- matches that of a single reference image. Recent approaches address this problem by training a diffusion model on a single image. But even in this setting, training is computationally expensive and requires hours of optimization. Instead, we model the image using a dataset of its patches at different scales. As this dataset is finite and the dimensionality of its patches is small, the score function for a noisy patch can be computed tractably using an optimal, closed-form denoiser, eliminating the need for neural network training. We integrate this patch-based denoiser into an efficient, training-free image diffusion model, and we describe how our method connects to classical patch-based image restoration techniques. Our approach achieves state-of-the-art generation quality and diversity compared to trained single-image diffusion models, and we demonstrate applications, including unconditional image generation, text-guided stylization, image symmetrization, and retargeting. Further, we show that our approach is compatible with latent space diffusion, and we show multiple additional acceleration techniques to achieve megapixel single-image generation in one second, and gigapixel generation in minutes.

Figures

Figures reproduced from arXiv: 2606.04299 by the authors.

Figure 1
Figure 1. We introduce an efficient, training-free diffusion model that generates images based on the internal structure of a single input [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Method overview. Our approach takes a single image as input, extracts patches, and uses the patches to generate new images using [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of single-scale sampling. This procedure [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Unconditional single-image generation results. Our training-free, coarse-to-fine image sampling procedure based on closed-form [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: High-resolution generation. The input image is 308 MP, and we generate an image of size [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: Structural analogies [4]. Our approach combines the style of one image and the structure of another image to generate a new image that combines both properties. Nano Banana Pro [25] preserves neither the structure nor the input patch distribution. denoising (η > 0) and…
Figure 8
Figure 8. Figure 8: Plot of SIFID vs diffusion timesteps T for coarse-to-fine image sampling across different η values. The SIFID converges in roughly 10 timesteps. patch size ρ ∞ 3.0 1.0 0.2 0.1 5 2.2 ± 0.9 2.2 ± 0.9 2.4 ± 1.2 2.7 ± 1.9 2.8 ± 2.0 7 2.7 ± 1.4 2.6 ± 1.3 2.1 ± 1.0 1.9 ± 1.0…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

79 extracted references · 4 canonical work pages

  1. [1]

    Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein

    Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie S. Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Cold diffusion: Inverting ar- bitrary image transforms without noise. InProc. NeurIPS,

  2. [2]

    Text2LIVE: Text-driven layered image and video editing

    Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kas- ten, and Tali Dekel. Text2LIVE: Text-driven layered image and video editing. InProc. ECCV, 2022. 12

    2022

  3. [3]

    Patchmatch: A randomized correspon- dence algorithm for structural image editing.ACM Trans

    Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspon- dence algorithm for structural image editing.ACM Trans. Graph., 28(3):24, 2009. 1, 2

    2009

  4. [4]

    Structural analogy from a single image pair.Computer Graphics Forum, 40(1):249–265, 2021

    Saguy Benaim, Ron Mokady, Amit Bermano, and Lior Wolf. Structural analogy from a single image pair.Computer Graphics Forum, 40(1):249–265, 2021. 6, 7

    2021

  5. [5]

    On the closed-form of flow matching: Gen- eralization does not arise from target stochasticity

    Quentin Bertrand, Anne Gagneux, Mathurin Massias, and R´emi Emonet. On the closed-form of flow matching: Gen- eralization does not arise from target stochasticity. InProc. NeurIPS, 2025. 3, 5

    2025

  6. [6]

    Large scale GAN training for high fidelity natural image synthesis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. InProc. ICLR, 2019. 1

    2019

  7. [7]

    A non-local algorithm for image denoising

    Antoni Buades, Bartomeu Coll, and Jean-Michel Morel. A non-local algorithm for image denoising. InProc. CVPR,

  8. [8]

    On the edge of memorization in diffusion models

    Sam Buchanan, Druv Pai, Yi Ma, and Valentin De Bortoli. On the edge of memorization in diffusion models. InProc. NeurIPS, 2025. 2, 3, 5

