Comprehensive Exploration of Synthetic Data Generation: A Survey

1 Introduction

• Training data scarcity is attributed to high labeling costs and privacy concerns, which hinder models’ ability to generalize effectively.SDG is presented as a source of artificial training samples.Source: §1 · p. 1
• Existing SDG overviews often omit recent advances, focus on one domain, classify few aspects, or cover architectures coarsely.Prior literature includes surveys of particular model types and domains, as well as broader reviews.Source: §1 · p. 2
• The survey examines 417 SDG models from the last decade, identifying 20 model types and 42 subtypes.It also surveys their applications and enhancements and classifies them by generated data, performance, privacy, and training processes.Source: §1 · p. 2
• Computer vision is the most popular application field, while GANs and diffusion models are top performers there; RNNs dominate sequential-data generation.Privacy-preserving generation commonly uses Markov chains, Bayesian networks, and GANs.Source: §1 · p. 2
• The survey reports increasing model complexity and performance, alongside non-standardized evaluation metrics and datasets that challenge comparisons.The authors mention predecessor graphs based on models’ performance evaluations as one possible response.Source: §1 · p. 2

2 Overview of Generative Models

• The overview focuses on approaches released in the last ten years, describing model architectures and surveying their recent use in SDG.Its coverage builds on Harshvardhan et al., adding missing models and selected implementations.Source: §2 · p. 2

2.1 Gaussian Mixture Models

• A GMM represents data with weighted Gaussian components and can generate samples by selecting a component and sampling from its distribution.The survey describes fitting component membership and parameters with expectation-maximization; a deep GMM instead concatenates transformations along a random path.Source: §2.1 · p. 3

2.2 Kernel Density Estimators

• KDEs estimate a probability density between observations, and the survey describes using component-wise KDEs to synthesize photoplethysmogram time series.Sampling from the learned component distributions allows new PPGs to be constructed.Source: §2.2 · p. 4

2.3 Markov Chain Models

• Markov chains model sequences through conditional transition probabilities; they are simple and interpretable but order-1 models miss long-term structure, while higher orders risk overfitting and require more data.The survey also describes their use in interactive music generation through phrase-level modeling and note-by-note continuation.Source: §2.3 · pp. 4–5

2.3.1 Hidden Markov Models

• HMMs add hidden states with emission probabilities to Markov chains, and the Viterbi algorithm can recursively find the most probable state sequence.Applications described include biological sequences, music chord generation, and private location traces.Source: §2.3.1 · pp. 5–6

2.3.2 N-Grams

• N-grams model a word’s probability from the previous n−1 words under a Markov assumption, with Laplace smoothing allowing unseen combinations low probability.The survey notes their use in language modeling and synthetic text generation, including constrained lyric generation.Source: §2.3.2 · p. 6

2.4 Bayesian Networks

• Bayesian networks support synthetic-data generation through hierarchical sampling, including PrivBayes’ noisy marginals for private high-dimensional data and BayesBoost’s targeted oversampling of under-represented cases,.BayesBoost identifies uncertain validation cases and generates similar samples for them to create a less biased dataset.Source: §2.4 · pp. 7–8

2.5 Genetic Algorithms

• Genetic algorithms evolve candidate solutions through selection, crossover, and mutation; applications include theory-based music fitness functions and iterative optimization of categorical-table statistics,.Table GAs are more computationally demanding and error-prone than string-data GAs because of the number of variables and their relationships.Source: §2.5 · pp. 8–9

2.6 Boltzmann Machines

• Boltzmann machines are undirected networks of binary visible and hidden nodes whose interaction parameters are trained by gradient ascent to assign probabilities to visible-unit states.The interaction parameters represent visible-to-hidden, visible-to-visible, and hidden-to-hidden connections.Source: §2.6 · p. 9

2.6.1 Restricted Boltzmann Machines

• Restricted Boltzmann machines permit only visible-to-hidden connections, enabling exact inference and more efficient learning; a convolutional RBM with constrained sampling generates music with local and global structure,.The music-generation process combines a convolutional RBM with self-similarity, tonality, and meter constraints.Source: §2.6.1 · pp. 9–10

2.6.2 Deep Belief Networks

• Convolutional DBNs learned hierarchical image features that improved classification on Caltech-101 and MNIST, while DBN music generation produced stylistically plausible but irregular samples and faced high training time and complex sampling.A two-layer DBN learned edges in its first layer and contours, corners, angles, and surface boundaries in its second; a separate DBN autoencoder copied 56.9% of piano-roll notes on average during reconstruction.Source: §2.6.2 · p. 11

