NoiseCutMix: A Novel Data Augmentation Approach by Mixing Estimated Noise in Diffusion Models

Ryoma Bise; Shinnosuke Matsuo; Shumpei Takezaki

REVIEW 3 major objections 5 minor 24 references

NoiseCutMix: A Novel Data Augmentation Approach by Mixing Estimated Noise in Diffusion Models

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read NoiseCutMix generates natural fused-class images by mixing the noise estimates of two class prompts inside Stable Diffusion's denoising loop, and improves fine-grained classification accuracy over CutMix and MixUp on CUB and Flower.

desk verdict NoiseCutMix is a clean, honestly written diffusion-based mixing augmentation with plausible gains on two of three datasets, held back by a size-confounded headline comparison and a missing Diff-Mix baseline. read the letter →

arxiv 2509.00378 v1 pith:P2OMSSWF submitted 2025-08-30 cs.CV

Shumpei Takezaki , Ryoma Bise , Shinnosuke Matsuo This is my paper

classification cs.CV

keywords dataaugmentationdiffusionmodelsCutMixnoisemixingfine-grainedclassificationStablesoftlabels

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

Data augmentation that mixes two classes usually pastes pixels together, leaving unnatural seams. NoiseCutMix instead moves CutMix into the diffusion process: at each denoising step it runs the UNet with each class prompt, combines the two predicted noises with a CutMix-style rectangular mask, and denoises with the mixture. Because every step re-synthesizes the whole image, the resulting images fuse the visual traits of both classes without hard boundaries, and the mask area still controls the soft label. On fine-grained CUB birds and Oxford flowers, training a ResNet-50 with these images improves accuracy over CutMix, MixUp, and naive Stable-Diffusion generation. The method inherits a key dependence: when Stable Diffusion cannot generate the target domain well, as on FGVC-Aircraft, the augmented images can hurt accuracy.

What carries the argument

The central object is the mask-weighted noise mixture epsilon = M * epsilon_A + (1-M) * epsilon_B, evaluated at every reverse step of a class-conditioned diffusion model. The binary mask M is sampled exactly as in CutMix, so the mixing ratio is precisely controllable and the soft training label stays CutMix's rule. This transfers CutMix's data-diversity mechanism from pixel space to the noise/latent space that the diffusion UNet uses, which is what produces smooth, semantically coherent fusions.

What would settle it

On a dataset where fine-tuned Stable Diffusion is known to generate high-fidelity images (e.g., CUB), generate two training sets with identical masks and lambda values, one using NoiseCutMix and one using SD-random images with CutMix applied in pixel space. If ResNet-50 accuracies are statistically indistinguishable across five runs, the paper's claim that mixing noise is better than mixing generated pixels is falsified.

Watch

Extended reading notes

Core claim

The paper's claim is that the estimated-noise field, not the pixel field, is the right place to apply CutMix. For two class prompts, the UNet predicts epsilon_A and epsilon_B from the same current latent, and the method replaces the noise in a rectangular region of one estimate with the corresponding region of the other, with the mask area ratio set to 1-lambda, then uses the mixed noise for the denoising step. The final image naturally blends features from both classes, and the soft label lambda*y_A + (1-lambda)*y_B matches the mixture proportion. Trained with these images, a ResNet-50 classifier gains 2.27% over CutMix on CUB and 2.98% on Flower, and outperforms both SD-random and SD-rando

Load-bearing premise

NoiseCutMix only helps if Stable Diffusion, after fine-tuning, can generate convincing images of the target classes; when it cannot, the mixed-noise images hurt classifier accuracy instead of helping.

Editorial extensions

If this is right

On datasets where a fine-tuned Stable Diffusion generates convincing class images, NoiseCutMix should deliver consistent gains over pixel-space mixing (CutMix/MixUp) and over naive 'generate with SD then CutMix' pipelines.
Because the label rule and mask sampling are inherited from CutMix, NoiseCutMix is a drop-in augmentation for any classifier that already uses CutMix; no new labeling or loss terms are required.
The same noise-mixing mechanism can be layered onto other Stable-Diffusion-based augmentation methods, since it only modifies the denoising trajectory.
On domains where the diffusion model misaligns with the target distribution, the method can backfire; accuracy on Aircraft drops below the no-augmentation baseline, so generator quality is a precondition.

