Diffusion-TS is a novel diffusion-based framework for generating high-quality multivariate time series samples. It uses an encoder-decoder transformer with disentangled temporal representations to capture the semantic meaning of time series while transformers mine detailed sequential information from the noisy model input. Unlike existing diffusion-based approaches, Diffusion-TS is trained to directly reconstruct the sample rather than the noise in each diffusion step, combining a Fourier-based loss term. This approach enables Diffusion-TS to generate time series that are both interpretable and realistic. It can be easily extended to conditional generation tasks such as forecasting and imputation without any model changes. The framework also demonstrates strong performance in various realistic time series analyses. Diffusion-TS achieves state-of-the-art results through both qualitative and quantitative experiments. The model uses a Fourier-based training objective and a deep decomposition architecture to learn meaningful temporal properties from the data. For conditional generation, an instance-aware guidance strategy is adopted to enable Diffusion-TS to adapt different controllable generative tasks. The model's ability to generate realistic time series with high diversity and novelty is validated through experiments. Diffusion-TS is effective in handling complex time series with long-term dependencies and seasonal patterns. The model's interpretable decomposition architecture allows for the capture of complex periodic dependencies and provides explainable disentangled representations. The Fourier-based loss term helps in reconstructing the samples rather than the noises in each diffusion step, leading to more accurate generation of the time series. The model's conditional generation method, called reconstruction-based sampling, makes it versatile for various conditional applications. The framework is evaluated on multiple real-world and simulated datasets, showing superior performance in terms of distribution similarity, temporal and feature dependency, and predictive usefulness. Diffusion-TS outperforms existing methods in unconditional and conditional time series generation tasks. The model is also effective in handling irregular settings and cold starts, maintaining performance even with limited data. The results demonstrate that Diffusion-TS is a robust and interpretable method for time series generation.Diffusion-TS is a novel diffusion-based framework for generating high-quality multivariate time series samples. It uses an encoder-decoder transformer with disentangled temporal representations to capture the semantic meaning of time series while transformers mine detailed sequential information from the noisy model input. Unlike existing diffusion-based approaches, Diffusion-TS is trained to directly reconstruct the sample rather than the noise in each diffusion step, combining a Fourier-based loss term. This approach enables Diffusion-TS to generate time series that are both interpretable and realistic. It can be easily extended to conditional generation tasks such as forecasting and imputation without any model changes. The framework also demonstrates strong performance in various realistic time series analyses. Diffusion-TS achieves state-of-the-art results through both qualitative and quantitative experiments. The model uses a Fourier-based training objective and a deep decomposition architecture to learn meaningful temporal properties from the data. For conditional generation, an instance-aware guidance strategy is adopted to enable Diffusion-TS to adapt different controllable generative tasks. The model's ability to generate realistic time series with high diversity and novelty is validated through experiments. Diffusion-TS is effective in handling complex time series with long-term dependencies and seasonal patterns. The model's interpretable decomposition architecture allows for the capture of complex periodic dependencies and provides explainable disentangled representations. The Fourier-based loss term helps in reconstructing the samples rather than the noises in each diffusion step, leading to more accurate generation of the time series. The model's conditional generation method, called reconstruction-based sampling, makes it versatile for various conditional applications. The framework is evaluated on multiple real-world and simulated datasets, showing superior performance in terms of distribution similarity, temporal and feature dependency, and predictive usefulness. Diffusion-TS outperforms existing methods in unconditional and conditional time series generation tasks. The model is also effective in handling irregular settings and cold starts, maintaining performance even with limited data. The results demonstrate that Diffusion-TS is a robust and interpretable method for time series generation.