2.6.3 Temporal Restricted Boltzmann Machines and Related Models

• SRTRBMs outperformed earlier temporal Boltzmann models across bouncing-ball, motion-capture, and weather prediction tasks, while RTRBMs enabled exact inference and feasible gradient computation.TRBM inference requires evaluating all states of two RBMs, making heuristics necessary; RTRBMs can be expressed as RNNs and trained with backpropagation through time.Source: §2.6.3 · p. 11

2.6.4 Deep Boltzmann Machines

• Deep Boltzmann machines generate synthetic handwritten digits by Gibbs sampling and produce grayscale toy images, with pixel-translation augmentation improving results on NORB,.The MNIST model was initialized with random binary states and sampled for 100,000 steps.Source: §2.6.4 · p. 12

2.6.5 Gated Boltzmann Machines

• When the number of factors F is similar to the number of hidden and visible variables, the model requires O(N^2) rather than O(N^3) parameters; like, it learns transformation representations on larger 40×40-pixel images and successfully extracts motion.Source: §2.6.5 · p. 13
• CRBMs condition on prior time steps, while stacked CDBNs and gated CRBMs support time-series generation, nonlinear interactions, and mixed or changing motion styles.Evaluated on the CMU motion data set, the models continued motions or created new ones; factorization reduces parameters and predefined style labels guide the gate process.Source: §2.6.5 · p. 13

2.7 Autoencoders

• Autoencoders map inputs to hidden representations and reconstruct them, learning parameters by minimizing reconstruction error with a data-compatible loss such as squared error.Hidden layers can be stacked to learn higher-level representations or pre-train weights for a later neural network; the decoder weights may optionally be tied to the encoder weights.Source: §2.7 · p. 13
• Unlike usually directed autoencoders, RBMs are undirected and sample binary hidden representations, whereas autoencoders use real-valued mappings and are deterministic feedforward networks.The two model types also differ in their training procedures.Source: §2.7 · pp. 13–14

2.7.1 Helmholtz Machines

• The Helmholtz machine uses wake-sleep training: inputs propagate bottom-up during wake, while sleep generates lower-layer states top-down and updates recognition weights.The model is a directed deep generative model with binary units, biases, and generative and recognition weights.Source: §2.7.1 · p. 14

2.7.2 Denoising Autoencoder

• Walkback-trained DAEs produce more natural-looking MNIST Markov-chain samples than DAEs trained with simple corruption.Walkback creates divergent negative examples through random-length walks in the model’s Markov chain, helping keep predictions within a plausible range.Source: §2.7.2 · pp. 14–15

2.7.3 Contractive Autoencoder

• CAE results were slightly better on TFD and worse on MNIST than a two-layer DBN in generated-sample log-likelihood, while reducing sensitivity to MNIST digit deformations.The CAE penalizes the input sensitivity of hidden representations using the Frobenius norm of the feature Jacobian.Source: §2.7.3 · pp. 15–16
• Sampling and interpolating at higher-level CAE and DBN representations yielded higher log-likelihood and more plausible digit samples, with greater class variation reported for sampling.The authors tested representations generated by Markov chains at multiple depths.Source: §2.7.3 · p. 16

2.7.4 Generative Stochastic Network

• A GSN generalizes deep denoising autoencoders as a Markov-chain framework trained with back-propagated gradients; its DBM-inspired adaptation trains across repeated encodings and reconstructions.The adaptation sums log-likelihoods against the starting training sample, following the walkback objective.Source: §2.7.4 · pp. 16–17

2.7.5 Variational Autoencoder

• VAEs encode inputs as Gaussian latent distributions and decode samples; their aggregated latent distribution is constrained to N(0, I) for random-vector generation.On MNIST and Frey Face, an MLP VAE achieved faster convergence and higher marginal likelihood than the wake-sleep reference.Source: §2.7.5 · p. 17
• DRAW generated highly realistic MNIST and SVHN images but overfit CIFAR-10 with 50,000 training samples; its LSTM encoder and decoder use attention at each step.The model iteratively refines images in a process modeled on human drawing.Source: §2.7.5 · p. 17
• β-VAE outperformed InfoGAN and DC-IGN on disentanglement, while a mixture-prior VAE achieved state-of-the-art log-likelihood against a normal VAE and avoided its local-optima problem.β-VAE uses an added β hyperparameter to encourage more independent latent factors; the mixture-prior model was evaluated across several datasets.Source: §2.7.5 · p. 18

