REVIEW 3 major objections 5 minor 1 cited by
Ambient Diffusion Omni: Training Good Models with Bad Data
T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read The paper claims that a diffusion model can be trained on a mixture of clean, corrupted, and out-of-distribution images and still generate clean samples, as long as each degraded image is used only at diffusion times where noise has made…
desk verdict A practical, honest paper with strong empirical results and a real gap between its theory and the trained objective; the method stands, the theory needs a rewrite. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the distance-contraction inequality of Theorem 4.2: for distributions $P, Q$ supported on a set of diameter $D$, $d_{TV}(P \circledast \mathcal{N}(0, \sigma^2 I), Q \circledast \mathcal{N}(0, \sigma^2 I)) \le d_{TV}(P, Q) \cdot D/(2\sigma)$. This shows additive Gaussian noise shrinks distribution mismatch linearly in $1/\sigma$, converting the bias introduced by corrupted samples into a controlled, decaying term while their contribution increases sample size and cuts variance. The second mechanism is the receptive-field-to-noise mapping $\mathrm{crop}(t)$: empirical loss curves show denoising at noise level $t$ needs only a local context, so matching patch marginals between target and out-of-distribution data justifies using those images at low $t$. Time-conditional and crop-conditional classifiers turn these two conditions into per-sample annotations $t_{\min}$ and $t_{\max}$.
What would settle it
Measure $d_{TV}(\text{noised clean}, \text{noised corrupted})$ directly on a synthetic blur-corruption dataset, for example by estimating densities or training a calibrated classifier, and compare the time at which it drops below a fixed $\epsilon$ with the $t_{\min}$ produced by Eq. (3.2); if the classifier crosses its threshold systematically before the true total-variation distance is below $\epsilon$, samples are being used while still biased and the annotation step fails its stated goal.
Extended reading notes
Core claim
The paper's central claim is that diffusion training does not need to choose between 'good' and 'bad' data; it can use everything, as long as each sample is annotated with the diffusion times at which it is safe. Theorem 4.2 formalizes the high-noise half: convolving two distributions with Gaussian noise of standard deviation $\sigma$ reduces their total-variation distance by a factor $D/(2\sigma)$, where $D$ is the diameter of the support, so at sufficiently large $t$ the noised corrupted distribution is close to the noised clean distribution and the corrupted samples contribute signal with bounded bias. For low noise, the paper exploits locality: the optimal denoiser at noise level $t$ only needs a crop of size $\mathrm{crop}(t)$, so any out-of-distribution or synthetic image whose crop marginals match the target can be used for $t$ below $t_{\max}$. A trained time-conditional classifier selects $t_{\min}$ by finding when the noised distributions merge, and a crops classifier selects $t_{\max}$ by finding the largest crop for which clean and out-of-distribution patches are indistinguishable. In between lies the 'donut paradox': a middle range with neither global merging nor local crop equivalence, where the method falls back to the clean subset. Experiments with Gaussian blur, JPEG, and motion blur on CIFAR-10 and FFHQ, plus ImageNet and a text-to-image model, support the claim that this annotation strategy improves both fidelity and diversity.
Load-bearing premise
The method's safety annotations rest on the assumption that a time-conditional classifier trained to distinguish clean from corrupted noised images has its confidence threshold crossing exactly when the two distributions merge, so that Eq. (3.2) picks the true $t_{\min}$; a miscalibrated classifier or a misset threshold would let biased samples in or discard useful signal.
Editorial extensions
If this is right
- Filtering approaches discard usable signal: the bias-variance trade-off shows that, at high diffusion times, training on a large corrupted set can beat training only on a small clean set.
- The requirement of knowing the exact degradation model can be relaxed; for arbitrary corruptions, a classifier that finds the merging time is enough to make the data usable.
- Synthetic data from older models can help a text-to-image model when treated as corrupted data rather than clean data, improving COCO FID, GenEval scores, and output diversity.