    2025

Show all 79 references
  1. [9]

    A multiresolution spline with application to image mosaics.ACM Trans

    Peter Burt and Edward Adelson. A multiresolution spline with application to image mosaics.ACM Trans. Graph., 2 (4):217–236, 1983. 5

    1983

  2. [10]

    NIFTY: a non-local image flow matching for texture syn- thesis.arXiv preprint arXiv:2509.22318, 2025

    Pierrick Chatillon, Julien Rabin, and David Tschumperl ´e. NIFTY: a non-local image flow matching for texture syn- thesis.arXiv preprint arXiv:2509.22318, 2025. 2

    2025

  3. [11]

    Diffusion pos- terior sampling for general noisy inverse problems

    Hyungjin Chung, Jeongsol Kim, Michael Thompson Mc- cann, Marc Louis Klasky, and Jong Chul Ye. Diffusion pos- terior sampling for general noisy inverse problems. InProc. ICLR, 2023. 8, 27

    2023

  4. [12]

    Image denoising by sparse 3-D transform- domain collaborative filtering.IEEE Trans

    Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-D transform- domain collaborative filtering.IEEE Trans. Image Process., 16(8):2080–2095, 2007. 1, 2

  5. [13]

    FlashAttention-2: Faster attention with better par- allelism and work partitioning

    Tri Dao. FlashAttention-2: Faster attention with better par- allelism and work partitioning. InProc. ICLR, 2024. 5, 6, 8

    2024

  6. [14]

    Fu, Stefano Ermon, Atri Rudra, and Christopher R´e

    Tri Dao, Daniel Y . Fu, Stefano Ermon, Atri Rudra, and Christopher R´e. FlashAttention: Fast and memory-efficient exact attention with IO-awareness. InProc. NeurIPS, 2022. 2, 5, 6, 8

    2022

  7. [15]

    Diffusion models beat GANs on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat GANs on image synthesis. InProc. NeurIPS, 2021. 1, 2, 8

    2021

  8. [16]

    Diffusion is spectral autoregression

    Sander Dieleman. Diffusion is spectral autoregression. https://sander.ai/2024/09/02/spectral- autoregression.html, 2024. Blog post. 27

    2024

  9. [17]

    The Faiss library

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazar´e, Maria Lomeli, Lucas Hosseini, and Herv´e J´egou. The Faiss library. IEEE Trans. Big Data, 12(2):346–361, 2025. 9

    2025

  10. [18]

    Image quilting for texture synthesis and transfer

    Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. InProc. SIGGRAPH, 2001. 1, 2

    2001

  11. [19]

    Texture synthesis by non-parametric sampling

    Alexei A Efros and Thomas K Leung. Texture synthesis by non-parametric sampling. InProc. ICCV, 1999. 1, 2

    1999

  12. [20]

    Image denoising via sparse and redundant representations over learned dictionar- ies.IEEE Trans

    Michael Elad and Michal Aharon. Image denoising via sparse and redundant representations over learned dictionar- ies.IEEE Trans. Image Process., 15(12):3736–3745, 2006. 2

    2006

  13. [21]

    Generating natural images with direct patch distributions matching

    Ariel Elnekave and Yair Weiss. Generating natural images with direct patch distributions matching. InProc. ECCV,

  14. [22]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. InProc. ICML, 2024. 2

    2024

  15. [23]

    Super- resolution from a single image

    Daniel Glasner, Shai Bagon, and Michal Irani. Super- resolution from a single image. InProc. ICCV, 2009. 1

    2009

  16. [24]

    Generative adversarial networks.Commun

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Commun. ACM, 63(11):139–144, 2020. 1

    2020

  17. [25]

    Introducing Nano Banana Pro (Gemini 3 Pro Image).https://blog.google/innovation- and-ai/products/nano-banana-pro/, 2025

    Google DeepMind. Introducing Nano Banana Pro (Gemini 3 Pro Image).https://blog.google/innovation- and-ai/products/nano-banana-pro/, 2025. 7, 8, 16

