Generative data augmentation adds samples from generative models to a classifier's training set. Many influential studies use generators pre-trained on large external datasets, and so mix the value of synthetic data with knowledge transferred through pretraining. We study the case without external data, where each generator learns only from the classifier's own training set, and ask when and why adding generated samples to that training set raises or lowers the classifier's test accuracy. We train classifiers with four backbones on six datasets, two of them medical, adding samples from four generators. We bound the excess risk of a classifier trained with synthetic data, that is, how far its population risk can lie above the minimum risk within its hypothesis class. The bound charges synthetic data for the total variation distance between the real and generated distributions and credits synthetic data for a larger training set. The gain from adding synthetic data, the fall in population risk, is at most the excess risk of the real-only classifier, which trains on the real images alone. Read through the bound, the experiments suggest that synthetic data help most when our classifiers detect only a small discrepancy between the generated and real data and the real-only classifier has room to improve, and that synthetic data help little or hurt when the detected discrepancy is large. Generators with a larger lower bound on the total variation distance, a bound that we estimate from our own classifiers, tend to help less on five of the six datasets, and on the CIFAR datasets the most accurate backbone gains the least. The bound gives these directions but not the size of the largest gains, most of which traditional augmentation also reaches at the same number of optimizer steps.
Figures & tables
Figure 1: Accuracy moves in both directions as we add synthetic data. Test accuracy (%) against the mixing ratio α , at real-data fraction ρ=1 . Rows: backbones. Columns: datasets, each with its own vertical scale. Lines: generators, traditional augmentation and the real-only classifier (black), as three-seed means with ±1 standard-deviation bands. Values: Tabs. A1 to A3.
Figure 2: Accuracy depends on how much real data the classifier has. Test accuracy (%) against the real-data fraction ρ at α=1.0 , with everything else as in Generative Data Augmentation Without External Data: An Empirical and Theoretical Study on Deep Networks . Tables A4 and A5 give the values for the four benchmark datasets.
Figure 3: The generator never sees data beyond the classifier’s training set. Black tags mark the two factors we vary, and ⊕ is the union of Eq. 1 . The last box gives top-1 test accuracy.
Figure 4: Real, synthetic and traditionally augmented images of the six datasets. Synthetic: samples from the flow-matching generator. Trad. aug.: images from the traditional augmentation pool that the classifiers train on.
Figure 5: At the same training budget, the VAE is the lower-fidelity generator. CIFAR-10 samples from the same sheet positions for every generator.
Figure 6: Generators with a larger estimated lower bound help less on five of the six datasets. Gain in test accuracy Err(0)−Err(α) of each mixture at ρ=1 , in points, against L(Gρ) , the largest gap between a classifier’s test error and its error on images from Gρ .
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
CIFAR-10
MNIST
Classifier
Generator
0
0.2
0.4
0.6
0.8
1.0
2.0
4.0
8.0
FID
0
0.2
0.4
0.6
0.8
1.0
2.0
4.0
8.0
FID
ViT
VAE
85.44
85.14
85.49
85.86
86.18
85.50
87.39
87.51
88.07
113.9
99.01
98.99
99.03
98.99
99.09
99.05
99.06
99.19
99.14
20.9
DDPM
85.44
86.78
87.41
87.93
88.83
88.82
90.12
91.96
93.31
11.0
99.01
98.88
98.97
98.91
98.90
98.86
98.85
98.92
98.94
5.0
VP-SDE
85.44
86.38
87.59
88.24
88.63
89.09
90.50
91.54
92.73
6.7
99.01
98.99
99.07
99.04
99.09
99.11
99.08
99.21
99.21
8.0
Flow matching
85.44
86.63
87.34
88.04
88.84
88.74
90.24
91.71
92.60
9.3
99.01
99.02
99.00
98.92
99.02
99.06
99.04
99.11
99.15
3.0
Trad. aug.