Reading between the lines

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

A testable extension: replace the hard rectangular mask with a soft or irregular mask in noise space; if the benefit comes from smooth transitions rather than label mixing, soft masks should improve accuracy further.
The method effectively manufactures 'between-class' exemplars, which may help disentangle shared and discriminative features; this suggests a use for low-data fine-grained domains where only a few reference images per class exist.
The Aircraft failure points to a practical screening rule researchers could adopt: measure generation quality (e.g., FID or per-class classifier confidence) before deciding whether NoiseCutMix will help.
One could also test whether the gain is from the natural boundaries or from the soft-label regularization by feeding NoiseCutMix images to a classifier trained with hard labels; the paper does not isolate this factor.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Alice,

Here's my read of NoiseCutMix. The concrete new thing is the algorithm: at each denoising step you run the UNet with two class prompts and combine the two estimated noises with a CutMix-style binary mask. That is genuinely different from Diff-Mix, which mixes at the image level after generation, and the paper describes it clearly. It also ships code, and the qualitative CUB examples show the intended effect: natural-looking blends that actually carry traits from both parent classes.

Credit where it's due: the authors do not hide the Aircraft failure, they explicitly flag the SD-distribution dependence as a limitation, and they include SD-random+CutMix and SD-random+MixUp conditions. Those last two are important because they partially control for the “extra data” confound—if NoiseCutMix used more data than plain CutMix and that were the whole story, it should not beat SD-random+CutMix. So the mechanism does appear to add something.

The soft spots are real, though none is fatal. First, the headline claim—outperforms CutMix and MixUp—is only true on CUB and Flower, not Aircraft; the conclusion overstates consistency. Second, the comparison is still not size-matched: NoiseCutMix adds 100% generated images to the training set, while CutMix/MixUp are applied on-the-fly to the original data only. Part of the 2–3 point gain could come from seeing twice as many unique samples. The SD-random+CutMix baseline is the right control and it points in the right direction, but the paper never reports a size-matched CutMix. Third, there are no significance tests. Five trials with overlapping standard deviations—Ours on CUB is 68.78±0.33, MixUp 67.57±0.88—makes the gain over MixUp look fragile. Fourth, the Beta hyperparameter alpha for NoiseCutMix is never stated, and Diff-Mix [20] is not compared against even though the authors use its fine-tuned SD weights. That is the most direct prior method and it needs to be in the table.

Who is this for? Anyone working on diffusion-based augmentation for fine-grained classification. It's a simple, testable idea that does not require retraining the diffusion model and can probably be plugged into existing pipelines. I would send it to a referee, not desk-reject it, with the expectation that the revision adds a Diff-Mix comparison, a matched-size control, and significance testing. My own verdict is conditional, not rejection.

Best, [Your name]

Referee Report

3 major / 5 minor

Summary. The paper proposes NoiseCutMix, a data augmentation method that applies CutMix-style mixing inside the reverse denoising process of Stable Diffusion. For two class prompts, the UNet estimates two noises at each denoising step; these are combined with a binary rectangular mask whose area ratio determines both the image blend and the soft label. The method is evaluated with a ResNet-50 classifier on CUB-200, Oxford Flowers, and FGVC-Aircraft, comparing against CutMix, MixUp, SD-random, SD-random+CutMix, and SD-random+MixUp. The authors report accuracy gains on CUB and Flower, show qualitative examples of fused bird images, and acknowledge that the method fails on Aircraft because Stable Diffusion does not align with that data distribution.

