Generative Model Research Articles

Inferring control objectives in a virtual balancing task in humans and monkeys.

Natural behaviors have redundancy, which implies that humans and animals can achieve their goals with different strategies. Given only observations of behavior, is it possible to infer the control objective that the subject is employing? This challenge is particularly acute in animal behavior because we cannot ask or instruct the subject to use a particular strategy. This study presents a three-pronged approach to infer an animal's control objective from behavior. First, both humans and monkeys performed a virtual balancing task for which different control strategies could be utilized. Under matched experimental conditions, corresponding behaviors were observed in humans and monkeys. Second, a generative model was developed that represented two main control objectives to achieve the task goal. Model simulations were used to identify aspects of behavior that could distinguish which control objective was being used. Third, these behavioral signatures allowed us to infer the control objective used by human subjects who had been instructed to use one control objective or the other. Based on this validation, we could then infer objectives from animal subjects. Being able to positively identify a subject's control objective from observed behavior can provide a powerful tool to neurophysiologists as they seek the neural mechanisms of sensorimotor coordination.

StoryForge: AI-Powered Narrative with Dynamic Imagery and Voice

StoryForge is an AI-powered narrative generation system that utilizes a combination of OpenAI’s GPT-3.5 Turbo model [1], Hugging Face’s Diffusion-based text-to-image gen- erative model Stable Diffusion XL (SDXL) [4], and Hugging Face’s Text-to-Speech Bark model [10] to create dynamic and engaging stories tailored to user-specified constraints. Users can input parameters such as genre, number of characters, character details, key events, mood, tone, point of view, and reader age to guide the narrative generation process. The system leverages the power of GPT-3.5 Turbo to generate the story text [1], SDXL to generate images for key events [4], and Bark to synthesize speech for the narrative [10], providing a comprehensive and immersive storytelling experience. A Streamlit-based user interface enables seamless interaction with the system, allowing users to input constraints, view generated text and images, and listen to the synthesized audio. It transforms user inputs into a multisen- sory narrative experience, marking a significant stride in the intersection of artificial intelligence and imaginative storytelling. StoryForge demonstrates the potential of AI to revolutionize storytelling, empowering users to create personalized and cap- tivating narratives with ease and embark on a journey where words, images, and voices converge to create truly one-of-a-kind tales. Index Terms—StoryForge, GPT-3.5 Turbo, Bark, Stable Dif- fusion XL, Streamlit, Hugging Face, Generative Model, Story- telling, Multisensory

FLAIR MRI sequence synthesis using squeeze attention generative model for reliable brain tumor segmentation

Manual segmentation of brain tumors using structural magnetic resonance imaging (MRI) is an arduous and time-consuming task. Therefore, automatic and robust segmentation will considerably influence neuro-oncological clinical trials by reducing excessive manual annotation time. Herein, we propose a deep learning model that automatically segments brain tumors even in cases of missing MRI sequences, which are common in practical clinical settings. To address this issue, we enhance a generative adversarial network (GAN) by incorporating a squeeze-and-excitation (SE) attention module into its generator and a PatchGAN into its discriminator. The SE module recalibrates channel responses by explicitly modeling interdependencies, enabling the network to focus on critical regions such as tumor areas. Our proposed generative model is optimized using a combination of adversarial, structural similarity, and mean absolute error losses to synthesize missing MRI sequences more effectively. This enhancement allows our model to synthesize the missing MRI sequence (fluid attenuated inversion recovery [FLAIR]) by leveraging information from other available sequences (T1-weighted, T2-weighted, or contrast-enhanced T1-weighted [T1ce]). For the segmentation task, we employ an optimized nnU-Net model, which is trained using existing sequences and evaluated using both available and synthesized sequences (including missing ones), mimicking real-world scenarios where often only limited MRI sequences are available or usable. Our findings reveal a notable enhancement in brain tumor segmentation, as indicated by a significant increase in overall the Dice similarity coefficient (DSC) from 0.688% (when FLAIR is missing) to 0.873% (when using synthesized FLAIR derived from T2). This improvement brings the segmentation performance closer to what was achieved when real FLAIR was available, where the DSC reaches 0.901%. Moreover, our synthesizing model was also tested on two additional datasets: the BraTS 2020 validation set and BraTS Africa 2023 training set, which produces results comparable to those of BraTS 2021, thereby proving its robustness and generalizability. In addition, the resulting tumor segmentations are subsequently employed to assess the response to treatment in cases where all sequences were available and when synthesis was employed, according to response assessment in neuro-oncology criteria.