2.7.6 Deep Latent Gaussian Models

• A three-layer Deep Latent Gaussian Model produced realistic-looking samples comparable to contemporary RBMs and DBNs and was also evaluated for imputing missing data.Its directed generative process uses layer-wise Gaussian latent variables and stochastic backpropagation.Source: §2.7.6 · p. 21

2.7.7 Gated Autoencoders

• A pyramid of stacked GAEs learned transformations between time-series observations and predicted future observations, outperforming an RNN and RBM in mean squared error.The models were tested on chirps and video frames of bouncing balls and NORB objects.Source: §2.7.7 · p. 21

2.7.8 Masked Autoencoders

• MADE estimates joint data probabilities autoregressively by masking future inputs, and can sample binary MNIST images from the computed probabilities.Masking removes computational paths from future inputs to outputs, supporting autoregressive tasks.Source: §2.7.8 · p. 21

2.8 Neural Autoregressive Distribution Estimators

• NADE models a joint distribution as ordered one-dimensional conditionals; RNADE uses a Gaussian mixture model for real-valued outputs, while deep NADE often improves test-set log-likelihood over single-layer models.Shared-weight training can optimize across variable orderings and determine a suitable ordering in constant time.Source: §2.8 · pp. 22–23

2.9 Sparse Coding

• Variational Sparse Coding combines sparse coding with a VAE recognition model to improve latent feature disentanglement, with evaluations on image and human-activity datasets.For image super-resolution, a separate sparse-representation approach assumes low- and high-resolution patches share a sparse code.Source: §2.9 · pp. 23–24

2.10 Recurrent Neural Networks

• Other RNN applications included realistic handwriting generation and image captioning; LSTM video prediction produced blurry future frames, although its representations worked well for action recognition.The image-captioning model reported state-of-the-art BLEU, METEOR, and CIDER scores for its time on several datasets.Source: §2.10 · p. 25
• For music generation, reinforcement-learning-optimized Note-RNN produced more coherent results complying with musical theory, while an LSTM piano model generated natural-sounding performances but did not choose a coherent play style.The piano model generated from a starting sequence or an empty sequence without copying training samples.Source: §2.10 · pp. 25–26
• RNN-based work spans image, text, video, and music generation; PixelRNN’s diagonal BiLSTM outperformed its row LSTM and PixelCNN, DRAW, and DLGMs on image density estimation.RNNs process sequential data while retaining a memory of the past; LSTMs and GRUs are described as addressing the vanishing-gradient problem.Source: §2.10 · pp. 24–25

2.11 Convolutional Neural Networks

• CNNs handle large inputs such as images or speech with significantly fewer trainable parameters than fully connected neural networks.The described architecture uses convolutions, nonlinearities, pooling, and a final fully connected layer.Source: §2.11 · p. 30
• PixelCNN masks future pixels and predicts discrete RGB values sequentially; it can parallelize training and test-image evaluation but performs worse than PixelRNN.It has also been adapted as a decoder for video continuation.Source: §2.11 · p. 31
• WaveNet uses dilated causal convolutions to model raw audio in order, enabling parallel training but requiring sequential generation; it achieved state-of-the-art text-to-speech results.It can condition on additional inputs such as text and speaker identity and can generate musical fragments.Source: §2.11 · p. 31, Figure 26 · p. 32
• CNN-based style transfer iteratively optimizes a white-noise image against style and content losses derived from reference images.The losses compare the style and content representations extracted by a deep CNN.Source: §2.11 · p. 31
• A recursively supervised CNN for image super-resolution outperformed SRCNN and produced clearer images, using reconstructed intermediate layers and a skip connection.The weighted average of layer predictions addresses vanishing or exploding gradients and uncertainty about recursion depth during training.Source: §2.11 · pp. 31–32
• PixelCNN++ reduced training cost by using rounded continuous RGB distributions instead of 256-way softmax and added downsampling, shortcut connections, and dropout.The implementation reported state-of-the-art results on class-conditional and unconditional CIFAR-10 image generation.Source: §2.11 · p. 32