- On ImageNet-512, the method reaches the best reported FID among the compared baselines and reduces memorization, as shown by larger gains on test FID than train FID.
- A middle range of diffusion times, the 'donut paradox,' has less usable data than either the very high or very low noise regimes, and the framework explicitly falls back to clean samples there.
Reading between the lines
- Beyond the paper: because the contraction factor in Theorem 4.2 is $D/(2\sigma)$, one can predict a data-scaling law: the number $n_2$ of corrupted samples needed to beat a small clean set should grow roughly quadratically with $1/\sigma$ at the crossover, a prediction a controlled experiment could test.
- Beyond the paper: the same crop-matching idea could be made patch-wise, annotating each image region with its own $t_{\max}$; this would recover signal from images that are only partly degraded, such as faces blurred for privacy, which the paper lists as future work.
- Beyond the paper: for corruptions that hit low frequencies, the method degrades to filtering; replacing isotropic Gaussian noise with colored or structured noise might extend the safe window to low-frequency corruption, a direction not explored here.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces Ambient Diffusion Omni (Ambient-o), a framework for training diffusion models on datasets that mix clean, corrupted, and out-of-distribution images. The method has two components: in the high-noise regime, a time-conditional classifier annotates each low-quality sample with a minimum diffusion time t_min above which the noised corrupted distribution is treated as approximately equal to the noised clean distribution; in the low-noise regime, a crop-based classifier identifies patch sizes for which out-of-distribution images can supply local high-frequency detail. The authors provide a theoretical analysis based on Gaussian kernel density estimation and a total-variation contraction bound under additive noise, together with controlled CIFAR-10 and FFHQ experiments, ImageNet-512 results, and text-to-image experiments on a Micro-Diffusion model.
Significance. If the theoretical justification holds, the paper makes a strong practical contribution: it shows that data normally discarded by filtering can be repurposed, and it reports state-of-the-art ImageNet FID and improved text-to-image generation without sacrificing diversity. The controlled experiments are convincing and internally consistent, especially the CIFAR-10 blur results (e.g., FID 5.34 for Ambient-o vs. 11.42 for all-data training at sigma_B=0.6). The theoretical results in Section 4 are self-contained and clearly proved in the appendix, and the paper promises code and model release, which will aid reproducibility. However, the gap between the total-variation bounds and the denoiser objective used in training is load-bearing, and the classifier-based annotation is asserted rather than rigorously connected to the theory.
major comments (3)
- [Section 4, Theorem 4.2 and Eqs. (2.1)-(2.2)] The theoretical comparison controls the total variation distance between noised marginals, but the diffusion training objective optimizes a conditional expectation (the denoiser), and small dTV(p_t, q_t) does not imply small denoiser bias. For example, take p_0 and q_0 to be narrow smooth densities supported in [0,1] with means 0 and 1. By Theorem 4.2, dTV(p_t,q_t) <= dTV(p_0,q_0)/(2 sigma_t) tends to 0 as sigma_t grows, while E_p[X_0 | X_t=x] and E_q[Y_0 | Y_t=x] remain approximately 0 and 1 for every fixed x. The ambient objective trained mostly on q can therefore have an O(1) bias even when the theorem's criterion is met. The limitation paragraph in Section 3.1 acknowledges that low-frequency corruptions are harder, but Section 4 does not restrict the theory to high-frequency corruptions. The authors should either prove a bound on the difference of the conditional expectations (or scores) under explicit corruption assumptions, or explicitly restrict the theoretical claim and state that the low-frequency cases are supported only empirically.
- [Section 3.1, Eq. (3.2), and Section 4, final paragraph] The assertion that an optimal classifier 'exactly tracks' dTV(~p_t, p_t) is not proven and is not generally true. For a Bayes classifier, the average predicted probability over the corrupted distribution is \int q_t(x) p_t(x)/(p_t(x)+q_t(x)) dx, which is related to Hellinger affinity rather than total variation. The first time this average crosses a fixed threshold tau = 0.5 - epsilon is therefore not equal to the first time dTV(p_t,q_t) <= epsilon. Because t_min controls which corrupted samples enter the ambient loss, this connection is load-bearing. Please provide a calibration argument, use a direct estimator of dTV, or validate the threshold choice on held-out data and report the sensitivity of the final FID to tau.