Portfolio Generator using Generative AI in Web Development

The rise of generative intelligence (AI) has led to significant transformations in various industries, notably web development. This paper introduces a unique strategy for creating personalized portfolios for web developers through the utilization of generative AI techniques. By exploiting generative AI models, developers can automate the process of generating distinct portfolio designs customized to individual preferences, resulting in saving time and effort. This paper delves into the methodology, execution, and potential advantages of utilizing generative AI in portfolio generation for web development.

DFUSNN: zero-shot dual-domain fusion unsupervised neural network for parallel MRI reconstruction

Objective. Recently, deep learning models have been used to reconstruct parallel magnetic resonance (MR) images from undersampled k-space data. However, most existing approaches depend on large databases of fully sampled MR data for training, which can be challenging or sometimes infeasible to acquire in certain scenarios. The goal is to develop an effective alternative for improved reconstruction quality that does not rely on external training datasets. Approach. We introduce a novel zero-shot dual-domain fusion unsupervised neural network (DFUSNN) for parallel MR imaging reconstruction without any external training datasets. We employ the Noise2Noise (N2N) network for the reconstruction in the k-space domain, integrate phase and coil sensitivity smoothness priors into the k-space N2N network, and use an early stopping criterion to prevent overfitting. Additionally, we propose a dual-domain fusion method based on Bayesian optimization to enhance reconstruction quality efficiently. Results. Simulation experiments conducted on three datasets with different undersampling patterns showed that the DFUSNN outperforms all other competing unsupervised methods and the one-shot Hankel-k-space generative model (HKGM). The DFUSNN also achieves comparable results to the supervised Deep-SLR method. Significance. The novel DFUSNN model offers a viable solution for reconstructing high-quality MR images without the need for external training datasets, thereby overcoming a major hurdle in scenarios where acquiring fully sampled MR data is difficult.

Discrete latent embedding of single-cell chromatin accessibility sequencing data for uncovering cell heterogeneity.

Single-cell epigenomic data has been growing continuously at an unprecedented pace, but their characteristics such as high dimensionality and sparsity pose substantial challenges to downstream analysis. Although deep learning models-especially variational autoencoders-have been widely used to capture low-dimensional feature embeddings, the prevalent Gaussian assumption somewhat disagrees with real data, and these models tend to struggle to incorporate reference information from abundant cell atlases. Here we propose CASTLE, a deep generative model based on the vector-quantized variational autoencoder framework to extract discrete latent embeddings that interpretably characterize single-cell chromatin accessibility sequencing data. We validate the performance and robustness of CASTLE for accurate cell-type identification and reasonable visualization compared with state-of-the-art methods. We demonstrate the advantages of CASTLE for effective incorporation of existing massive reference datasets in a weakly supervised or supervised manner. We further demonstrate CASTLE's capacity for intuitively distilling cell-type-specific feature spectra that unveil cell heterogeneity and biological implications quantitatively.

Multi-scale tissue fluorescence mapping with fiber optic ultraviolet excitation and generative modeling

Cellular imaging of thick samples requires physical sectioning or laser scanning microscopy, which can be restrictive, involved, and generally incompatible with high-throughput requirements. We developed fiber optic microscopy with ultraviolet (UV) surface excitation (FUSE), a portable and quantitative fluorescence imaging platform for thick tissue that enabled quick sub-cellular imaging without thin sections. We substantially advanced prior UV excitation approaches with illumination engineering and computational methods. Optical fibers delivered <300nm light with directional control, enabling unprecedented 50× widefield imaging on thick tissue with sub-nuclear clarity, and 3D topography of surface microstructure. Probabilistic modeling of high-magnification images using our normalizing flow architecture FUSE-Flow (made freely available as open-source software) enhanced low-magnification imaging with measurable localized uncertainty via variational inference. Comprehensive validation comprised multi-scale fluorescence histology compared with standard H&E histology, and quantitative analyses of senescence, antibiotic toxicity, and nuclear DNA content in tissue models via efficient sampling of thick slices from entire murine organs up to 0.4×8×12mm and 1.3 million cells per surface. This technology addresses long-standing laboratory gaps in high-throughput studies for rapid cellular insights.