2.12 Transformers

• Transformers outperform previous state-of-the-art architectures, including RNNs, CNNs, and GANs, in processible image size and quality.The Image Transformer predicts each pixel’s RGB values from a local neighborhood using 1D or 2D attention.Source: §2.12 · pp. 33–34
• Sparse attention reduces time and memory requirements from O(n^2) to O(n√n) without performance loss, with state-of-the-art density modeling results.The sparse transformer also generates large unconditional sequences in image and audio domains.Source: §2.12 · p. 34
• VideoBERT trains on paired video representations and text, supporting action classification, video captioning, and future-token forecasting with coherent results and state-of-the-art captions.Source: §2.12 · pp. 34–35
• A graph transformer replacing GraphRNN’s edge-output network with a self-attention decoder achieves competitive or better results across varied metrics.Source: §2.12 · p. 35

2.13 Generative Adversarial Networks

• The original GAN uses a generator and discriminator in an adversarial game, but joint training can be unstable and mode collapse can reduce output diversity.The original experiments achieved first- and second-place log-likelihood results on MNIST and TFD, respectively, against DBN, deep GSN, and stacked CAE baselines.Source: §2.13 · pp. 35–36
• DCGAN adapts convolutional networks for unsupervised GAN learning, while WaveGAN generates raw-waveform audio and produces higher sound quality than SpecGAN.SpecGAN has the higher Inception Score and human label accuracy in the reported comparison.Source: §2.13 · pp. 36–37

2.13.1 Conditional GANs

• Conditional GANs condition the generator and discriminator on additional information, such as class labels, to control the type of generated data.Source: §2.13.1 · p. 43
• SAGAN raises Inception Score from 36.8 to 52.52 and lowers FID from 27.62 to 18.65 versus previous state-of-the-art models.Source: §2.13.1 · p. 45
• BigGAN achieves state-of-the-art class-conditional image synthesis at resolutions up to 512 × 512, with a controllable trade-off between detail and variety.Source: §2.13.1 · p. 45

2.13.2 Deep/Stacked GANs

• LAPGAN’s coarse-to-fine image generation produces more realistic results than a standard GAN in log-likelihood estimates and human evaluation.Source: §2.13.2 · p. 49
• StackGAN separates text-to-image generation into stages and outperforms GAN-INT-CLS and GAWWN in resolution and realism, reaching 256 × 256 pixels.Source: §2.13.2 · p. 49

2.13.3 Bidirectional GANs

• BigBiGAN achieves state-of-the-art unsupervised representation learning and unconditional image generation, outperforming BiGAN and BigGAN on those respective tasks.Source: §2.13.3 · p. 50

2.13.4 Adversarial Autoencoders

• An adversarial autoencoder matches encoded latent representations to a user-defined prior, enabling configurable representations for tasks including classification and clustering.On MNIST, SVHN, and TFD, the AAE achieved state-of-the-art test log-likelihood against GAN, GMMN, DBN, and GSN baselines.Source: §2.13.4 · p. 50

• GMMNs map samples from an easy-to-sample distribution through a neural network to produce data samples.Source: §2.14 · p. 51
• Unlike GAN generators, GMMNs train by minimizing the Maximum Mean Discrepancy criterion, which compares the generating distributions of two sample sets.Its differentiable Gaussian-kernel criterion permits gradient backpropagation, and higher-dimensional data requires more training data to estimate the distribution.Source: §2.14 · p. 51
• Training a GMMN on autoencoder latent representations reduces data dimensionality and, consequently, the training data needed.Source: §2.14 · p. 52
• MMD GAN replaces the fixed Gaussian kernel with an adversarially learned kernel, aiming to reduce required training data and improve accuracy and efficiency.Source: §2.14 · p. 52

2.15 Plug & Play Generative Networks

• DGN-AM synthesizes images by maximizing a classifier neuron’s activation through a pretrained generator, producing state-of-the-art quality but little diversity.Source: §2.15 · p. 52
• Learning a latent-variable prior with a denoising autoencoder improves DGN-AM’s diversity; generalized PPGNs achieve higher sample quality and diversity than DGN-AM.Source: §2.15 · p. 52

2.16 Copulas

• DPCopula synthesizes differentially private multidimensional data and outperforms PSD and P-HP, especially on high-dimensional datasets.The study analyzes maximum likelihood estimation and Kendall’s τ correlation for privacy guarantees and computational complexity on census datasets.Source: §2.16 · p. 53
• Vine copulas model mobility-trajectory statistical and semantic/geographic similarities at a fraction of neural-network computational cost, while less accurate RNN and GAN models perform better in privacy tests.Source: §2.16 · p. 53