- [Section 4 vs. Section 3] The theory analyzes Algorithms 1 and 2 with a single fixed annotation time t_n and no per-sample annotation, whereas the actual method uses sample-dependent t_min^i and crop-based t_max^i, producing the 'donut' data-availability structure shown in Figure 14. The low-noise part in Section 3.2 relies on Eq. (3.3) for crop marginals, but no theorem quantifies the effect of approximate equality of crop marginals on the denoiser error, nor the error of the crops classifier in Eq. (3.4). Please either extend the theory to the actual algorithm or explicitly state that the theoretical results cover only the fixed-annotation high-noise variant and treat the crop-based extension as an empirical contribution.
minor comments (5)
- [Appendix A.1, Eq. (A.18)] In Eq. (A.18), the bound is written as |\hat p_sigma(l/L) - p_sigma(x)|, but the preceding union-bound argument requires comparing \hat p_sigma(l/L) with p_sigma(l/L). Please correct the variable to avoid confusion.
- [Table 2a] The 'All data' row is hard to read because four FID values and their corresponding sigma_B values are packed into single cells. Please format these as separate rows or separate columns.
- [Section 3.1, Eq. (3.2) and Figure 2] The threshold is defined as tau = 0.5 - epsilon, but the value epsilon = 0.05 used in Figure 2 is not given in the main text; please state it explicitly when discussing the figure.
- [Appendix E.3 and Appendix E.1] There are typos: 'paramemeters' should be 'parameters' and 'wethers' should be 'whether'.
- [Section 5, ImageNet results] The reported ImageNet FID improvements over EDM-2 are small (e.g., Train FID 1.91 vs. 1.93 for Ambient-o-XXL+crops with CFG), and no statistical significance or multiple-seed variability is reported. Please add a discussion of run-to-run variation or confidence intervals for at least the main ImageNet comparison.
Circularity Check
No significant circularity: the theoretical bounds are proven in-paper from independent coupling and KDE arguments, and the classifier-based annotation is a heuristic whose dTV-equivalence is asserted rather than used to force a result.
full rationale
I walked the derivation chain and found no step in which a claimed prediction reduces by construction to its inputs. The load-bearing theory consists of Theorem 4.1 and Theorem 4.2. Theorem 4.1 is a Gaussian kernel density estimation bound proven in Appendix A.1 via Hoeffding's inequality and a Lipschitzness argument; it does not assume the target result. Theorem 4.2 is proven in Appendix A.2 via the optimal coupling theorem, Fact 2, Fact 3, and a product coupling; its conclusion dTV(P*N(0,sigma^2 I), Q*N(0,sigma^2 I)) <= dTV(P,Q)*D/(2 sigma) follows from the coupling construction, not from the training objective or from the paper's own claims. The reuse of the ambient objective (2.2) from the authors' prior work [16] is a real self-citation, but it is not load-bearing for the new claims: the objective is stated explicitly and the new analysis compares Algorithms 1 and 2 using the in-paper theorems. The classifier annotation in Eq. (3.2) defines t_min as a threshold crossing of a trained classifier, and the paper asserts that an optimal classifier 'exactly tracks' dTV(pt, qt). That assertion is an unproved heuristic connecting classifier confidence to total variation, and it is a correctness or overreach risk rather than a circularity: the annotation is a learned selection rule, not a fitted constant of the final model, and the theory does not define t_min in terms of the experiments' success. The same holds for the crops classifier in Eq. (3.4), where the definition of t_max as the point of classifier confusion is exactly a heuristic annotation, and the experimental FID improvements are external evidence. The skeptic's concern that small dTV(p_t, q_t) does not imply small denoiser bias is a legitimate gap between the theorem's quantity and the L2 conditional-expectation objective, and the paper's own Section 3.1 limitation paragraph concedes that low-frequency corruptions are 'more challenging'; but this is an incompleteness in the theoretical justification, not a circular derivation. Against external benchmarks (FID, CLIP-FD, GenEval), the results are not forced by construction. Accordingly, the paper receives a low score reflecting only a minor, non-load-bearing self-citation lineage.
