**TIME WEAVER: A Conditional Time Series Generation Model** **Abstract:** TIME WEAVER is a novel diffusion-based model designed to generate realistic time series data conditioned on heterogeneous metadata, including categorical, continuous, and time-variant variables. The model addresses the challenge of generating time series that accurately reflect real-world conditions, such as electricity demand patterns based on weather and location. Traditional approaches often ignore this metadata, leading to suboptimal results. TIME WEAVER leverages a diffusion process to generate time series while incorporating metadata through a preprocessing module. The model is evaluated using a new metric, the Joint Frechet Time Series Distance (J-FTSD), which captures the specificity of the generated time series relative to the paired metadata. Experiments on various datasets, including energy, healthcare, air quality, and traffic, demonstrate that TIME WEAVER outperforms state-of-the-art GAN models by up to 27% in downstream classification tasks. **Introduction:** Generating synthetic time series data is crucial for various applications, such as stress-testing systems, creating realistic private data, and training models. Current methods often fail to incorporate rich contextual metadata, which can significantly improve the quality and realism of generated data. TIME WEAVER addresses this gap by integrating diverse metadata conditions into the generation process. The model's architecture is tailored to handle categorical, continuous, and time-variant metadata, making it more flexible and effective than existing methods. **Background and Related Works:** Generative models, particularly GANs, have been widely used for time series generation. However, they struggle with mode collapse and instability, especially when dealing with continuous conditions. Diffusion Models (DMs) offer a more stable and realistic approach, but they have not been fully explored in the context of time series data with heterogeneous metadata. Previous works have attempted to incorporate metadata into GANs, but they often suffer from limitations such as mode collapse and poor handling of continuous conditions. **Problem Formulation:** The problem formulation involves generating multivariate time series conditioned on paired metadata, which includes both categorical and continuous features. The goal is to develop a conditional generation model that can produce samples that closely match the real data distribution given the metadata conditions. **Conditional Time Series Generation using TIME WEAVER:** TIME WEAVER uses a diffusion process to generate time series while incorporating metadata through a preprocessing module. The metadata is processed separately and then combined using a self-attention layer to create a metadata embedding. This embedding is then fed into a denoiser model, which is trained to generate samples that match the conditional distribution of the time series given the metadata. **Joint Frechet Time Series Distance (J-FTSD):** J-FTSD is a new metric designed to evaluate conditional time series generation models. It calculates the Frechet distance between the real and generated joint distributions of the time series and the paired metadata. This metric effectively penalizes models that fail to reproduce metadata-specific features in the**TIME WEAVER: A Conditional Time Series Generation Model** **Abstract:** TIME WEAVER is a novel diffusion-based model designed to generate realistic time series data conditioned on heterogeneous metadata, including categorical, continuous, and time-variant variables. The model addresses the challenge of generating time series that accurately reflect real-world conditions, such as electricity demand patterns based on weather and location. Traditional approaches often ignore this metadata, leading to suboptimal results. TIME WEAVER leverages a diffusion process to generate time series while incorporating metadata through a preprocessing module. The model is evaluated using a new metric, the Joint Frechet Time Series Distance (J-FTSD), which captures the specificity of the generated time series relative to the paired metadata. Experiments on various datasets, including energy, healthcare, air quality, and traffic, demonstrate that TIME WEAVER outperforms state-of-the-art GAN models by up to 27% in downstream classification tasks. **Introduction:** Generating synthetic time series data is crucial for various applications, such as stress-testing systems, creating realistic private data, and training models. Current methods often fail to incorporate rich contextual metadata, which can significantly improve the quality and realism of generated data. TIME WEAVER addresses this gap by integrating diverse metadata conditions into the generation process. The model's architecture is tailored to handle categorical, continuous, and time-variant metadata, making it more flexible and effective than existing methods. **Background and Related Works:** Generative models, particularly GANs, have been widely used for time series generation. However, they struggle with mode collapse and instability, especially when dealing with continuous conditions. Diffusion Models (DMs) offer a more stable and realistic approach, but they have not been fully explored in the context of time series data with heterogeneous metadata. Previous works have attempted to incorporate metadata into GANs, but they often suffer from limitations such as mode collapse and poor handling of continuous conditions. **Problem Formulation:** The problem formulation involves generating multivariate time series conditioned on paired metadata, which includes both categorical and continuous features. The goal is to develop a conditional generation model that can produce samples that closely match the real data distribution given the metadata conditions. **Conditional Time Series Generation using TIME WEAVER:** TIME WEAVER uses a diffusion process to generate time series while incorporating metadata through a preprocessing module. The metadata is processed separately and then combined using a self-attention layer to create a metadata embedding. This embedding is then fed into a denoiser model, which is trained to generate samples that match the conditional distribution of the time series given the metadata. **Joint Frechet Time Series Distance (J-FTSD):** J-FTSD is a new metric designed to evaluate conditional time series generation models. It calculates the Frechet distance between the real and generated joint distributions of the time series and the paired metadata. This metric effectively penalizes models that fail to reproduce metadata-specific features in the