2.17 Normalizing Flow Models

• Normalizing flows are invertible transformations between a base distribution and observational space, enabling exact density calculation or sampling in observation space.Source: §2.17 · p. 53
• GraphAF generates graphs sequentially, predicting each next node before its edges and using valency checking to reject invalid edges.It also uses reinforcement learning for goal-directed molecule generation.Source: §2.17 · p. 54

2.18 Reinforcement Learning

• An RL agent observes states, chooses actions, and uses rewards to learn a policy that maximizes cumulative reward.The policy may be deterministic or stochastic, and the agent need not have complete knowledge or control of its environment.Source: §2.18 · p. 54
• Oh et al. [449] condition Atari future-frame generation on previous frames and player actions, producing realistic 100-step predictions.A CNN, optionally paired with an RNN, encodes frames; the prediction is upsampled into a full-size frame.Source: §2.18 · p. 55
• Paintbot [450] selects canvas strokes from local reference and canvas patches, updating the canvas until a similarity threshold is reached.Strokes specify angle, length, color, and width; training also fits a neural-network reward predictor.Source: §2.18 · p. 55, Figure 40 · p. 55
• Krishna et al. [451] synthesize full-resolution CT images by having an RL agent create anatomical masks that a conditional GAN fills with textures.A CNN classifier predicts rewards for masks using human feedback.Source: §2.18 · p. 55

2.19 Diffusion Models

• The first diffusion implementation generated and inpainted images and binary sequences, performing worse than GANs but better than DBNs, GSNs, and CAEs in log-likelihood.Diffusion models iteratively add Gaussian noise in a forward Markov process and learn a reverse process that maps noise back to the data distribution.Source: §2.19 · p. 56, Figure 41 · p. 56
• DDPM [461] simplifies diffusion training by predicting added noise, while improved DDPM [463] restores trainable variance, uses a cosine noise schedule, and reduces sampling steps.The improved model reports better log-likelihood, and increasing model size also increases performance.Source: §2.19 · pp. 56–57
• ADM-G [464] adds multiscale attention, class conditioning, a deeper architecture, and classifier guidance; Ramesh et al. [465] enable text-conditioned image generation and manipulation.Text-conditioned manipulation reconstructs images from noise using new or interpolated text embeddings.Source: §2.19 · p. 57

2.20 Virtual Environments

• Virtual environments range from standalone 3D models to complex virtual worlds and interactive rule-based settings; multistage training and real-camera flaws can improve rendered-data performance.More realistic lighting does not necessarily help, according to Mayer et al..Source: §2.20 · p. 57

2.20.1 Graphic Models

• Graphic-model datasets include MPI-Sintel for optical flow and RGB-D room renderings for visual odometry, reconstruction, and SLAM evaluation.SceneNet generates annotated indoor environments, while its renderings have been used to train depth-based semantic segmentation models.Source: §2.20.1 · pp. 57–58
• Su et al. [472] report state-of-the-art viewpoint-estimation performance using rendered 3D models over real backgrounds at negligible human cost.The approach uses existing 3D model repositories.Source: §2.20.1 · p. 57
• Peng et al. [473] find that detectors trained on synthetic images of non-photorealistic CAD models outperform models trained on real data from a different domain.The detectors target novel object categories absent from the real training data.Source: §2.20.1 · p. 57

2.20.2 Virtual Worlds

• Game-derived training data from Grand Theft Auto V boosts segmentation accuracy and reduces reliance on labeled real images; another simulated-image-only model achieves state-of-the-art vehicle detection on real data.The approaches obtain semantic labels or annotations through game rendering and GPU data.Source: §2.20.2 · p. 58
• SYNTHIA provides semantically segmented urban scenes and seasonal driving sequences, and combining it with real data can significantly improve segmentation accuracy.The dataset includes views from multiple perspectives and a virtual car equipped with multiple cameras and depth sensors.Source: §2.20.2 · pp. 58–59, Figure 42 · p. 58
• Other virtual-world resources include computer-game data for segmentation and depth estimation, Virtual KITTI, VIPER, and Meta-Sim.VIPER contains more than 250,000 annotated high-resolution video frames across perception tasks.Source: §2.20.2 · p. 59