Assumptions & free parameters
free parameters (4)
- Classifier threshold tau =
0.45 (epsilon = 0.05)
- DiffusionDB annotation noise sigma_min =
2
- CLIP-IQA quality split =
top 10% vs bottom 90% of ImageNet
- Ambient loss buffer/clip =
clip at 2.0 or buffer at 4.0
assumptions (5)
- domain assumption Natural images have power-law spectral decay
- domain assumption At low diffusion times, optimal denoising needs only local context
- domain assumption The mixture model ~p0 = (n1/(n1+n2)) p0 + (n2/(n1+n2)) q0 with known clean/corrupted labels
- standard math p0 supported on [0,1] and lambda-Lipschitz
- ad hoc to paper An optimal time-conditional classifier's confidence tracks dTV(pt,qt)
Cite this review
Pith. "Pith review of Ambient Diffusion Omni: Training Good Models with Bad Data." pith.science (2026). https://pith.science/paper/VPVLN6NW
@misc{pith2026250610038,
author = {Pith},
title = {Pith review of: Ambient Diffusion Omni: Training Good Models with Bad Data},
year = {2026},
howpublished = {\url{https://pith.science/paper/VPVLN6NW}},
note = {Machine review of arXiv:2506.10038}
}
read the original abstract
We show how to use low-quality, synthetic, and out-of-distribution images to improve the quality of a diffusion model. Typically, diffusion models are trained on curated datasets that emerge from highly filtered data pools from the Web and other sources. We show that there is immense value in the lower-quality images that are often discarded. We present Ambient Diffusion Omni, a simple, principled framework to train diffusion models that can extract signal from all available images during training. Our framework exploits two properties of natural images -- spectral power law decay and locality. We first validate our framework by successfully training diffusion models with images synthetically corrupted by Gaussian blur, JPEG compression, and motion blur. We then use our framework to achieve state-of-the-art ImageNet FID, and we show significant improvements in both image quality and diversity for text-to-image generative modeling. The core insight is that noise dampens the initial skew between the desired high-quality distribution and the mixed distribution we actually observe. We provide rigorous theoretical justification for our approach by analyzing the trade-off between learning from biased data versus limited unbiased data across diffusion times.
Figures
Figures from the paper (25 more)
Forward citations
Cited by 1 Pith paper
-
Riemannian AmbientFlow: Towards Simultaneous Manifold Learning and Generative Modeling from Corrupted Data
Riemannian AmbientFlow jointly learns a deformed-Gaussian generative model and a pullback-Riemannian autoencoder from corrupted measurements, with Wasserstein recovery and linear-convergence inverse-problem guarantees.