Significance. If the reported gains are robust, NoiseCutMix is a simple and interesting contribution to diffusion-based data augmentation: it performs class mixing in the noise domain, producing natural boundaries while retaining explicit control over the mixing ratio. The public code and the inclusion of SD-random and SD-random+CutMix/MixUp controls are strengths, as is the transparent discussion of the Aircraft failure. However, the headline comparison with CutMix/MixUp is not size-matched, and the statistical evidence is weak, so the significance of the empirical claim depends on additional confirmatory experiments.

major comments (3)

[§4.3, Table 1] The headline comparison with CutMix/MixUp is confounded by training-set size. The paper states that for the diffusion-based conditions (including NoiseCutMix) it generated 100% additional images and added them to the training set, whereas CutMix and MixUp are applied as on-the-fly augmentations to the original images only. Thus NoiseCutMix trains on roughly 2N unique images per epoch, while CutMix/MixUp see N images. The reported gains of 2.27% on CUB and 2.98% on Flower over CutMix could therefore be caused by the extra data rather than by mixing in the noise domain. The SD-random+CutMix/MixUp controls partially address this, but they do not exactly isolate the proposed mechanism under matched data size. Please add a size-matched control, e.g., apply CutMix/MixUp to the same original+generated training set, or train all methods for the same total number of samples, and report those numb
[§5.1, Table 1] No statistical significance tests are reported. With only five trials, several differences are small relative to the reported standard deviations. For example, on CUB, Ours (68.78 ± 0.33) vs. Original (67.78 ± 1.39) and vs. MixUp (67.57 ± 0.88); on Flower, Ours (92.91 ± 0.32) vs. SD-random (92.52 ± 1.26). Please provide paired significance tests (e.g., paired t-test or Wilcoxon over the same seeds/folds) or confidence intervals so the reader can judge whether the improvements are reliable rather than seed noise.
[§6, Table 1] The conclusion states that the experiments show 'consistent gains over standard augmentation', but the Aircraft results directly contradict this: Ours (79.69) is below CutMix (83.28), MixUp (82.71), and the Original baseline (81.69). While the limitation paragraph acknowledges the dependence on Stable Diffusion, the abstract's unqualified claim and the conclusion's 'consistent gains' are too broad. Please qualify the central claim to CUB and Flower, or reframe the Aircraft result as an explicit negative case with analysis of why the method underperforms.

minor comments (5)

[§4.3] The value of the Beta distribution parameter α for NoiseCutMix is not reported, although α is specified for CutMix (α=1.0) and MixUp (α=0.2). Please state the α value(s) used and whether it was tuned per dataset.
[Footnote 2] The parenthetical use of Diff-Mix's publicly available fine-tuned weights is vague. Please specify which exact weights were used, which dataset they were fine-tuned on, and how they relate to the LoRA/Textual Inversion description in the main text.
[References [8] and [9]] References [8] and [9] are duplicates of the same Diffusemix paper; one should be removed and the citation details merged correctly.
[§3.3, Eqs. (3)–(4)] The mask construction should be clarified: when W and H differ, the formulas r_w = W√(1−λ) and r_h = H√(1−λ) give the desired area ratio only if the rectangle remains inside the image. Please state explicitly how boundary cases are handled and define the direction of the mask (which region is set to 0).
[§4.3] It is unclear whether NoiseCutMix uses the generated images in every epoch alongside the original images, and whether the 'augmentation probability 0.5' mentioned for CutMix/MixUp also applies to NoiseCutMix. Please describe the per-epoch data composition for each condition.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: the contribution is an empirical augmentation method with no self-citation or construction-reduced prediction.