    2025

  18. [26]

    Drop the GAN: In defense of patches near- est neighbors as single image generative models

    Niv Granot, Ben Feinstein, Assaf Shocher, Shai Bagon, and Michal Irani. Drop the GAN: In defense of patches near- est neighbors as single image generative models. InProc. CVPR, 2022. 1, 6, 19

    2022

  19. [27]

    Image analogies

    Aaron Hertzmann, Charles Jacobs, Nuria Oliver, Brian Cur- less, and David Salesin. Image analogies. InProc. SIG- GRAPH, 2001. 2

    2001

  20. [28]

    Improved techniques for training single-image GANs

    Tobias Hinz, Matthew Fisher, Oliver Wang, and Stefan Wermter. Improved techniques for training single-image GANs. InProc. WACV, 2021. 1, 2

    2021

  21. [29]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. InProc. NeurIPS, 2020. 1, 2, 3

    2020

  22. [30]

    Blurring diffusion models

    Emiel Hoogeboom and Tim Salimans. Blurring diffusion models. InProc. ICLR, 2023. 27

    2023

  23. [31]

    Estimation of non- normalized statistical models by score matching.J

    Aapo Hyv ¨arinen and Peter Dayan. Estimation of non- normalized statistical models by score matching.J. Mach. Learn. Res., 6(4):695–709, 2005. 2

    2005

  24. [32]

    Approximate nearest neighbors: Towards removing the curse of dimensionality

    Piotr Indyk and Rajeev Motwani. Approximate nearest neighbors: Towards removing the curse of dimensionality. InProc. STOC, 1998. 2, 5

    1998

  25. [33]

    Billion- scale similarity search with GPUs.IEEE Trans

    Jeff Johnson, Matthijs Douze, and Herv ´e J ´egou. Billion- scale similarity search with GPUs.IEEE Trans. Big Data, 7(3):535–547, 2019. 9 9

    2019

  26. [34]

    Frey, and Anitha Kannan

    Nebojsa Jojic, Brendan J. Frey, and Anitha Kannan. Epito- mic analysis of appearance and shape. InProc. ICCV, 2003. 1

    2003

  27. [35]

    An analytic theory of cre- ativity in convolutional diffusion models

    Mason Kamb and Surya Ganguli. An analytic theory of cre- ativity in convolutional diffusion models. InProc. ICML,

  28. [36]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. InProc. NeurIPS, 2022. 2, 3, 5

    2022

  29. [37]

    MUSIQ: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-scale image quality transformer. InProc. ICCV, 2021. 6, 7

    2021

  30. [38]

    SinDDM: A single image denoising diffusion model

    Vladimir Kulikov, Shahar Yadin, Matan Kleiner, and Tomer Michaeli. SinDDM: A single image denoising diffusion model. InProc. ICML, 2023. 1, 2, 6, 7, 8, 12, 13, 16, 17

    2023

  31. [39]

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dock- horn, Jack English, Zion English, Patrick Esser, et al. FLUX. 1 Kontext: Flow matching for in-context image generation and editing in latent space.arXiv preprint arX...

    arXiv

  32. [40]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matthew Le. Flow matching for generative modeling. InProc. ICLR, 2023. 6

    2023

  33. [41]

    Locality in image diffusion models emerges from data statistics

    Artem Lukoianov, Chenyang Yuan, Justin Solomon, and Vincent Sitzmann. Locality in image diffusion models emerges from data statistics. InProc. NeurIPS, 2025. 2, 3, 5

    2025

  34. [42]

    Tiled diffusion

    Or Madar and Ohad Fried. Tiled diffusion. InProc. CVPR,

  35. [43]

    Non-local sparse models for image restoration

    Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro, and Andrew Zisserman. Non-local sparse models for image restoration. InProc. ICCV, 2009. 1

    2009

  36. [44]

    Sym- metrization.ACM Trans

    Niloy J Mitra, Leonidas J Guibas, and Mark Pauly. Sym- metrization.ACM Trans. Graph., 26(3):63–es, 2007. 2

