Latent Feature Space Research Articles

Ablation Study of Diffusion Model with Transformer Backbone for Low-count PET Denoising.

Diffusion models (DM) built from a hierarchy of denoising autoencoders have achieved remarkable progress in image generation, and are increasingly influential in the field of image restoration (IR) tasks. In the meantime, its backbone of autoencoders also evolved from UNet to vision transformer, e.g. Restormer. Therefore, it is important to disentangle the contribution of backbone networks and the additional generative learning scheme. Notably, DM shows varied performance across IR tasks, and the performance of recent advanced transformer-based DM on PET denoising is under-explored. In this study, we further raise an intuitive question, "{if we have a sufficiently powerful backbone, whether DM can be a general add-on generative learning scheme to further boost PET denoising}". Specifically, we investigate one of the best-in-class IR models, i.e., DiffIR, which is a latent DM based on the Restormer backbone. We provide a qualitative and quantitative comparison with UNet, SR3 (UNet+pixel DM), and Restormer, on the 25% low dose 18F-FDG whole-body PET denoising task, aiming to identify the best practices. We trained and tested on 93 and 12 subjects, and each subject has 644 slices. It appears that Restormer outperforms UNet in terms of PSNR and MSE. However, additional latent DM over Restormer does not contribute to better MSE, SSIM, or PSNR in our task, which is even inferior to the conventional UNet. In addition, SR3 with pixel space DM is not stable to synthesize satisfactory results. The results are consistent with the natural image super-resolution tasks, which also suffer from limited spatial information. A possible reason would be the denoising iteration at latent feature space cannot well support detailed structure and texture restoration. This issue is more crucial in the IR tasks taking inputs with limited details, e.g., SR and PET denoising.

Mutual stacked autoencoder for unsupervised fault detection under complex multi-residual correlations

Due to the increasing complexity of variable relationships, fault detection has garnered significant attention, as it is crucial for ensuring industrial safety and engineering reliability. Traditional detection methods can be classified as twofold: global-based and local-based strategies, which respectively focus on mining macro- and micro-level information. However, our theoretical derivation and experiment results reveal that some spurious assumptions, such as local groups and their provided information are mutually independent are implicitly adhered to but are hardly satisfied in unsupervised fault detection under real industrial scenarios. Hence, this study introduces a novel mutual stacked autoencoder (M-SAE) which can be divided into three sub-networks: L-Net, R-Net, and M-Net. L-Net enriches local information learning through multiple local backbones by incorporating the unsupervised clustering algorithm. R-Net, employing a multi-scale attention mechanism, leverages complete local information for residual strength calculation and utilizes local features to capture residual information within the latent feature space. M-Net fuses the multi-scale local feature information to perform a reconstruction for each local. A multitask entropy-aided loss function is introduced to enrich local details, the global structure, and the residual associations. Finally, results on eleven datasets validate the high-performance of the proposed M-SAE and the ablation experiments demonstrate the efficacy of each component in M-SAE, confirming that this research effectively and accurately addresses multivariable industrial fault detection tasks, thereby enabling timely interventions that are crucial for maintaining operational safety in real-world scenarios.

ERAT:Eyeglasses removal with attention

Eyeglasses removal has witnessed substantial progress in recent years, the key to eyeglasses removal is to accurately detect the eyeglasses and generate visually realistic new content in the eyeglasses removal region. However, current methods either cannot remove eyeglasses cleanly due to inability to accurately detect eyeglasses or fail to synthesize visually realistic eye information in the eyeglasses removal region. In this paper, we propose a new solution for high-quality eyeglasses removal that employs the attention mechanism to focus on solving two key challenges of the eyeglasses removal task. Specifically, our proposed method divides the eyeglasses removal task into two stages. The first stage learns the eyeglasses attention map by simultaneously imposing label filtering strategy and attention mask filtering strategy on both the latent feature space and image space, which mainly solves the challenges of eyeglass detection. The second stage fuses the attention map into a novel three-path network to remove eyeglasses and synthesize visually realistic content in the eyeglasses removal region. Experiments show that our method owns superior performance than almost all existing techniques on the task of eyeglasses removal.

NeRF-Texture: Synthesizing Neural Radiance Field Textures.