Reference graph
Works this paper leans on
-
[1]
Solving inverse problems with score-based generative priors learned from noisy data
Asad Aali, Marius Arvinte, Sidharth Kumar, and Jonathan I Tamir. Solving inverse problems with score-based generative priors learned from noisy data. arXiv preprint arXiv:2305.01166, 2023
arXiv 2023
-
[2]
Asad Aali, Giannis Daras, Brett Levac, Sidharth Kumar, Alex Dimakis, and Jon Tamir. Ambient diffusion posterior sampling: Solving inverse problems with diffusion models trained on corrupted data. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=qeXcMutEZY
work page 2025
-
[3]
Self-consuming generative models go mad
Sina Alemohammad, Josue Casco-Rodriguez, Lorenzo Luzi, Ahmed Imtiaz Humayun, Hossein Babaei, Daniel LeJeune, Ali Siahkoohi, and Richard G Baraniuk. Self-consuming generative models go mad. arXiv preprint arXiv:2307.01850, 4:14, 2023
arXiv 2023
-
[4]
Self-improving diffusion models with synthetic data
Sina Alemohammad, Ahmed Imtiaz Humayun, Shruti Agarwal, John Collomosse, and Richard Baraniuk. Self-improving diffusion models with synthetic data. arXiv preprint arXiv:2408.16333, 2024
arXiv 2024
-
[5]
Weimin Bai, Yifei Wang, Wenzheng Chen, and He Sun. An expectation-maximization al- gorithm for training clean diffusion models from corrupted observations. arXiv preprint arXiv:2407.01014, 2024
arXiv 2024
-
[6]
Procedural image programs for representation learning
Manel Baradad, Chun-Fu Chen, Jonas Wulff, Tongzhou Wang, Rogerio Feris, Antonio Torralba, and Phillip Isola. Procedural image programs for representation learning. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022. URL https://openreview.net/forum?id=wJwHTgIoE0P
work page 2022
-
[7]
Ambientgan: Generative models from lossy measurements
Ashish Bora, Eric Price, and Alexandros G Dimakis. Ambientgan: Generative models from lossy measurements. In International conference on learning representations, 2018
work page 2018
-
[8]
Kernel density estimation via diffusion
Zdravko I Botev, Joseph F Grotowski, and Dirk P Kroese. Kernel density estimation via diffusion. 2010
work page 2010
Show all 71 references
-
[9]
Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions
Sitan Chen, Sinho Chewi, Jerry Li, Yuanzhi Li, Adil Salim, and Anru R Zhang. Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions. arXiv preprint arXiv:2209.11215, 2022
2022 arXiv
-
[10]
Restoration-degradation beyond linear diffusions: A non-asymptotic analysis for DDIM-type samplers
Sitan Chen, Giannis Daras, and Alex Dimakis. Restoration-degradation beyond linear diffusions: A non-asymptotic analysis for DDIM-type samplers. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, Proceedings of t...
2023
-
[11]
Stargan v2: Diverse image synthesis for multiple domains
Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8188–8197, 2020
2020
-
[12]
Emu: Enhancing image generation models using photogenic needles in a haystack, 2023
Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, Matthew Yu, Abhishek Kadian, Filip Radenovic, Dhruv Mahajan, Kunpeng Li, Yue Zhao, Vladan Petrovic, Mitesh Kumar Singh, Simran Motwani, Yi W...
2023
-
[13]
Consistent diffusion models: Mitigating sampling drift by learning to be consistent
Giannis Daras, Yuval Dagan, Alexandros G Dimakis, and Constantinos Daskalakis. Consistent diffusion models: Mitigating sampling drift by learning to be consistent. arXiv preprint arXiv:2302.09057, 2023
2023 arXiv
-
[14]
Soft diffusion: Score matching with general corruptions
Giannis Daras, Mauricio Delbracio, Hossein Talebi, Alex Dimakis, and Peyman Milanfar. Soft diffusion: Score matching with general corruptions. Transactions on Machine Learning Re- search, 2023. ISSN 2835-8856. URL https://openreview.net/forum?id=W98rebBxlQ
2023
-
[15]
Ambient diffusion: Learning clean distributions from corrupted data
Giannis Daras, Kulin Shah, Yuval Dagan, Aravind Gollakota, Alex Dimakis, and Adam Klivans. Ambient diffusion: Learning clean distributions from corrupted data. In Thirty-seventh Con- ference on Neural Information Processing Systems, 2023. URL https://openreview.net/ forum?id=w...