full rationale

The paper proposes NoiseCutMix, which mixes the estimated noise of two class-conditioned diffusion models during denoising. The central claim—that this improves classification accuracy over CutMix/MixUp on CUB and Flower—is an empirical result, not a derivation that reduces to its inputs. The method's equations (Eq. 1 for noise mixing, Eq. 2 for the soft label, Eqs. 3–4 for mask sampling) define the augmentation procedure; they do not fit any parameter to the evaluation data. The label rule is borrowed from CutMix, but that is standard practice and not circular. There are no self-citations: all cited prior works (e.g., Diff-Mix, DiffuseMix, GenMix) are by other author groups. No uniqueness theorem is invoked, and no ansatz is smuggled in via a self-citation. The internal comparisons against SD-random+CutMix and SD-random+MixUp use the same number of generated images as NoiseCutMix, providing a controlled test of the mixing mechanism. The headline comparison to standard CutMix/MixUp is not training-set-size matched (NoiseCutMix adds 100% generated images while CutMix/MixUp do not), which is a potential experimental confound but not a form of circular reasoning. The duplicate reference [8]/[9] is a minor editorial error. Therefore, the derivation chain is self-contained and no circularity is present.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The ledger shows that the method rests on three main assumptions: the linear mixing of noise estimates is meaningful, the CutMix label rule transfers to noise-mixed samples, and the external fine-tuned diffusion model is well aligned with the target domains. The only free parameter is the Beta hyperparameter α, which is not reported for the proposed method.

free parameters (1)

Beta distribution parameter α for NoiseCutMix
Controls the distribution of the mask area ratio λ in Eq. (2). The paper specifies α for CutMix (1.0) and MixUp (0.2) but never states the value used for NoiseCutMix, leaving a free hyperparameter unreported.

assumptions (3)

domain assumption The UNet's noise estimate ε conditioned on a class prompt encodes class-specific features that can be linearly combined via a mask to produce a valid mixed image.
The core mechanism of Eq. (1) assumes that mixing noise estimates yields a sample containing features of both classes. This is the central design choice and is only supported by qualitative examples, not by a theoretical justification.
domain assumption The soft label rule Ῡ = λ yA + (1-λ) yB, borrowed from CutMix, is valid for images generated by noise mixing.
The paper assumes that the mask area ratio directly corresponds to the feature contribution in the final generated image (Section 3.2). No analysis is given to support that the noise-mixing process preserves this ratio in the final image.
domain assumption The fine-tuned Stable Diffusion model, obtained from Diff-Mix's publicly available weights, produces images whose distribution is sufficiently aligned with each target dataset for augmentation to help.
The method's success depends on the quality of the generator. The paper states this as a limitation in Section 6 and shows that when alignment fails (Aircraft), the method does not improve accuracy.

how reviews work

0 comments

Cite this review

Pith. "Pith review of NoiseCutMix: A Novel Data Augmentation Approach by Mixing Estimated Noise in Diffusion Models." pith.science (2026). https://pith.science/paper/P2OMSSWF

@misc{pith2026250900378,
  author       = {Pith},
  title        = {Pith review of: NoiseCutMix: A Novel Data Augmentation Approach by Mixing Estimated Noise in Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P2OMSSWF}},
  note         = {Machine review of arXiv:2509.00378}
}

read the original abstract

In this study, we propose a novel data augmentation method that introduces the concept of CutMix into the generation process of diffusion models, thereby exploiting both the ability of diffusion models to generate natural and high-resolution images and the characteristic of CutMix, which combines features from two classes to create diverse augmented data. Representative data augmentation methods for combining images from multiple classes include CutMix and MixUp. However, techniques like CutMix often result in unnatural boundaries between the two images due to contextual differences. Therefore, in this study, we propose a method, called NoiseCutMix, to achieve natural, high-resolution image generation featuring the fused characteristics of two classes by partially combining the estimated noise corresponding to two different classes in a diffusion model. In the classification experiments, we verified the effectiveness of the proposed method by comparing it with conventional data augmentation techniques that combine multiple classes, random image generation using Stable Diffusion, and combinations of these methods. Our codes are available at: https://github.com/shumpei-takezaki/NoiseCutMix

Figures

Figures reproduced from arXiv: 2509.00378 by the authors.