Texture synthesis is a fundamental problem in computer graphics that would benefit various applications. Existing methods are effective in handling 2D image textures. In contrast, many real-world textures contain meso-structure in the 3D geometry space, such as grass, leaves, and fabrics, which cannot be effectively modeled using only 2D image textures. We propose a novel texture synthesis method with Neural Radiance Fields (NeRF) to capture and synthesize textures from given multi-view images. In the proposed NeRF texture representation, a scene with fine geometric details is disentangled into the meso-structure textures and the underlying base shape. This allows textures with meso-structure to be effectively learned as latent features situated on the base shape, which are fed into a NeRF decoder trained simultaneously to represent the rich view-dependent appearance. Using this implicit representation, we can synthesize NeRF-based textures through patch matching of latent features. However, inconsistencies between the metrics of the reconstructed content space and the latent feature space may compromise the synthesis quality. To enhance matching performance, we further regularize the distribution of latent features by incorporating a clustering constraint. In addition to generating NeRF textures over a planar domain, our method can also synthesize NeRF textures over curved surfaces, which are practically useful. Experimental results and evaluations demonstrate the effectiveness of our approach.

Delving into CLIP latent space for Video Anomaly Recognition

We tackle the complex problem of detecting and recognising anomalies in surveillance videos at the frame level, utilising only video-level supervision. We introduce the novel method AnomalyCLIP, the first to combine Vision and Language Models (VLMs), such as CLIP, with multiple instance learning for joint video anomaly detection and classification. Our approach specifically involves manipulating the latent CLIP feature space to identify the normal event subspace, which in turn allows us to effectively learn text-driven directions for abnormal events. When anomalous frames are projected onto these directions, they exhibit a large feature magnitude if they belong to a particular class. We also leverage a computationally efficient Transformer architecture to model short- and long-term temporal dependencies between frames, ultimately producing the final anomaly score and class prediction probabilities. We compare AnomalyCLIP against state-of-the-art methods considering three major anomaly detection benchmarks, i.e. ShanghaiTech, UCF-Crime, and XD-Violence, and empirically show that it outperforms baselines in recognising video anomalies. Project website and code are available at https://lucazanella.github.io/AnomalyCLIP/.

Efficient and user-friendly visualization of neural relightable images for cultural heritage applications

We introduce an innovative multiresolution framework for encoding and interactively visualizing large relightable images using a neural reflectance model derived from a state-of-the-art technique. The framework is seamlessly integrated into a scalable multi-platform framework that supports adaptive streaming and exploration of multi-layered relightable models in web settings. To enhance efficiency, we optimized the neural model, simplified decoding, and implemented a custom WebGL shader specific to the task, eliminating the need for deep-learning library integration in the code. Additionally, we introduce an efficient level-of-detail management system supporting fine-grained adaptive rendering through on-the-fly resampling in latent feature space. The resulting viewer facilitates interactive neural relighting of large images. Its modular design allows the incorporation of functionalities for Cultural Heritage analysis, such as loading and simultaneous visualization of multiple relightable layers with arbitrary rotations.

Recognition of Forward Head Posture Through 3D Human Pose Estimation With a Graph Convolutional Network: Development and Feasibility Study.

Prolonged improper posture can lead to forward head posture (FHP), causing headaches, impaired respiratory function, and fatigue. This is especially relevant in sedentary scenarios, where individuals often maintain static postures for extended periods-a significant part of daily life for many. The development of a system capable of detecting FHP is crucial, as it would not only alert users to correct their posture but also serve the broader goal of contributing to public health by preventing the progression of chronic injuries associated with this condition. However, despite significant advancements in estimating human poses from standard 2D images, most computational pose models do not include measurements of the craniovertebral angle, which involves the C7 vertebra, crucial for diagnosing FHP. Accurate diagnosis of FHP typically requires dedicated devices, such as clinical postural assessments or specialized imaging equipment, but their use is impractical for continuous, real-time monitoring in everyday settings. Therefore, developing an accessible, efficient method for regular posture assessment that can be easily integrated into daily activities, providing real-time feedback, and promoting corrective action, is necessary. The system sequentially estimates 2D and 3D human anatomical key points from a provided 2D image, using the Detectron2D and VideoPose3D algorithms, respectively. It then uses a graph convolutional network (GCN), explicitly crafted to analyze the spatial configuration and alignment of the upper body's anatomical key points in 3D space. This GCN aims to implicitly learn the intricate relationship between the estimated 3D key points and the correct posture, specifically to identify FHP. The test accuracy was 78.27% when inputs included all joints corresponding to the upper body key points. The GCN model demonstrated slightly superior balanced performance across classes with an F1-score (macro) of 77.54%, compared to the baseline feedforward neural network (FFNN) model's 75.88%. Specifically, the GCN model showed a more balanced precision and recall between the classes, suggesting its potential for better generalization in FHP detection across diverse postures. Meanwhile, the baseline FFNN model demonstrates a higher precision for FHP cases but at the cost of lower recall, indicating that while it is more accurate in confirming FHP when detected, it misses a significant number of actual FHP instances. This assertion is further substantiated by the examination of the latent feature space using t-distributed stochastic neighbor embedding, where the GCN model presented an isotropic distribution, unlike the FFNN model, which showed an anisotropic distribution. Based on 2D image input using 3D human pose estimation joint inputs, it was found that it is possible to learn FHP-related features using the proposed GCN-based network to develop a posture correction system. We conclude the paper by addressing the limitations of our current system and proposing potential avenues for future work in this area.