85.44
86.27
86.49
86.85
87.18
87.60
89.39
90.46
91.89
37.5
99.01
99.23
99.28
99.42
99.46
99.42
99.49
99.60
99.66
8.9
Appendix
Table A1 : Test accuracy (%) on CIFAR-10 and MNIST at ρ=1 . Three-seed means, largest standard deviation 1.55 (CIFAR-10) and 0.17 (MNIST). Columns: mixing ratio α ( 0 : real-only) and FID. Trad. aug.: traditional augmentation. Bold: best per backbone and dataset.
CIFAR-100
Fashion-MNIST
Classifier
Generator
0
0.2
0.4
0.6
0.8
1.0
2.0
4.0
8.0
FID
0
0.2
0.4
0.6
0.8
1.0
2.0
4.0
8.0
FID
ViT
VAE
58.69
58.13
58.75
58.39
58.70
58.97
61.18
63.27
63.80
101.2
93.72
93.60
93.84
93.71
93.94
93.77
93.74
93.84
93.77
47.4
DDPM
58.69
59.84
61.29
61.18
61.79
63.06
65.53
69.62
73.09
9.4
93.72
93.68
93.87
93.88
93.61
93.82
93.87
93.79
93.65
4.3
VP-SDE
58.69
59.85
61.42
61.46
62.74
63.07
66.37
69.09
72.02
6.3
93.72
93.80
93.87
93.89
94.03
94.02
93.85
94.22
94.10
5.8
Flow matching
58.69
59.79
61.01
61.17
61.54
62.40
65.23
68.13
71.31
11.0
93.72
93.78
93.76
93.80
93.81
93.76
93.74
93.73
93.47
5.0
Trad. aug.
58.69
59.10
60.60
60.41
60.70
61.29
62.65
65.42
68.55
39.7
93.72
93.65
93.88
93.88
94.01
94.08
94.12
94.22
94.48
19.1
Appendix
Table A2 : Test accuracy (%) on CIFAR-100 and Fashion-MNIST at ρ=1 . Layout as in Tab. A1 . Largest standard deviation 1.66 (CIFAR-100) and 0.64 (Fashion-MNIST).
DermaMNIST
BloodMNIST
Classifier
Generator
0
0.2
0.4
0.6
0.8
1.0
2.0
4.0
8.0
FID
0
0.2
0.4
0.6
0.8
1.0
2.0
4.0
8.0
FID
ViT
VAE
77.64
76.43
76.92
76.79
77.52
77.92
77.42
78.50
77.97
54.2
97.83
97.87
97.97
97.90
97.79
97.85
98.06
97.95
97.88
127.3
DDPM
77.64
76.41
75.33
75.46
74.91
75.34
74.90
74.88
72.62
33.3
97.83
97.88
97.85
97.81
97.63
98.10
97.79
97.82
97.91
5.3
VP-SDE
77.64
76.97
76.69
76.21
76.77
76.31
75.71
76.04
75.66
25.2
97.83
97.87
98.10
98.20
98.16
98.19
98.30
98.26
98.34
4.8
Flow matching
77.64
77.69
77.11
76.82
76.23
76.44
75.48
74.78
75.05
21.8
97.83
98.06
98.02
98.19
98.02
98.19
98.29
98.35
98.35
3.4
Trad. aug.
77.64
76.82
77.67
77.97
77.01
78.49
77.24
77.32
78.40
132.5
97.83
97.96
97.87
98.01
98.00
98.16
98.36
97.91
98.50
110.9
Appendix
Table A3 : Test accuracy (%) on DermaMNIST and BloodMNIST at ρ=1 . Layout as in Tab. A1 . Largest standard deviation 1.80 (DermaMNIST) and 0.55 (BloodMNIST).
MNIST
CIFAR-10
Classifier
ρ
α=0
VAE
DDPM
VP-SDE
FM
Trad.
α=0
VAE
DDPM
VP-SDE
FM
Trad.