**Figure 2.** The proposed NoiseCutmix mixes estimated noise of diffusion models in the denoising process. 3. NoiseCutMix: Data Augmentation by Mixing Estimated Noise in Diffusion Models We propose NoiseCutMix, a novel data augmentation method that leverages a pre-trained diffusion model, such as Stable Diffusion, to generate highly natural and diverse samples of a desired class, conditioned on text prompts. As shown in [PITH_F… view at source ↗

**Figure 3.** Examples from the CUB dataset: real images, images generated by our data augmentation method, and the masks used. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 20 canonical work pages

[20]

Enhance im- age classification via inter-class image mixup with diffusion model

Zhicai Wang, Longhui Wei, Tan Wang, Heyu Chen, Yanbin Hao, Xiang Wang, Xiangnan He, and Qi Tian. Enhance im- age classification via inter-class image mixup with diffusion model. In Computer Vision and Pattern Recognition, pages 17223–17233, 2024. 2, 3

work page 2024
[1]

Advances in diffusion models for image data augmentation: A review of methods, models, evaluation metrics and future research directions

Panagiotis Alimisis, Ioannis Mademlis, Panagiotis Radoglou-Grammatikis, Panagiotis Sarigiannidis, and Georgios Th Papadopoulos. Advances in diffusion models for image data augmentation: A review of methods, models, evaluation metrics and future research directions. Artificial Intelligence Review, pages 1–55, 2025. 2

work page 2025
[2]

Shekoofeh Azizi, Simon Kornblith, Chitwan Saharia, Mo- hammad Norouzi, and David J. Fleet. Synthetic data from diffusion models improves imagenet classification. Transac- tions on Machine Learning Research, 2023. 2

work page 2023
[3]

An image is worth one word: Personalizing text-to-image gen- eration using textual inversion

Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image gen- eration using textual inversion. In International Conference on Learning Representations, 2023. 3

work page 2023
[4]

Deep residual learning for image recognition

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition, pages 770–778, 2016. 3

work page 2016
[5]

Classifier-free diffusion guidance

Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In Advances in Neural Information Processing Systems Workshop on Deep Generative Models and Down- stream Applications, 2021. 3

work page 2021
[6]

Denoising dif- fusion probabilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. In Advances in Neural Informa- tion Processing Systems, pages 6840–6851, 2020. 1

work page 2020
[7]

Lora: Low-rank adaptation of large language models

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3

arXiv 2021

Show all 24 references

[9]

Diffusemix: Label- preserving data augmentation with diffusion models

Khawar Islam, Muhammad Zaigham Zaheer, Arif Mah- mood, and Karthik Nandakumar. Diffusemix: Label- preserving data augmentation with diffusion models. In Computer Vision and Pattern Recognition , pages 27621– 27630, 2024. 2

2024
[10]

Genmix: effective data augmentation with generative diffusion model image editing

Khawar Islam, Muhammad Zaigham Zaheer, Arif Mah- mood, Karthik Nandakumar, and Naveed Akhtar. Genmix: effective data augmentation with generative diffusion model image editing. arXiv preprint arXiv:2412.02366, 2024. 2

2024
[11]

A comprehensive survey of recent trends in deep learn- ing for digital images augmentation

Nour Eldeen Khalifa, Mohamed Loey, and Seyedali Mir- jalili. A comprehensive survey of recent trends in deep learn- ing for digital images augmentation. Artificial Intelligence Review, pages 2351–2377, 2022. 1, 2

2022
[12]

Adam: A method for stochastic optimization

Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. International Conference on Learn- ing Representations, 2015. 3

2015
[13]

Dpm-solver: A fast ode solver for dif- fusion probabilistic model sampling in around 10 steps

Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan LI, and Jun Zhu. Dpm-solver: A fast ode solver for dif- fusion probabilistic model sampling in around 10 steps. In Advances in Neural Information Processing Systems , pages 5775–5787, 2022. 3

2022
[14]

Fine-grained visual classi- fication of aircraft

Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classi- fication of aircraft. arXiv preprint arXiv:1306.5151 , 2013. 3, 4

2013 arXiv
[15]

Automated flower classification over a large number of classes

Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In In- dian Conference on Computer Vision, Graphics and Image Processing, 2008. 3, 4

2008
[16]