2.20.3 Interactive Environments

• Interactive environments support RL training and evaluation, from Atari’s ALE and VizDoom to CAD2RL, which trains in randomized virtual settings and demonstrates real-drone flight indoors.SC2LE agents behaved like novice players in mini-games and made no noticeable progress in the main StarCraft game.Source: §2.20.3 · p. 59
• Additional interactive platforms include DeepMind Lab, TorchCraft, and Project Malmo for designing or interfacing with varied game-based tasks.The described tasks include navigation, survival, collaboration, and problem-solving.Source: §2.20.3 · p. 59

3.1 Criteria for Classification

• The classification records metadata, data structure and type, conditioning inputs, sampling process, training losses and optimization, datasets, performance comparisons, and privacy status.Because performance metrics are not commonly shared and often depend on data type or domain, the survey records which predecessors each model outperformed when available.Source: §3.1 · p. 60

3.3 Metadata

• GANs quickly surpassed other model categories after their 2014 proposal, while Markov chains and Boltzmann machines declined and RL and diffusion gained popularity.GANs also received the most citations over the last decade; RNNs, CNNs, and autoencoders often received comparable attention.Source: §3.3 · p. 61, Figure 44 · p. 61, Figure 45 · p. 62
• CNNs are the most common submodels, and surveyed architectures show strong cross-category reuse, including RNNs in 70% of RL models.GANs and GMMNs often use autoencoder decoders; diffusion, GAN, and normalizing-flow models heavily use CNNs, while four of five surveyed diffusion models depend significantly on one another.Source: §3.3 · p. 61, Figure 46 · p. 63, Figure 47 · p. 64

3.4 Data Structure

• Markov chains, RNNs, and transformers mainly generate arbitrarily sized data, whereas fixed-size architectures can support sequences through recurrent components.RL also handles unlimited lengths, aided by reward-based learning during generation; modern Boltzmann machines such as TRBM are linked to RNNs.Source: §3.4 · pp. 61–62, Figure 48 · p. 65

3.5 Data Types

• Natural image generation is the most prominent data type; RNNs mainly generate arbitrary-length sequences, while autoencoders and GANs span nearly all data types.Boltzmann machines are limited to time series, video, and images; virtual environments generate visual data, and RL suits sequential generation.Source: §3.5 · p. 63, Figure 49 · p. 65, Figure 50 · p. 66

3.6 Sampling Requirements

• Genetic algorithms, PPGNs, sparse coding models, and CNNs can always condition generation on additional information, while many other model types also support conditioning.Images and class labels are the most common inputs; Markov chains suit model constraints, RL suits reward functions, and segmentation masks are used exclusively with CNNs and GANs in the surveyed literature.Source: §3.6 · pp. 63, 65, 67, Figure 51 · p. 66, Figure 52 · p. 67

3.7 Sampling Process

• Most surveyed models use one-shot sampling, while PPGNs, genetic algorithms, and diffusion models always sample iteratively.Autoencoders, CNNs, Markov chains, NADEs, RL, and RNNs can also iteratively refine outputs, but are less often used this way.Source: §3.7 · p. 67, Figure 53 · p. 68

3.8 Training Process

• Training approaches include distribution-matching and adversarial objectives, reconstruction and cycle-consistency losses, reward or rule-based guidance, and varied optimization adaptations.Examples include MMD for GMMNs, Wasserstein distance alongside GAN adversarial loss, and gradient or parameter restrictions to improve training.Source: §3.8 · pp. 67–68

3.9 Data Sets

• 355 of 388 models used at least one public dataset for evaluation, 33 did not disclose their data, and papers used 1.74 datasets on average.Forty-seven models used both private and public datasets; five of the 33 models not disclosing data were used for privacy preservation.Source: §3.9 · p. 69

3.10 Model Performance

• Newer models usually outperform older ones; GANs and transformers often lead, RNNs perform well for graph and molecule generation, and diffusion models outperform GANs in image generation.GANs were compared against most often (83 times), followed by autoencoders (62) and RNNs (44).Source: §3.10 · p. 70, Figure 56 · p. 71

3.11 Privacy

• Privacy-preserving generation usually relies on simpler, human-inspectable models; GANs are an exception, while privacy protections can reduce data utility and complex models may encode training samples.Differentially private models use privacy budgets to limit disclosure, and teacher forcing helps RNNs avoid error propagation; private SDG in the survey covered tabular data, graphs, and time series, not higher-dimensional data.Source: §3.8 · p. 69, §3.11 · pp. 72–73, §3.12 · p. 73