ECG-image-kit: a synthetic image generation toolbox to facilitate deep learning-based electrocardiogram digitization

Abstract Objective: Cardiovascular diseases are a major cause of mortality globally, and electrocardiograms (ECGs) are crucial for diagnosing them. Traditionally, ECGs are stored in printed formats. However, these printouts, even when scanned, are incompatible with advanced ECG diagnosis software that require time-series data. Digitizing ECG images is vital for training machine learning models in ECG diagnosis, leveraging the extensive global archives collected over decades. Deep learning models for image processing are promising in this regard, although the lack of clinical ECG archives with reference time-series data is challenging. Data augmentation techniques using realistic generative data models provide a solution.

Approach: We introduce ECG-Image-Kit, an open-source toolbox for generating synthetic multi-lead ECG images with realistic artifacts from time-series data, aimed at automating the conversion of scanned ECG images to ECG data points. The tool synthesizes ECG images from real time-series data, applying distortions like text artifacts, wrinkles, and creases on a standard ECG paper background.

Main results: As a case study, we used ECG-Image-Kit to create a dataset of 21,801 ECG images from the PhysioNet QT database. We developed and trained a combination of a traditional computer vision and deep neural network model on this dataset to convert synthetic images into time-series data for evaluation. We assessed digitization quality by calculating the signal-to-noise ratio (SNR) and compared clinical parameters like QRS width, RR, and QT intervals recovered from this pipeline, with the ground truth extracted from ECG time-series. The results show that this deep learning pipeline accurately digitizes paper ECGs, maintaining clinical parameters, and highlights a generative approach to digitization.

Significance: The toolbox has broad applications, including model development for ECG image digitization and classification. The toolbox currently supports data augmentation for the 2024 PhysioNet Challenge, focusing on digitizing and classifying paper ECG images.

Two-stage fine-tuning with ChatGPT data augmentation for learning class-imbalanced data

Classification of long-tailed distributed data is a challenging problem, which suffers from serious class imbalance and hence poor performance on tail classes, which have only a few samples. Owing to this paucity of samples, learning on the tail classes is especially challenging for fine-tuning when transferring a pretrained model to a downstream task. In this work, we present a simple modification of standard fine-tuning to cope with these challenges. Specifically, we propose a two-stage fine-tuning. In Stage 1, we fine-tune the final layer of the pretrained model with class-balanced augmented data, generated using ChatGPT. As a large generative language model, ChatGPT is capable of generating novel and contextually similar responses to a given prompt, which makes it an excellent candidate for data augmentation. In Stage 2, we perform the standard fine-tuning. Our modification has several benefits: (1) it leverages pretrained representations by only fine-tuning a small portion of the model parameters while keeping the rest untouched; (2) it allows the model to learn an initial representation of the specific task; and importantly (3) it protects the learning of tail classes from being at a disadvantage during the model updating. We conduct extensive experiments on synthetic datasets of both two-class and multi-class tasks of text classification as well as a real-world application to ADME (i.e., absorption, distribution, metabolism, and excretion) semantic drug labeling. The experimental results show that the proposed two-stage fine-tuning outperforms vanilla fine-tuning and state-of-the-art methods on the above datasets.

DesignFusion: Integrating Generative Models for Conceptual Design Enrichment

Abstract Conceptual design is a pivotal phase of product design and development, encompassing user requirement exploration and informed solution generation. Recent generative models with their powerful content generation capabilities have been applied to conceptual design to support designers' ideation. However, the lack of transparency in their generation process and the shallow nature of their generated solutions constrain their performance in complex conceptual design tasks. In this study, we first introduce a conceptual design generation approach that combines generative models with classic design theory. This approach decomposes the conceptual design task based on design process and design attributes, and uses the Who, What, Where, When, Why, How (5W1H) method, Function-Behavior-Structure (FBS) model, and Kansei Engineering to guide generative models to generate conceptual design solutions through multi-step reasoning. Then we present an interactive system using a mind-map layout to visualize multi-step reasoning, called DesignFusion. This empowers designers to track the generation process and control inputs/outputs at each reasoning step. Two user studies show that our approach significantly enhances the quality of generated design solutions and enriches designer experience in human-AI co-creation.