Source-free domain adaptation for semantic image segmentation using internal representations.

Semantic segmentation models trained on annotated data fail to generalize well when the input data distribution changes over extended time period, leading to requiring re-training to maintain performance. Classic unsupervised domain adaptation (UDA) attempts to address a similar problem when there is target domain with no annotated data points through transferring knowledge from a source domain with annotated data. We develop an online UDA algorithm for semantic segmentation of images that improves model generalization on unannotated domains in scenarios where source data access is restricted during adaptation. We perform model adaptation by minimizing the distributional distance between the source latent features and the target features in a shared embedding space. Our solution promotes a shared domain-agnostic latent feature space between the two domains, which allows for classifier generalization on the target dataset. To alleviate the need of access to source samples during adaptation, we approximate the source latent feature distribution via an appropriate surrogate distribution, in this case a Gaussian mixture model (GMM).

Physics-informed generator-encoder adversarial networks with latent space matching for stochastic differential equations

We propose a new class of physics-informed neural networks, called Physics-Informed Generator-Encoder Adversarial Networks, to effectively address the challenges posed by forward, inverse, and mixed problems in stochastic differential equations (SDEs). In these scenarios, while governing equations are known, the available data consist of only a limited set of snapshots for system parameters. Our model consists of two key components: the generator and the encoder, both updated alternately by gradient descent. In contrast to previous approaches that directly match approximated solutions with real snapshots, we employ an indirect matching operating within the lower-dimensional latent feature space. This method circumvents challenges associated with high-dimensional inputs and complex data distributions in solving SDEs. Numerical experiments indicate that, compared to existing deep learning solvers, our proposed approach not only demonstrates superior accuracy but also exhibits advantages in both computational efficiency and model complexity.

Adversarial learning-based camera pose-to-image mapping network for synthesizing new view in real indoor environments

Given a new 6DoF camera pose in an indoor environment, we investigate the challenging problem of predicting the view from that pose based on a set of reference RGBD views. Existing explicit or implicit 3D geometry construction methods are computationally expensive, while those based on learning have primarily focused on isolated views of object categories with regular geometric structures. In contrast to the traditional render-inpaint approach for new view synthesis in real indoor environments, we propose a conditional generative adversarial neural network (P2I-NET) that directly predicts the new view from the given pose. P2I-NET learns the conditional distribution of environment images to establish the correspondence between the camera pose and its view, achieved through innovative designs in its architecture and training loss function. Two auxiliary discriminator constraints are incorporated to ensure consistency between the pose of the generated image and that of the corresponding real-world image in both the latent feature space and the real-world pose space. Moreover, a deep convolutional neural network (CNN) is introduced to further reinforce this consistency in the pixel space. Extensive new view synthesis experiments and pose estimation experiments have been conducted on real indoor datasets. The results demonstrate that P2I-NET outperforms a number of strong baseline models based on NeRF. In particular, it is shown that P2I-NET is 40 to 100 times faster than these competitor techniques while producing images of similar quality. The results also reveal that the discriminator in P2I-NET achieves comparable positioning performance to other positioning methods, despite having a smaller model size. Furthermore, we contribute a new publicly available indoor environment dataset containing 22 high resolution RGBD videos where each frame also has accurate camera pose parameters.

Heterogeneous Prototype Learning From Contaminated Faces Across Domains via Disentangling Latent Factors.