    2007

  37. [45]

    completely blind

    Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Mak- ing a “completely blind” image quality analyzer.IEEE Sig- nal Process. Lett., 20(3):209–212, 2012. 6, 7

    2012

  38. [46]

    Spectral normalization for generative ad- versarial networks

    Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative ad- versarial networks. InProc. ICLR, 2018. 1, 2

    2018

  39. [47]

    Null-text inversion for editing real images using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. InProc. CVPR, 2023. 8

    2023

  40. [48]

    Scale space diffusion.arXiv preprint arXiv:2603.08709, 2026

    Soumik Mukhopadhyay, Prateksha Udhayanan, and Abhi- nav Shrivastava. Scale space diffusion.arXiv preprint arXiv:2603.08709, 2026. 27

    2026

  41. [49]

    Towards a mechanistic explana- tion of diffusion model generalization

    Matthew Niedoba, Berend Zwartsenberg, Kevin Patrick Murphy, and Frank Wood. Towards a mechanistic explana- tion of diffusion model generalization. InProc. ICML, 2025. 2, 3, 5, 27

    2025

  42. [50]

    SinFusion: Training diffusion models on a single image or video

    Yaniv Nikankin, Niv Haim, and Michal Irani. SinFusion: Training diffusion models on a single image or video. In Proc. ICML, 2023. 1, 2, 6

    2023

  43. [51]

    Multi-scale patch-based image restoration.IEEE Trans

    Vardan Papyan and Michael Elad. Multi-scale patch-based image restoration.IEEE Trans. Image Process., 25(1):249– 261, 2015. 2

    2015

  44. [52]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. InProc. ICML, 2021. 1, 2, 8

    2021

  45. [53]

    Gener- ating diverse high-fidelity images with VQ-V AE-2

    Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Gener- ating diverse high-fidelity images with VQ-V AE-2. InProc. NeurIPS, 2019. 1, 2

    2019

  46. [54]

    Gener- ative modelling with inverse heat dissipation

    Severi Rissanen, Markus Heinonen, and Arno Solin. Gener- ative modelling with inverse heat dissipation. InProc. ICLR,

  47. [55]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InProc. CVPR, 2022. 2, 5

    2022

  48. [56]

    Closed-form diffusion models.Trans

    Christopher Scarvelis, Haitz S ´aez de Oc ´ariz Borde, and Justin Solomon. Closed-form diffusion models.Trans. Mach. Learn. Res., 2025. 2, 3, 5

    2025

  49. [57]

    Flashattention-3: Fast and ac- curate attention with asynchrony and low-precision

    Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, and Tri Dao. Flashattention-3: Fast and ac- curate attention with asynchrony and low-precision. InProc. NeurIPS, 2024. 6, 8

    2024

  50. [58]

    Sin- GAN: Learning a generative model from a single natural im- age

    Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli. Sin- GAN: Learning a generative model from a single natural im- age. InProc. ICCV, 2019. 1, 2, 6, 7

    2019

  51. [59]

    Zero- shot super-resolution using deep internal learning

    Assaf Shocher, Nadav Cohen, and Michal Irani. Zero- shot super-resolution using deep internal learning. InProc. CVPR, 2018. 1

    2018

  52. [60]

    InGAN: Capturing and remapping the “DNA” of a natural image

    Assaf Shocher, Shai Bagon, Phillip Isola, and Michal Irani. InGAN: Capturing and remapping the “DNA” of a natural image. InProc. ICCV, 2019. 2

    2019

  53. [61]

    Video Google: A text retrieval ap- proach to object matching in videos

    Sivic and Zisserman. Video Google: A text retrieval ap- proach to object matching in videos. InProc. ICCV, 2003. 5, 9

    2003

  54. [62]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InProc. ICML, 2015. 3

    2015

  55. [63]

    Denois- ing diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. InProc. ICLR, 2021. 1, 2, 3

    2021

  56. [64]