2023
-
[16]
Consistent diffusion meets tweedie: Training exact ambient diffusion models with noisy data
Giannis Daras, Alexandros G Dimakis, and Constantinos Daskalakis. Consistent diffusion meets tweedie: Training exact ambient diffusion models with noisy data. arXiv preprint arXiv:2404.10177, 2024
2024 arXiv
-
[17]
How much is a noisy image worth? data scaling laws for ambient diffusion
Giannis Daras, Yeshwanth Cherapanamjeri, and Constantinos Costis Daskalakis. How much is a noisy image worth? data scaling laws for ambient diffusion. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum? id=qZwtPEw2qN
2025
-
[18]
Inversion by direct iteration: An alternative to denoising diffusion for image restoration
Mauricio Delbracio and Peyman Milanfar. Inversion by direct iteration: An alternative to denoising diffusion for image restoration. arXiv preprint arXiv:2303.11435, 2023
2023 arXiv
-
[19]
Imagenet: A large- scale hierarchical image database
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large- scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. doi: 10.1109/CVPR.2009.5206848
2009
-
[20]
Diffusion models beat gans on image synthesis
Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021
2021
-
[21]
Diffusion is spectral autoregression, 2024
Sander Dieleman. Diffusion is spectral autoregression, 2024. URL https://sander.ai/ 2024/09/02/spectral-autoregression.html
2024
-
[22]
Optimizing ml training with metagradient descent
Logan Engstrom, Andrew Ilyas, Benjamin Chen, Axel Feldmann, William Moses, and Aleksander Madry. Optimizing ml training with metagradient descent. arXiv preprint arXiv:2503.13751, 2025
2025 arXiv
-
[23]
Self-consuming generative models with curated data provably optimize human preferences
Damien Ferbach, Quentin Bertrand, Avishek Joey Bose, and Gauthier Gidel. Self-consuming generative models with curated data provably optimize human preferences. arXiv preprint arXiv:2407.09499, 2024
2024 arXiv
-
[24]
Datacomp: In search of the next generation of multimodal datasets
Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023
2023 arXiv
-
[25]
Geneval: An object-focused framework for evaluating text-to-image alignment, 2023
Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment, 2023. URL https://arxiv.org/abs/2310.11513
2023 arXiv
-
[26]
Scaling laws for data filtering–data curation cannot be compute agnostic
Sachin Goyal, Pratyush Maini, Zachary C Lipton, Aditi Raghunathan, and J Zico Kolter. Scaling laws for data filtering–data curation cannot be compute agnostic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22702–22711, 2024
2024
-
[27]
Benchmarking neural network robustness to common corruptions and perturbations
Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019
1903 arXiv
-
[28]
Gans trained by a two time-scale update rule converge to a local nash equilibrium
Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017
2017
-
[29]
Denoising diffusion probabilistic models
Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840–6851, 2020
2020
-
[30]
Jacobs, Michael I
Robert A. Jacobs, Michael I. Jordan, Steven J. Nowlan, and Geoffrey E. Hinton. Adaptive Mixtures of Local Experts.Neural Computation, 3(1):79–87, March 1991. ISSN 0899-7667. doi: 10.1162/neco.1991.3.1.79. URL https://doi.org/10.1162/neco.1991.3.1.79. _eprint: https://direct.mi...