This article studies an emerging practical problem called heterogeneous prototype learning (HPL). Unlike the conventional heterogeneous face synthesis (HFS) problem that focuses on precisely translating a face image from a source domain to another target one without removing facial variations, HPL aims at learning the variation-free prototype of an image in the target domain while preserving the identity characteristics. HPL is a compounded problem involving two cross-coupled subproblems, that is, domain transfer and prototype learning (PL), thus making most of the existing HFS methods that simply transfer the domain style of images unsuitable for HPL. To tackle HPL, we advocate disentangling the prototype and domain factors in their respective latent feature spaces and then replacing the source domain with the target one for generating a new heterogeneous prototype. In doing so, the two subproblems in HPL can be solved jointly in a unified manner. Based on this, we propose a disentangled HPL framework, dubbed DisHPL, which is composed of one encoder-decoder generator and two discriminators. The generator and discriminators play adversarial games such that the generator embeds contaminated images into a prototype feature space only capturing identity information and a domain-specific feature space, while generating realistic-looking heterogeneous prototypes. Experiments on various heterogeneous datasets with diverse variations validate the superiority of DisHPL.

ScTPC: a novel semisupervised deep clustering model for scRNA-seq data.

Continuous advancements in single-cell RNA sequencing (scRNA-seq) technology have enabled researchers to further explore the study of cell heterogeneity, trajectory inference, identification of rare cell types, and neurology. Accurate scRNA-seq data clustering is crucial in single-cell sequencing data analysis. However, the high dimensionality, sparsity, and presence of "false" zero values in the data can pose challenges to clustering. Furthermore, current unsupervised clustering algorithms have not effectively leveraged prior biological knowledge, making cell clustering even more challenging. This study investigates a semisupervised clustering model called scTPC, which integrates the triplet constraint, pairwise constraint, and cross-entropy constraint based on deep learning. Specifically, the model begins by pretraining a denoising autoencoder based on a zero-inflated negative binomial distribution. Deep clustering is then performed in the learned latent feature space using triplet constraints and pairwise constraints generated from partial labeled cells. Finally, to address imbalanced cell-type datasets, a weighted cross-entropy loss is introduced to optimize the model. A series of experimental results on 10 real scRNA-seq datasets and five simulated datasets demonstrate that scTPC achieves accurate clustering with a well-designed framework. scTPC is a Python-based algorithm, and the code is available from https://github.com/LF-Yang/Code or https://zenodo.org/records/10951780.

Unsupervised face image deblurring via disentangled representation learning

Unsupervised face image deblurring aims to restore latent clear face images with explicit structure and facial details without relying on labels. Despite the recent progress, most existing approaches ignore the domain shift issue between clear and blur domains, leading to color distortion and loss of detail in deblurred images. To tackle this issue, we propose a symmetrical unsupervised framework including a decoupling stage and a reconstruction stage. In the decoupling stage, we employ disentangled representation and adversarial domain translation to establish discriminative boundaries between content-invariant domains and feature-shift domains. Besides, a clear feature encoder for clear images is introduced to obtain a latent feature space containing more detailed clarity, providing reliable input for subsequent reconstruction. In the reconstruction stage, similar modules are adopted to ensure the integrity of the mutual mapping between clear and blurred images, while domain latent supervision loss is introduced to ensure that the decoupled features and content information from both stages belong to the same domain. Extensive experimental results on popular benchmarks validate the effectiveness of our method by surpassing existing works.

Unsupervised model adaptation for source-free segmentation of medical images

The recent prevalence of deep neural networks has led semantic segmentation networks to achieve human-level performance in the medical field, provided they are given sufficient training data. However, these networks often fail to generalize when tasked with creating semantic maps for out-of-distribution images, necessitating re-training on new distributions. This labor-intensive process requires expert knowledge for generating training labels. In the medical field, distribution shifts can naturally occur due to the choice of imaging devices, such as MRI or CT scanners. To mitigate the need for labeling images in a target domain after successful model training in a fully annotated source domain with a different data distribution, unsupervised domain adaptation (UDA) can be employed. Most UDA approaches ensure target generalization by generating a shared source/target latent feature space, allowing a source-trained classifier to maintain performance in the target domain. However, such approaches necessitate joint source and target data access, potentially leading to privacy leaks with respect to patient information. We propose a UDA algorithm for medical image segmentation that does not require access to source data during adaptation, thereby preserving patient data privacy. Our method relies on approximating the source latent features at the time of adaptation and creates a joint source/target embedding space by minimizing a distributional distance metric based on optimal transport. We demonstrate that our approach is competitive with recent UDA medical segmentation works, even with the added requirement of privacy. 11Early partial results of this work has been presented in 2022 British Machine Vision Conference (Stan and Rostami, 2022a).