ResNet-18
0.2
99.09
99.01
98.55
98.67
98.85
99.32
84.20
84.74
85.55
86.26
85.79
87.25
0.4
99.36
99.39
98.96
99.14
99.20
99.55
90.24
89.44
90.95
90.90
90.61
91.35
0.6
99.43
99.45
99.24
99.36
99.35
99.63
92.50
91.51
92.57
92.61
92.62
93.36
0.8
99.57
99.60
99.28
99.29
99.46
99.63
93.76
92.78
93.52
93.54
93.56
94.24
1.0
99.58
99.64
99.25
99.43
99.43
99.67
94.85
93.38
94.51
94.44
94.31
95.11
Appendix
Table A4 : Test accuracy (%) at each real-data fraction ρ on MNIST and CIFAR-10. Mixtures at α=1.0 , three-seed means. α=0 : real-only classifier at that ρ . FM: flow matching. Trad.: traditional augmentation.
CIFAR-100
Fashion-MNIST
Classifier
ρ
α=0
VAE
DDPM
VP-SDE
FM
Trad.
α=0
VAE
DDPM
VP-SDE
FM
Trad.
ResNet-18
0.2
51.07
49.00
53.46
55.39
52.83
55.56
91.99
92.35
92.06
92.17
92.06
92.04
0.4
64.94
60.37
64.76
64.68
63.11
66.91
93.55
93.68
93.27
93.53
93.54
93.49
0.6
71.10
66.89
69.78
69.77
67.79
71.83
94.45
94.34
93.91
94.01
94.01
94.24
0.8
74.22
70.46
72.44
72.53
71.03
75.07
94.91
94.75
94.37
94.44
94.38
94.78
1.0
76.66
72.62
75.93
75.85
74.92
76.85
95.07
94.78
94.65
95.02
94.75
95.04
Appendix
Table A5 : Test accuracy (%) at each real-data fraction ρ on CIFAR-100 and Fashion-MNIST. Layout as in Tab. A4 .
Dataset
VAE
DDPM
VP-SDE
FM
MNIST
0.03
0.13
0.14
0.03
Fashion-MNIST
0.12
0.10
0.15
0.06
CIFAR-10
0.77
0.10
0.03
0.06
CIFAR-100
0.70
0.17
0.08
0.16
DermaMNIST
0.40
0.40
0.23
0.21
BloodMNIST
0.12
0.05
0.01
0.01
Appendix
Table A6: Estimated lower bounds L(Gρ) on dTV(P,Gρ) at ρ=1 . From 36 seed-0 classifiers per generator (4 backbones × 9 values of α ) and about 5,000 fresh draws from Gρ .
Dataset
ρ
VAE
DDPM
VP-SDE
FM
MNIST
0.2
0.06
0.72
0.52
0.31
0.4
0.06
0.28
0.18
0.11
0.6
0.05
0.19
0.15
0.04
0.8
0.03
0.16
0.14
0.04
1.0
0.03
0.13
0.14
0.03
Fashion-MNIST
0.2
0.16
0.20
0.18
0.14
Appendix
Table A7: With less real data, the estimated lower bound on dTV(P,Gρ) rises for the higher-fidelity generators, most on MNIST and CIFAR-100. The classifiers of Tab. A6 , trained at ρ=1 , on 5,000 fresh draws from Gρ refitted at each ρ . Rows at ρ=1 repeat Tab. A6 .
Figure A1 : Samples from every generator on the six datasets. CIFAR-10: preview grids of the runs behind the synthetic datasets. Other rows: the same checkpoints with each generator’s default sampler (DDPM: 1000 ancestral steps, not the 250 DDIM steps of its synthetic datasets).
Datasets
MNIST, Fashion-MNIST, CIFAR-10,
CIFAR-100, DermaMNIST, BloodMNIST
Classifiers
ResNet-18, ViT, Swin-T, ConvNeXt-V2
Generators
VAE, DDPM, VP-SDE, flow matching;
traditional augmentation baseline (no generator)
Splits
official train/test, stratified tenth
of train held out for validation
Appendix
Table A8: Experimental protocol. α : mixing ratio ( α=0 : real-only classifier, shared by all generators). ρ : real-data fraction.