1991 doi
-
[31]
Adaptive data optimization: Dynamic sample selection with scaling laws
Yiding Jiang, Allan Zhou, Zhili Feng, Sadhika Malladi, and J Zico Kolter. Adaptive data optimization: Dynamic sample selection with scaling laws. arXiv preprint arXiv:2410.11820, 2024
2024 arXiv
-
[32]
An analytic theory of creativity in convolutional diffusion models
Mason Kamb and Surya Ganguli. An analytic theory of creativity in convolutional diffusion models. arXiv preprint arXiv:2412.20292, 2024
2024 arXiv
-
[33]
A style-based generator architecture for generative adversarial networks
Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401–4410, 2019
2019
-
[34]
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. arXiv preprint arXiv:2206.00364, 2022. 15
2022 arXiv
-
[35]
Analyzing and improving the training dynamics of diffusion models
Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. In Proc. CVPR, 2024
2024
-
[36]
PyTorch Im- age Quality: Metrics and measure for image quality assessment, 2019
Sergey Kastryulin, Dzhamil Zakirov, and Denis Prokopenko. PyTorch Im- age Quality: Metrics and measure for image quality assessment, 2019. URL https://github.com/photosynthesis-team/piq. Open-source software available at https://github.com/photosynthesis-team/piq
2019
-
[37]
Sergey Kastryulin, Jamil Zakirov, Denis Prokopenko, and Dmitry V . Dylov. Pytorch image quality: Metrics for image quality assessment, 2022. URL https://arxiv.org/abs/2208. 14818
2022
-
[38]
Ambient- flow: Invertible generative models from incomplete, noisy measurements
Varun A Kelkar, Rucha Deshpande, Arindam Banerjee, and Mark A Anastasio. Ambient- flow: Invertible generative models from incomplete, noisy measurements. arXiv preprint arXiv:2309.04856, 2023
2023 arXiv
-
[39]
Adam: A method for stochastic optimization
Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014
2014 arXiv
-
[40]
Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick
Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything, 2023. URL https://arxiv.org/abs/2304.02643
2023 arXiv
-
[41]
Learning multiple layers of features from tiny images
Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. 2009
2009
-
[42]
Dimakis, Yair Carmon, Achal Dave, Ludwig Schmidt, and Vaishaal Shankar
Jeffrey Li, Alex Fang, Georgios Smyrnis, Maor Ivgi, Matt Jordan, Samir Gadre, Hritik Bansal, Etash Guha, Sedrick Keh, Kushal Arora, Saurabh Garg, Rui Xin, Niklas Muennighoff, Reinhard Heckel, Jean Mercat, Mayee Chen, Suchin Gururangan, Mitchell Wortsman, Alon Albalak, Yonatan ...
2024
-
[43]
Lawrence Zitnick, and Piotr Dollár
Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dollár. Microsoft coco: Common objects in context, 2015. URL https://arxiv.org/abs/1405.0312
2015 arXiv
-
[44]
Imaging an evolving black hole by leveraging shared structure
Yvette Y Lin, Angela F Gao, and Katherine L Bouman. Imaging an evolving black hole by leveraging shared structure. ICASSP, 2024
2024
-
[45]
Adg: Ambient diffusion-guided dataset recovery for corruption-robust offline reinforcement learning
Zeyuan Liu, Zhihe Yang, Jiawei Xu, Rui Yang, Jiafei Lyu, Baoxiang Wang, Yunjian Xu, and Xiu Li. Adg: Ambient diffusion-guided dataset recovery for corruption-robust offline reinforcement learning. arXiv preprint arXiv:2505.23871, 2025
2025 arXiv
-
[46]
SFBD: A method for training diffusion models with noisy data
Haoye Lu, Qifan Wu, and Yaoliang Yu. SFBD: A method for training diffusion models with noisy data. In Frontiers in Probabilistic Inference: Learning meets Sampling , 2025. URL https://openreview.net/forum?id=6HN14zuHRb
2025
-
[47]
completely blind
Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209–212, 2012
2012
-
[48]
Scalable diffusion models with transformers, 2023
William Peebles and Saining Xie. Scalable diffusion models with transformers, 2023. URL https://arxiv.org/abs/2212.09748
2023 arXiv
-
[49]
The fineweb datasets: Decanting the web for the finest text data at scale
Guilherme Penedo, Hynek Kydlíˇcek, Anton Lozhkov, Margaret Mitchell, Colin Raffel, Leandro V on Werra, Thomas Wolf, et al. The fineweb datasets: Decanting the web for the finest text data at scale. arXiv preprint arXiv:2406.17557, 2024
2024 arXiv
-
[50]
Learning diffusion priors from observations by expectation maximization
François Rozet, Gérôme Andry, François Lanusse, and Gilles Louppe. Learning diffusion priors from observations by expectation maximization. arXiv preprint arXiv:2405.13712, 2024
2024
-
[51]
Laion- 5b: An open large-scale dataset for training next generation image-text models
Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion- 5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Inform...