3.12 Summary

• Direct model comparison is often impossible because studies use different metrics and datasets; training and sampling costs, time, and data requirements are also seldom reported.The survey recommends standardized benchmarks using common models, datasets, and predetermined metrics.Source: §3.10 · p. 70, §3.12 · p. 73

3.13 Guideline for Synthetic Data Generation

• For image generation, diffusion models and GANs are quality-focused options, GANs suit limited training data, and autoencoders offer a less resource-intensive alternative; transformers and RNNs suit sequential data.Markov chains are a simple, efficient, inspectable option for sequential data, while autoencoders, especially VAEs, can learn and recombine features to produce data with desired properties.Source: §3.13 · p. 73
• CNNs help process high-dimensional data, while reinforcement learning and classifier-guided GANs can steer generation toward desired properties and reduce training-data needs.Virtual environments suit visual data requiring accurate labels and configurability, but can require substantial human effort to build and configure.Source: §3.13 · p. 74
• For private data, the guidance recommends modified-probability or noise-injected Markov chains and Bayesian networks, or correctly applied GANs.The survey also identifies GANs as flexible and able to generate large amounts of data without directly training on the training data.Source: §3.12 · p. 73, §3.13 · p. 74

4 Related Work

• Prior work includes focused reviews of particular SDG domains or model families, while this section distinguishes those from comprehensive reviews.The focused reviews are described as limited in scope and model selection.Source: §4 · p. 74, §4.1 · p. 76

4.1 Domain- or Model-Specific Overviews and Comparisons

• Domain- or model-specific reviews cover topics such as synthetic image pipelines, graph generation, healthcare, and privacy-enhancing data.They also examine needs such as image-generation requirements and the trade-off between privacy and data utility.Source: §4.1 · pp. 75–76

4.2 Comprehensive Reviews

• Comprehensive reviews classify generative models and discuss their applications, while some also compare models or identify evaluation criteria across domains.Examples include model taxonomies and reviews of which evaluation criteria matter for NLP, audio, healthcare, and mobility.Source: §4.2 · pp. 76–77
• Existing comprehensive reviews often miss transformers or domains such as music, and extensive model-selection guidance is uncommon.The passage identifies extensive guidance in and coarser guidance in.Source: §4.2 · p. 77

5 Conclusion

• The survey covers 417 SDG model papers, classifying them into 20 distinct types and 42 subtypes.The classification criteria were drawn from related work and identified during the survey.Source: §5 · p. 77
• Computer vision is the most popular SDG application field, and GANs are the most popular SDG models.RNNs and transformers are more suitable for sequential data, while CNNs, GANs, and autoencoders are mostly used for static-size data.Source: §5 · p. 77
• SDG models are often combined: GANs and autoencoders can frame RNN or CNN components, while RNNs may be paired with reinforcement learning.Source: §5 · p. 78
• The survey reports that SDG performance evaluation lacks standardized metrics and datasets, limiting comparability among models.It identifies common metrics, datasets, and reference models as needed for enhanced comparability.Source: §5 · p. 78
• Privacy-preserving data generation is at a nascent stage, with simpler models prevalent and GANs the only more complex neural-network approach identified.The simpler models include Markov chains, Bayesian Networks, and genetic algorithms.Source: §5 · p. 78
• The authors recommend future work on training and sampling costs and systematic evaluation to improve understanding, usability, and comparability.Source: §5 · p. 78

A Acronyms

• AI means Artificial Intelligence, AIC means Akaike information criterion, ALI means Adversarially Learned Inference, and BN means Bayesian Network.Source: A Acronyms · p. 102
• BPTT means Backpropagation Through Time, BiGAN means Bidirectional GAN, CAD means Computer Aided Design, and CNN means Convolutional Neural Network.Source: A Acronyms · p. 102
• CAE means Contractive Autoencoder, CT means Computed Tomography, DBM means Deep Boltzmann Machine, and DDPM means Denoising Diffusion Probabilistic Model.Source: A Acronyms · p. 102
• GAN means Generative Adversarial Net, GMM means Gaussian Mixture Model, GPU means Graphics Processing Unit, and GRU means Gated Recurrent Unit.Source: A Acronyms · p. 102
• LSTM means Long Short-Term Memory, MMD means Maximum Mean Discrepancy, NLP means Natural Language Processing, and VAE means Variational Autoencoder.Source: A Acronyms · p. 103