Dual-space disentangled-multimodal network (DDM-net) for glioma diagnosis and prognosis with incomplete pathology and genomic data

Objective. Effective fusion of histology slides and molecular profiles from genomic data has shown great potential in the diagnosis and prognosis of gliomas. However, it remains challenging to explicitly utilize the consistent-complementary information among different modalities and create comprehensive representations of patients. Additionally, existing researches mainly focus on complete multi-modality data and usually fail to construct robust models for incomplete samples. Approach. In this paper, we propose a dual-space disentangled-multimodal network (DDM-net) for glioma diagnosis and prognosis. DDM-net disentangles the latent features generated by two separate variational autoencoders (VAEs) into common and specific components through a dual-space disentangled approach, facilitating the construction of comprehensive representations of patients. More importantly, DDM-net imputes the unavailable modality in the latent feature space, making it robust to incomplete samples. Main results. We evaluated our approach on the TCGA-GBMLGG dataset for glioma grading and survival analysis tasks. Experimental results demonstrate that the proposed method achieves superior performance compared to state-of-the-art methods, with a competitive AUC of 0.952 and a C-index of 0.768. Significance. The proposed model may help the clinical understanding of gliomas and can serve as an effective fusion model with multimodal data. Additionally, it is capable of handling incomplete samples, making it less constrained by clinical limitations.

ExplainLFS: Explaining neural architectures for similarity learning from local perturbations in the latent feature space

Despite the increasing development in recent years of explainability techniques for deep neural networks, only some are dedicated to explaining the decisions made by neural networks for similarity learning. While existing approaches can explain classification models, their adaptation to generate visual similarity explanations is not trivial. Neural architectures devoted to this task learn an embedding that maps similar examples to nearby vectors and non-similar examples to distant vectors in the feature space. In this paper, we propose a post-hoc agnostic technique that explains the inference of such architectures on a pair of images. The proposed method establishes a relation between the most important features of the abstract feature space and the input feature space (pixels) of an image. For this purpose, we employ a relevance assignment and a perturbation process based on the most influential latent features in the inference. Then, a reconstruction process of the images of the pair is carried out from the perturbed embedding vectors. This process relates the latent features to the original input features. The results indicate that our method produces “continuous” and “selective” explanations. A sharp drop in the value of the function (summarized by a low value of the area under the curve) indicates its superiority over other explainability approaches when identifying features relevant to similarity learning. In addition, we demonstrate that our technique is agnostic to the specific type of similarity model, e.g., we show its applicability in two similarity learning tasks: face recognition and image retrieval.

Multidomain Adaptation With Sample and Source Distillation.

Unsupervised multidomain adaptation attracts increasing attention as it delivers richer information when tackling a target task from an unlabeled target domain by leveraging the knowledge attained from labeled source domains. However, it is the quality of training samples, not just the quantity, that influences transfer performance. In this article, we propose a multidomain adaptation method with sample and source distillation (SSD), which develops a two-step selective strategy to distill source samples and define the importance of source domains. To distill samples, the pseudo-labeled target domain is constructed to learn a series of category classifiers to identify transfer and inefficient source samples. To rank domains, the agreements of accepting a target sample as the insider of source domains are estimated by constructing a domain discriminator based on selected transfer source samples. Using the selected samples and ranked domains, transfer from source domains to the target domain is achieved by adapting multilevel distributions in a latent feature space. Furthermore, to explore more usable target information which is expected to enhance the performance across domains of source predictors, an enhancement mechanism is built by matching selected pseudo-labeled and unlabeled target samples. The degrees of acceptance learned by the domain discriminator are finally employed as source merging weights to predict the target task. Superiority of the proposed SSD is validated on real-world visual classification tasks.