Used for
Package and version
Source
Tensors, training and sampling
PyTorch 2.8.0, Python 3.9, CUDA 12
pytorch/pytorch
ResNet-18, Swin-T, MNIST, Fashion-MNIST and CIFAR
torchvision 0.23.0
pytorch/vision
Crop, flip, affine and RandAugment
torchvision 0.23.0
pytorch/vision
ViT
vit-pytorch 1.24.1
lucidrains/vit-pytorch
ConvNeXt-V2 Femto
timm 1.0.28
huggingface/pytorch-image-models
Flow-matching path and ODE solver
flow_matching 1.0.10
facebookresearch/flow_matching
Appendix
Table A9 : Software, versions and sources. Versions of every reported run. Sources: GitHub owner and repository. Only InceptionV3 in pytorch-fid loads pretrained weights.
Dataset
Official training
Training ( N )
Validation
Test
MNIST
60,000
54,000
6,000
10,000
Fashion-MNIST
60,000
54,000
6,000
10,000
CIFAR-10
50,000
45,000
5,000
10,000
CIFAR-100
50,000
45,000
5,000
10,000
DermaMNIST
7,007
6,306
701
2,005
BloodMNIST
11,959
10,763
1,196
3,421
Appendix
Table A10: Image counts of the three splits. Validation: a stratified tenth of the official training set. Training ( N ): the rest. Test: the official test set.
Table A11 : Backbone settings, identical for every generator and mixing ratio.
Setting
MNIST
Fashion-MNIST
CIFAR-10/100
DermaMNIST / BloodMNIST
base channels
32
32
128
128
channel multipliers
(1, 2)
(1, 2)
(1, 2, 2, 2)
(1, 2, 2, 4)
residual blocks per resolution
2
2
2
2
attention resolutions
(16, 14)
(16, 14)
(16, 14)
(16, 14)
dropout
0.1
0.1
0.1
0.1
optimizer
AdamW
AdamW
AdamW
AdamW
Appendix
Table A12 : Generator network and training settings per dataset. Network rows: the shared U-Net. Training rows: all four generators.
Generator
Setting
Value
DDPM
timesteps T
1000
noise schedule
cosine
label-drop probability (CFG training)
0.1
guidance scale at sampling
0.0
sampler
DDIM, 250 steps
VAE
latent dimension
128 ( 28×28 datasets); 256 (others)
Appendix
Table A13 : Generator-specific settings, identical across datasets unless noted.
Generator
MNIST
Fashion-MNIST
CIFAR-10
CIFAR-100
DermaMNIST
BloodMNIST
VAE
20.9
47.4
113.9
101.2
54.2
127.3
DDPM
5.0
4.3
11.0
9.4
33.3
5.3
VP-SDE
8.0
5.8
6.7
6.3
25.2
4.8
Flow matching
3.0
5.0
9.3
11.0
21.8
3.4
Trad. aug.
8.9
19.1
37.5
39.7
132.5
110.9
Appendix
Table A14: FID of every synthetic dataset at ρ=1 . Seed-0 draws against the full real training set ( Appendix E ). Trad. aug.: traditional augmentation. Bold: lowest per dataset.
Symbol
Meaning
Defined in
Data and distributions
N
number of images in the full real training set Dtrain(1)
Sec. 3.1
ρ
real-data fraction, the share of each class of Dtrain(1) that Dtrain(ρ) keeps
Sec. 3.1
n
number of real training points, n=ρN
Sec. 3.1
α
mixing ratio, so αρN synthetic points join the ρN real ones
Sec. 3.1
P
true distribution of labeled images (x,y)
Sec. 4
Appendix
Table A15 : Notation of Secs. 4 and J . Hats mark estimates. Subscript α : mixing ratio. Dependence on ρ is implicit except in Gρ .