Enterprise violation risk deduction combining generative AI and event evolution graph

AbstractIn the current realms of scientific research and commercial applications, the risk inference of regulatory violations by publicly listed enterprises has attracted considerable attention. However, there are some problems in the existing research on the deduction and prediction of violation risk of listed enterprises, such as the lack of analysis of the causal logic association between violation events, the low interpretability and effectiveness of the deduction and the lack of training data. To solve these problems, we propose a framework for enterprise violation risk deduction based on generative AI and event evolution graphs. First, the generative AI technology was used to generate a new text summary of the lengthy and complex enterprise violation announcement to realize a concise overview of the violation matters. Second, by fine‐tuning the generative AI model, an event entity and causality extraction framework based on automated data augmentation are proposed, and the UIE (Unified Structure Generation for Universal Information Extraction) event entity extraction model is used to create the event entity extraction for listed enterprises ‘violations. Then, a causality extraction model CDDP‐GAT (Event Causality Extraction Based on Chinese Dictionary and Dependency Parsing of GAT) is proposed. This model aims to identify and analyse the causal links between corporate breaches, thereby deepening the understanding of the event logic. Then, the merger of similar events was realized, and the causal correlation weights between enterprise violation‐related events were evaluated. Finally, the listed enterprise's violation risk event evolution graph was constructed, and the enterprise violation risk deduction was carried out to form an expert system of financial violations. The deduction results show that the method can effectively reveal signs of enterprise violations and adverse consequences.

Deep Learning Based Over-the-Air Training of Wireless Communication Systems without Feedback

In trainable wireless communications systems, the use of deep learning for over-the-air training aims to address the discontinuity in backpropagation learning caused by the channel environment. The primary methods supporting this learning procedure either directly approximate the backpropagation gradients using techniques derived from reinforcement learning, or explicitly model the channel environment by training a generative channel model. In both cases, over-the-air training of transmitter and receiver requires a feedback channel to sound the channel environment and obtain measurements of the learning objective. The use of continuous feedback not only demands extra system resources but also makes the training process more susceptible to adversarial attacks. Conversely, opting for a feedback-free approach to train the models over the forward link, exclusively on the receiver side, could pose challenges to reliably end the training process without intermittent testing over the actual channel environment. In this article, we propose a novel method for the over-the-air training of wireless communication systems that does not require a feedback channel to train the transmitter and receiver. Random samples are transmitted through the channel environment to train a mixture density network to approximate the channel distribution on the receiver side of the network. The transmitter and receiver models are trained with the resulting channel model, and the transmitter can be deployed after training. We show that the block error rate measurements obtained with the simulated channel are suitable for monitoring as a stopping criterion during the training process. The resulting method is demonstrated to have equivalent performance to the end-to-end autoencoder training on small message sequences.

InstructPatentGPT: training patent language models to follow instructions with human feedback

AbstractIn this research, patent prosecution is conceptualized as a system of reinforcement learning from human feedback. The objective of the system is to increase the likelihood for a language model to generate patent claims that have a higher chance of being granted. To showcase the controllability of the language model, the system learns from granted patents and pre-grant applications with different rewards. The status of “granted” and “pre-grant” are perceived as labeled human feedback implicitly. In addition, specific to patent drafting, the experiments in this research demonstrate the model’s capability to learn from adjusting claim length and inclusion of limiting terms for narrowing claim scope. As proof of concept, the experiments focus on claim ones only and the training data originates from a patent dataset tailored specifically for artificial intelligence. Although the available human feedback in patent prosecution are limited and the quality of generated patent text requires improvement, the experiments following the 3-stage reinforcement learning from human feedback have demonstrated that generative language models are capable of reflecting the human feedback or intent in patent prosecution. To enhance the usability of language models, the implementation in this research utilizes modern techniques that enable execution on a single consumer-grade GPU. The demonstrated proof of concept, which reduces hardware requirements, will prove valuable in the future as more human feedback in patent prosecution become available for broader use, either within patent offices or in the public domain.