2022
-
[52]
Stretch- ing each dollar: Diffusion training from scratch on a micro-budget
Vikash Sehwag, Xianghao Kong, Jingtao Li, Michael Spranger, and Lingjuan Lyu. Stretch- ing each dollar: Diffusion training from scratch on a micro-budget. arXiv preprint arXiv:2407.15811, 2024
2024 arXiv
-
[53]
Klivans, and Giannis Daras
Kulin Shah, Alkis Kalavasis, Adam R. Klivans, and Giannis Daras. Does generation require memorization? creative diffusion models using ambient diffusion, 2025
2025
-
[54]
Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning
Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Iryna Gurevych and Yusuke Miyao, editors, Proceedings of the 56th Annual Meeting of the Association for Computati...
2018 doi
-
[55]
Outrageously large neural networks: The sparsely-gated mixture-of-experts layer, 2017
Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer, 2017. URL https://arxiv.org/abs/1701.06538
2017 arXiv
-
[56]
Diffusion art or digital forgery? investigating data replication in diffusion models
Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. arXiv preprint arXiv:2212.03860, 2022
2022 arXiv
-
[57]
Understanding and mitigating copying in diffusion models
Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. arXiv preprint arXiv:2305.20086, 2023
2023 arXiv
-
[58]
Denoising diffusion implicit models
Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020
2010 arXiv
-
[59]
Generative modeling by estimating gradients of the data distribution
Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019
2019
-
[60]
Score-based generative modeling through stochastic differential equations
Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020
2011 arXiv
-
[61]
Journeydb: A benchmark for generative image understanding, 2023
Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, Jifeng Dai, Yu Qiao, Limin Wang, and Hongsheng Li. Journeydb: A benchmark for generative image understanding, 2023. URL https://arxiv. org/abs/2307.00716
2023 arXiv
-
[62]
Diffusion with forward models: Solving stochastic inverse problems without direct supervision
Ayush Tewari, Tianwei Yin, George Cazenavette, Semon Rezchikov, Josh Tenenbaum, Frédo Durand, Bill Freeman, and Vincent Sitzmann. Diffusion with forward models: Solving stochastic inverse problems without direct supervision. Advances in Neural Information Processing Systems, 3...
2023
-
[63]
Foundations of computer vision
Antonio Torralba, Phillip Isola, and William T Freeman. Foundations of computer vision. MIT Press, 2024
2024
-
[64]
Exploring clip for assessing the look and feel of images
Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023
2023
-
[65]
Exploring clip for assessing the look and feel of images
Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 2555–2563, 2023
2023
-
[66]
Multiscale structural similarity for im- age quality assessment
Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for im- age quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, volume 2, pages 1398–1402. Ieee, 2003
2003
-
[67]
Image quality assessment: from error visibility to structural similarity
Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4): 600–612, 2004
2004
-
[68]
Diffusiondb: A large-scale prompt gallery dataset for text-to-image genera- tive models
Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image genera- tive models. arXiv preprint arXiv:2210.14896, 2022
2022 arXiv
-
[69]
Restoration score distillation: From corrupted diffusion pretraining to one-step high-quality generation
Yasi Zhang, Tianyu Chen, Zhendong Wang, Ying Nian Wu, Mingyuan Zhou, and Oscar Leong. Restoration score distillation: From corrupted diffusion pretraining to one-step high-quality generation. arXiv preprint arXiv:2505.13377, 2025. 17 A Theoretical Results A.1 Kernel Estimation...
2025
-
[70]
sample (X, Y) ∼ γ (as specified in Fact 1), 2) sample (Z, Z′) ∼ γX,Y (as specified in Fact 3) and
-
[71]
Half-Unet
output ( ˜X, ˜Y ) := (X + Z, Y+ Z ′). Let us argue the following: Lemma A.3. The afore-described sampling procedure ˜γ is a valid coupling of ˜P and ˜Q. Proof. We need to establish that the marginals of ˜γ are ˜P and ˜Q. We will only show that for ( ˜X, ˜Y ) ∼ ˜γ according to ...
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.