Symmetry and Graph Bi-Regularized Non-Negative Matrix Factorization for Precise Community Detection

Community is a fundamental and highly desired pattern in a Large-scale Undirected Network (LUN). Community detection is a vital issue when LUN representation learning is performed. Owing to its good scalability and interpretability, a Symmetric and Non-negative Matrix Factorization model is frequently utilized to tackle this issue. It adopts a unique Latent Factor (LF) matrix for precisely representing LUN’s symmetry, which, unfortunately, leads to a reduced LF space that decreases its representation learning ability to a target LUN. Motivated by this discovery, this study proposes a Symmetry and Graph Bi-regularized Non-negative Matrix Factorization (B-NMF) method that: a) leverages multiple LF matrices when representing LUN, thereby boosting the representation learning ability; b) constructs a symmetry regularization term that implies the equality constraint among its multiple LF matrices, thereby illustrating LUN’s intrinsic symmetry; and c) incorporates graph regularization into its learning objective, thereby illustrating LUN’s local geometry. A theoretical proof is given to theoretically validate B-NMF’s convergence ability. The regularization hyperparameters are selected by validating model modularity, thereby guaranteeing B-NMF’s practicability in addressing real application issues. Extensive experimental results on ten LUNs from real applications demonstrate that the proposed B-NMF-based community detector significantly outperforms several baseline and state-of-the-art models in achieving highly-accurate community detection results. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Note to Practitioners</i> —LUNs are very-commonly seen in real applications like a social network system. Communities in LUNs are vital for various knowledge discovery-related applications. For accurately detecting them, a detector should guarantee its high representation learning ability to a target LUN. To do so, this paper presents a B-NMF model that is able to perform precise representation learning to LUNs, thereby achieving accurate community detection results. In comparison with conventional Symmetric and Non-negative Matrix Factorization-based community detectors, a B-NMF-based community detector enjoys its enlarged latent feature space, which ensures its higher representation ability to a target LUN. It depends on two regularization hyperparameters, which can be selected by performing grid-search on the target LUN via its modularity evaluation. This paper gives the empirical values of B-NMF’s regularization hyperparameters based on the parametersensitivity tests on the involved experimental datasets. The proposed B-NMF model is shown to be highly suitable for addressing community detection and clustering tasks on LUNs from real applications.

Contrastive domain generalization convolution neural network correcting the drift of gas sensors

The drift problems of gas sensors caused by aging ingredients and unexpected environmental impacts, which can severely decrease the performance and service life of electronic noses, have been widely discussed. Many scholars believe that the essential reason for sensor drift can be attributed to the different data distributions in the latent feature space. However, traditional drift compensation methods like OSC, GasNet, SniffMultinose, and SniffConv are costly and complex. This paper has introduced an algorithm called contrastive domain generalization convolution neural network (CDCNN) to resolve this problem. For the first time, domain generalization and contrastive learning were used as the drift compensation methods for the gas sensor. A novel data augmentation was proposed to enrich the datasets. Feature generation module is used to simulate the drift of gas sensors. Contrastive learning adapts to the unseen areas in the latent feature space. The artificially generated features and the original features are drawn closer in the feature space to improve the algorithm's generalization ability. Experiments on long-term drift data show that CDCNN achieves high accuracy (0.7230). The experimental results show that the CDCNN algorithm is more suitable for practical applications due to less resource consumption and looser constraints.

A Learnable Discrete-Prior Fusion Autoencoder with Contrastive Learning for Tabular Data Synthesis

The actual collection of tabular data for sharing involves confidentiality and privacy constraints, leaving the potential risks of machine learning for interventional data analysis unsafely averted. Synthetic data has emerged recently as a privacy-protecting solution to address this challenge. However, existing approaches regard discrete and continuous modal features as separate entities, thus falling short in properly capturing their inherent correlations. In this paper, we propose a novel contrastive learning guided Gaussian Transformer autoencoder, termed GTCoder, to synthesize photo-realistic multimodal tabular data for scientific research. Our approach introduces a transformer-based fusion module that seamlessly integrates multimodal features, permitting for mining more informative latent representations. The attention within the fusion module directs the integrated output features to focus on critical components that facilitate the task of generating latent embeddings. Moreover, we formulate a contrastive learning strategy to implicitly constrain the embeddings from discrete features in the latent feature space by encouraging the similar discrete feature distributions closer while pushing the dissimilar further away, in order to better enhance the representation of the latent embedding. Experimental results indicate that GTCoder is effective to generate photo-realistic synthetic data, with interactive interpretation of latent embedding, and performs favorably against some baselines on most real-world and simulated datasets.