Prediction of the microstructure of cold-compacted Astaloy 85Mo with deep generative models

In powder metallurgy, iron-based powders are compacted to so-called green bodies, which undergo a subsequent sintering process. The microstructure of the green bodies affects the sintering process and the resulting properties of the produced component. Assessing the microstructure experimentally is time consuming and costly, while simulation approaches rely on simplifications that significantly affect relevant microstructural information. This work investigates the suitability of deep generative models to synthetically generate realistic micrographs of green bodies as an alternative to experiments and simulations. For that purpose experiments comparing Generative Adversarial Neural Networks and Variational Autoencoders were conducted. The two deep learning frameworks were used to train models which produce synthetic micrographs in dependence of two parameters, namely powder particle size and green body porosity. The data utilized to train and evaluate the models was acquired from compacted water-atomized Astaloy 85Mo via scanning electron microscopy. Besides visual inspection, the trained models were evaluated via quantitative metrics. Distributions of relevant microstructural parameters such as pore perimeter or minimum and maximum pore feret diameter of the synthetically generated micrographs were compared to those of experimentally acquired micrographs. The results indicate that Generative Adversarial Networks represent a promising approach to train models, which produce synthetic micrographs that capture relevant properties of green body micrographs in dependence of given parameters such as powder particle size. In contrast, the Variational Autoencoders trained for this study did not capture microstructural properties well.

Multi-plane denoising diffusion-based dimensionality expansion for 2D-to-3D reconstruction of microstructures with harmonized sampling

Acquiring reliable microstructure datasets is a pivotal step toward the systematic design of materials with the aid of integrated computational materials engineering (ICME) approaches. However, obtaining three-dimensional (3D) microstructure datasets is often challenging due to high experimental costs or technical limitations, while acquiring two-dimensional (2D) micrographs is comparatively easier. To deal with this issue, this study proposes a novel framework called ‘Micro3Diff’ for 2D-to-3D reconstruction of microstructures using diffusion-based generative models (DGMs). Specifically, this approach solely requires pre-trained DGMs for the generation of 2D samples, and dimensionality expansion (2D-to-3D) takes place only during the generation process (i.e., reverse diffusion process). The proposed framework incorporates a concept referred to as ‘multi-plane denoising diffusion’, which transforms noisy samples (i.e., latent variables) from different planes into the data structure while maintaining spatial connectivity in 3D space. Furthermore, a harmonized sampling process is developed to address possible deviations from the reverse Markov chain of DGMs during the dimensionality expansion. Combined, we demonstrate the feasibility of Micro3Diff in reconstructing 3D samples with connected slices that maintain morphologically equivalence to the original 2D images. To validate the performance of Micro3Diff, various types of microstructures (synthetic or experimentally observed) are reconstructed, and the quality of the generated samples is assessed both qualitatively and quantitatively. The successful reconstruction outcomes inspire the potential utilization of Micro3Diff in upcoming ICME applications while achieving a breakthrough in comprehending and manipulating the latent space of DGMs.

Evaluation is key: a survey on evaluation measures for synthetic time series

Synthetic data generation describes the process of learning the underlying distribution of a given real dataset in a model, which is, in turn, sampled to produce new data objects still adhering to the original distribution. This approach often finds application where circumstances limit the availability or usability of real-world datasets, for instance, in health care due to privacy concerns. While image synthesis has received much attention in the past, time series are key for many practical (e.g., industrial) applications. To date, numerous different generative models and measures to evaluate time series syntheses have been proposed. However, regarding the defining features of high-quality synthetic time series and how to quantify quality, no consensus has yet been reached among researchers. Hence, we propose a comprehensive survey on evaluation measures for time series generation to assist users in evaluating synthetic time series. For one, we provide brief descriptions or - where applicable - precise definitions. Further, we order the measures in a taxonomy and examine applicability and usage. To assist in the selection of the most appropriate measures, we provide a concise guide for fast lookup. Notably, our findings reveal a lack of a universally accepted approach for an evaluation procedure, including the selection of appropriate measures. We believe this situation hinders progress and may even erode evaluation standards to a “do as you like”-approach to synthetic data evaluation. Therefore, this survey is a preliminary step to advance the field of synthetic data evaluation.

Diffusion models as stochastic quantization in lattice field theory

In this work, we establish a direct connection between generative diffusion models (DMs) and stochastic quantization (SQ). The DM is realized by approximating the reversal of a stochastic process dictated by the Langevin equation, generating samples from a prior distribution to effectively mimic the target distribution. Using numerical simulations, we demonstrate that the DM can serve as a global sampler for generating quantum lattice field configurations in two-dimensional ϕ4 theory. We demonstrate that DMs can notably reduce autocorrelation times in the Markov chain, especially in the critical region where standard Markov Chain Monte-Carlo (MCMC) algorithms experience critical slowing down. The findings can potentially inspire further advancements in lattice field theory simulations, in particular in cases where it is expensive to generate large ensembles.