    Selective underfitting in dif- fusion models.arXiv preprint arXiv:2510.01378, 2025

    Kiwhan Song, Jaeyeon Kim, Sitan Chen, Yilun Du, Sham Kakade, and Vincent Sitzmann. Selective underfitting in dif- fusion models.arXiv preprint arXiv:2510.01378, 2025. 2, 3, 5

    2025

  57. [65]

    Generative modeling by es- timating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by es- timating gradients of the data distribution. InProc. NeurIPS,

  58. [66]

    Score-based generative modeling through stochastic differential equa- tions

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Ab- hishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equa- tions. InProc. ICLR, 2021. 3

    2021

  59. [67]

    NIMA: Neural image assessment.IEEE Trans

    Hossein Talebi and Peyman Milanfar. NIMA: Neural image assessment.IEEE Trans. Image Process., 27(8):3998–4011,

  60. [68]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. InProc. NeurIPS,

  61. [69]

    A connection between score matching and denoising autoencoders.Neural Comput., 23(7):1661–1674,

    Pascal Vincent. A connection between score matching and denoising autoencoders.Neural Comput., 23(7):1661–1674,

  62. [70]

    SinDiffu- sion: Learning a diffusion model from a single natural im- age.IEEE Trans

    Weilun Wang, Jianmin Bao, Wengang Zhou, Dongdong Chen, Dong Chen, Lu Yuan, and Houqiang Li. SinDiffu- sion: Learning a diffusion model from a single natural im- age.IEEE Trans. Pattern Anal. Mach. Intell., 47(5):3412– 3423, 2025. 2, 6

    2025

  63. [71]

    Texture synthesis over arbitrary manifold surfaces

    Li-Yi Wei and Marc Levoy. Texture synthesis over arbitrary manifold surfaces. InProc. SIGGRAPH, 2001. 1

    2001

  64. [72]

    Geometry distribu- tions

    Biao Zhang, Jing Ren, and Peter Wonka. Geometry distribu- tions. InProc. ICCV, 2025. 8

    2025

  65. [73]

    Internal statistics of a single natural image

    Maria Zontak and Michal Irani. Internal statistics of a single natural image. InProc. CVPR, 2011. 1, 2

    2011

  66. [74]

    From learning models of natu- ral image patches to whole image restoration

    Daniel Zoran and Yair Weiss. From learning models of natu- ral image patches to whole image restoration. InProc. ICCV,

  67. [75]

    Connections to Prior Single-Image Generative Methods

    1, 2, 3 11 Efficient and Training-Free Single-Image Diffusion Models Supplementary Material Contents S1. Connections to Prior Single-Image Generative Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 S1.1. Comparison with GPNN . . . . . . . . . . . . . ...

  68. [76]

    x(i) t 1 # ,k (j) =

    (S12) To find the denoiser that produces the minimum mean squared error estimate, we set the gradient of the loss with respect to 3 the denoiser to be0: 0=∇ DL(D;x t, t)(S13) =∇ D wt Y YX j=1 N(x t;α ty(j), σ2 t I)∥D(xt, t)−y (j)∥2 2 (S14) = wt Y YX j=1 N(x t;α ty(j), σ2 t I)∇...

  69. [77]

    finding the closestn probe centroids and their associated clusters

  70. [78]

    circshift

    performing an exact search for theknearest neighbours within the union of thosen probe clusters. The approximation comes from step (1): if we setnprobe =n list, the method reduces to exact nearest neighbour search over all patches. In our experiments, we follow a standard choi...

  71. [79]

    Van Gogh

    The momentum term helps prevent CLIP updates from being overridden by the denoiser [38]; we setλ= 0.1for all experiments. Also, following [38], we do not use the raw CLIP gradient but instead compute gt ← ∇ λˆxt+(1−λ)ˆxt+1,CLIP LCLIP,(S38) mt ←1 pixel-norm(gt)>(1−f)-quantile ,...

Pith tools

Reviewed June 28, 2026 · model on record in the stance chip above.