High-quality training data are a major bottleneck in machine learning, motivating increasing use of artificially generated data. We study when a binary classifier trained on synthetic data transfers successfully to a real distribution, under a shared conditional label distribution. First, we derive finite-sample excess-error bounds when the real covariate distribution is absolutely continuous with respect to the synthetic distribution under a controlled density ratio. Second, we study the partially and fully singular case, where the real distribution has mass outside the synthetic one. If the hypothesis class is sufficiently expressive, synthetic risk alone cannot uniformly control real-world error: a classifier can be nearly optimal on synthetic data while approaching worst-case error on the uncovered real region, even when the distributions are arbitrarily close in Wasserstein distance. For function classes where hypotheses and the log-odds of the conditional label probability are Lipschitz, we recover a bound in terms of the probability that a real sample lies outside the reach of the synthetic data, restoring a transfer guarantee even in the fully singular case. Experiments confirm this in controlled low-dimensional settings and on 32,200 images generated as counterparts to real ImageNet images by four state-of-the-art generators: an adversarial classifier reaches 0.5% error on the synthetic domain but 99.25% error on the real domain, while ordinarily trained architectures show positively correlated synthetic and real test errors. Synthetic data should be budgeted at high sample counts and models validated on real data.
Leroy Odunlami, Tobias Riedlinger, Hanno Gottschalk
Institute of Mathematics, Technische Universität Berlin, Berlin, Germany
Generative Adversarial Networks (GANs) can help overcome data scarcity in computer vision tasks by generating additional training samples. In this work, we explore generative data augmentation in two low-resource domains: Bangla handwritten character recognition and chest X-ray image analysis. We use DCGAN-based models trained on 64x64 images to generate synthetic samples and evaluate their quality using Inception Score (IS), Fréchet Inception Distance (FID), and visualization methods such as t-SNE and UMAP. To measure practical usefulness, we train image classifiers using real data and a combination of real and synthetic data. Experimental results show that synthetic augmentation improves data diversity and consistently increases classification performance in limited-data settings. We also investigate training stability techniques, including gradient penalty and spectral normalization, and perform ablation studies on synthetic-to-real data ratios and sample filtering strategies. In addition, we discuss challenges related to medical image evaluation, dataset licensing, and privacy concerns of synthetic data. Our approach is simple, reproducible, and provides a strong baseline for generative augmentation in resource-constrained imaging applications.
Md. Sohanuzzaman Soad, Mahady Al Hady, S M Rafiuddin Rifat +1
Department of Computer Science and Engineering University of Asia Pacific Dhaka, Bangladesh
Generative data augmentation is widely used to mitigate class imbalance, yet its theoretical effect on downstream generalization remains poorly understood. In this work, we develop a statistical framework for conditional generative augmentation and analyze its impact on classification risk. We formalize augmentation as a distribution-mixing process and show that the resulting risk distortion is controlled by both the augmentation strength and the class-conditional Wasserstein discrepancy between real and generated distributions. We further derive a capacity-dependent generalization bound based on Rademacher complexity, revealing an explicit trade-off between hypothesis complexity, augmentation intensity, and generative fidelity. Empirically, we evaluate the framework on binary and multiclass imbalanced classification tasks using Conditional GAN and Conditional WGAN-GP augmentation. Across datasets, CWGAN-GP consistently achieves lower Wasserstein discrepancies than CGAN, indicating improved distributional fidelity. However, improved fidelity does not necessarily translate into superior classification performance, with classical oversampling methods often remaining competitive. These findings support the central theoretical prediction that augmentation reliability is governed by distributional approximation error rather than predictive performance alone. Overall, this work establishes generative augmentation as a distributional perturbation process whose reliability can be quantified through Wasserstein-based measures and supported by finite-sample generalization guarantees. The proposed framework provides a principled foundation for evaluating synthetic data quality beyond classification accuracy alone.
Chathurika S Abeykoon, Mathias Nthiani Muia, Mallory Goldstein