RoUIE: A Method for Constructing Knowledge Graph of Power Equipment Based on Improved Universal Information Extraction

The current state evaluation of power equipment often focuses solely on changes in electrical quantities while neglecting basic equipment information as well as textual information such as system alerts, operation records, and defect records. Constructing a device-centric knowledge graph by extracting information from multiple sources related to power equipment is a valuable approach to enhance the intelligence level of asset management. Through the collection of pertinent authentic datasets, we have established a dataset for the state evaluation of power equipment, encompassing 35 types of relationships. To better suit the characteristics of concentrated relationship representations and varying lengths in textual descriptions, we propose a generative model called RoUIE, which is a method for constructing a knowledge graph of power equipment based on improved Universal Information Extraction (UIE). This model first utilizes a pre-trained language model based on rotational position encoding as the text encoder in the fine-tuning stage. Subsequently, we innovatively leverage the Distribution Focal Loss (DFL) to replace Binary Cross-Entropy Loss (BCE) as the loss function, further enhancing the model’s extraction performance. The experimental results demonstrate that compared to the UIE model and mainstream joint extraction benchmark models, RoUIE exhibits superior performance on the dataset we constructed. On a general Chinese dataset, the proposed model also outperforms baseline models, showcasing the model’s universal applicability.

Enhancing X-ray Security Image Synthesis: Advanced Generative Models and Innovative Data Augmentation Techniques

This study addresses the field of X-ray security screening and focuses on synthesising realistic X-ray images using advanced generative models. Insufficient training data in this area pose a major challenge, which we address through innovative data augmentation techniques. We utilise the power of generative adversarial networks (GANs) and conditional GANs (cGANs), in particular the Pix2Pix and Pix2PixHD models, to investigate the generation of X-ray images from various inputs such as masks and edges. Our experiments conducted on a Korean dataset containing dangerous objects relevant to security screening show the effectiveness of these models in improving the quality and realism of image synthesis. Quantitative evaluations based on metrics such as PSNR, SSIM, LPIPS, FID, and FSIM, with scores of 19.93, 0.71, 0.12, 29.36, and 0.54, respectively, show the superiority of our strategy, especially when integrated with hybrid inputs containing both edges and masks. Overall, our results highlight the potential of advanced generative models to overcome the challenges of data scarcity in X-ray security screening and pave the way for more efficient and accurate inspection systems.

Continuous Glucose, Insulin and Lifestyle Data Augmentation in Artificial Pancreas Using Adaptive Generative and Discriminative Models.

Artificial pancreas requires data from multiple sources for accurate insulin dose estimation. These include data from continuous glucose sensors, past insulin dosage information, meal quantity and time and physical activity data. The effectiveness of closed-loop diabetes management systems might be hampered by the absence of these data caused by device error or lack of compliance by patients. In this study, we demonstrate the effect of output sequence length-driven generative and discriminative model selection in high quality data generation and augmentation. This novel generative adversarial network (GAN) based architecture automatically selects the generator and discriminator architecture based on the desired output sequence length. The proposed model is able to generate glucose, physical activity, meal information data for individual patients. The discriminative scores for Ohio T1DM (2018) dataset were 0.17 ±0.03 (Inputs: CGM, CHO, Insulin) and 0.15 ±0.02 (Inputs: CGM, CHO, Insulin, Heart Rate, Steps) and for Ohio T1D (2020) dataset was 0.16 ±0.02 (Inputs: CGM, CHO, Insulin) and 0.15 ±0.02 (Inputs: CGM, CHO, Insulin, acceleration). A mixture of generated and real data was used to test predictive scores for glucose forecasting models. The best RMSE and MARD achieved for OhioT1DM patients were 17.19 ±3.22 and 7.14 ±1.76 for PH=30min with CGM, CHO, Insulin, heartrate and steps as inputs. Similarly, the RMSE and MARD for real+synthetic data were 15.63 ±2.57 and 5.86 ±1.69 respectively. Compared to existing generative models, we demonstrate that sequence length based architecture selection leads to better synthetic data generation for multiple output sequences (CGM, CHO, Insulin) and forecasting accuracy.