TIME WEAVER is a novel diffusion-based model for generating realistic multivariate time series conditioned on paired heterogeneous metadata. The model addresses the challenge of generating time series data that incorporates diverse metadata, including categorical, continuous, and time-varying features. Unlike existing approaches that ignore such metadata, TIME WEAVER leverages these features to significantly improve the quality of generated time series. Additionally, the paper introduces a new evaluation metric, Joint Frechet Time Series Distance (J-FTSD), which accurately captures the specificity of conditional generation and the realism of the generated time series. J-FTSD incorporates time series and metadata conditions using feature extractors trained with contrastive learning. The model outperforms state-of-the-art benchmarks, such as Generative Adversarial Networks (GANs), by up to 27% in downstream classification tasks on real-world energy, medical, air quality, and traffic datasets. The paper also highlights the importance of evaluating conditional generation models with metrics that consider paired metadata, as existing metrics fail to penalize models for their inability to reproduce metadata-specific features. TIME WEAVER's architecture includes a denoiser backbone and a preprocessing module that processes time-varying categorical and continuous metadata variables. The model is trained to iteratively reconstruct less noisy samples, ultimately generating realistic time series data. The J-FTSD metric is shown to be highly sensitive to differences between real and generated time series data, making it a reliable indicator of the quality of generated data. The experiments demonstrate that TIME WEAVER significantly outperforms baseline models in synthesizing time series data across all evaluated benchmarks. The model's ability to generate realistic time series samples that are specific to the corresponding metadata condition is validated through both quantitative and qualitative analysis. The results show that TIME WEAVER's generated data closely matches the real data distribution, with a clear correlation between lower J-FTSD scores and higher TSTR performance. The paper concludes that TIME WEAVER is a state-of-the-art model for conditional time series generation, capable of generating realistic time series data that matches the real data distribution and captures metadata-specific features.TIME WEAVER is a novel diffusion-based model for generating realistic multivariate time series conditioned on paired heterogeneous metadata. The model addresses the challenge of generating time series data that incorporates diverse metadata, including categorical, continuous, and time-varying features. Unlike existing approaches that ignore such metadata, TIME WEAVER leverages these features to significantly improve the quality of generated time series. Additionally, the paper introduces a new evaluation metric, Joint Frechet Time Series Distance (J-FTSD), which accurately captures the specificity of conditional generation and the realism of the generated time series. J-FTSD incorporates time series and metadata conditions using feature extractors trained with contrastive learning. The model outperforms state-of-the-art benchmarks, such as Generative Adversarial Networks (GANs), by up to 27% in downstream classification tasks on real-world energy, medical, air quality, and traffic datasets. The paper also highlights the importance of evaluating conditional generation models with metrics that consider paired metadata, as existing metrics fail to penalize models for their inability to reproduce metadata-specific features. TIME WEAVER's architecture includes a denoiser backbone and a preprocessing module that processes time-varying categorical and continuous metadata variables. The model is trained to iteratively reconstruct less noisy samples, ultimately generating realistic time series data. The J-FTSD metric is shown to be highly sensitive to differences between real and generated time series data, making it a reliable indicator of the quality of generated data. The experiments demonstrate that TIME WEAVER significantly outperforms baseline models in synthesizing time series data across all evaluated benchmarks. The model's ability to generate realistic time series samples that are specific to the corresponding metadata condition is validated through both quantitative and qualitative analysis. The results show that TIME WEAVER's generated data closely matches the real data distribution, with a clear correlation between lower J-FTSD scores and higher TSTR performance. The paper concludes that TIME WEAVER is a state-of-the-art model for conditional time series generation, capable of generating realistic time series data that matches the real data distribution and captures metadata-specific features.