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. In Computer Vision and Pattern Recognition, pages 10684–10695, 2022. 1, 3, 4

2022
[17]

A survey on image data augmentation for deep learning

Connor Shorten and Taghi M Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of big data, 6(1):1–48, 2019. 1, 2

2019
[18]

Effective data augmentation with diffu- sion models

Brandon Trabucco, Kyle Doherty, Max A Gurinas, and Rus- lan Salakhutdinov. Effective data augmentation with diffu- sion models. In International Conference on Learning Rep- resentations, 2024. 2

2024
[19]

The caltech-ucsd birds-200 dataset

Catherine Wah, Steve Branson, Peter Welinder, Pietro Per- ona, and Serge Belongie. The caltech-ucsd birds-200 dataset. Technical report, California Institute of Technology, 2011. 3, 4

2011
[21]

Image data aug- mentation for deep learning: A survey

Suorong Yang, Weikang Xiao, Mengchen Zhang, Suhan Guo, Jian Zhao, and Furao Shen. Image data aug- mentation for deep learning: A survey. arXiv preprint arXiv:2204.08610, 2022. 1, 2

2022 arXiv
[22]

Real-fake: Effective training data synthesis through distribution matching

Jianhao Yuan, Jie Zhang, Shuyang Sun, Philip Torr, and Bo Zhao. Real-fake: Effective training data synthesis through distribution matching. In International Conference on Learning Representations, 2024. 2

2024
[23]

Cutmix: Regu- larization strategy to train strong classifiers with localizable features

Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regu- larization strategy to train strong classifiers with localizable features. In International Conference on Computer Vision , pages 6023–6032, 2019. 1, 2, 3, 4

2019
[24]

mixup: Beyond empirical risk minimiza- tion

Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimiza- tion. In International Conference on Learning Representa- tions, 2018. 1, 2, 3, 4

2018
[25]

A survey on data augmentation in large model era

Yue Zhou, Chenlu Guo, Xu Wang, Yi Chang, and Yuan Wu. A survey on data augmentation in large model era. arXiv preprint arXiv:2401.15422, 2024. 2 5

2024 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [20]

Enhance im- age classification via inter-class image mixup with diffusion model

Zhicai Wang, Longhui Wei, Tan Wang, Heyu Chen, Yanbin Hao, Xiang Wang, Xiangnan He, and Qi Tian. Enhance im- age classification via inter-class image mixup with diffusion model. In Computer Vision and Pattern Recognition, pages 17223–17233, 2024. 2, 3

work page 2024

[2] [1]

Advances in diffusion models for image data augmentation: A review of methods, models, evaluation metrics and future research directions

Panagiotis Alimisis, Ioannis Mademlis, Panagiotis Radoglou-Grammatikis, Panagiotis Sarigiannidis, and Georgios Th Papadopoulos. Advances in diffusion models for image data augmentation: A review of methods, models, evaluation metrics and future research directions. Artificial Intelligence Review, pages 1–55, 2025. 2

work page 2025

[3] [2]

Shekoofeh Azizi, Simon Kornblith, Chitwan Saharia, Mo- hammad Norouzi, and David J. Fleet. Synthetic data from diffusion models improves imagenet classification. Transac- tions on Machine Learning Research, 2023. 2

work page 2023

[4] [3]

An image is worth one word: Personalizing text-to-image gen- eration using textual inversion

Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image gen- eration using textual inversion. In International Conference on Learning Representations, 2023. 3

work page 2023

[5] [4]

Deep residual learning for image recognition

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition, pages 770–778, 2016. 3

work page 2016

[6] [5]

Classifier-free diffusion guidance

Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In Advances in Neural Information Processing Systems Workshop on Deep Generative Models and Down- stream Applications, 2021. 3

work page 2021

[7] [6]

Denoising dif- fusion probabilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. In Advances in Neural Informa- tion Processing Systems, pages 6840–6851, 2020. 1

work page 2020

[8] [7]

Lora: Low-rank adaptation of large language models

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3

arXiv 2021