T2S: High-resolution Time Series Generation with Text-to-Series Diffusion Models

1 Introduction

Time series generation (TSG) enables the creation of highquality data in scenarios with limited available datasets, thus serving as a way to simulate diverse, multimodal temporal dynamics, which offers significant value in real-world applications. Generating required objects from text (Text-to-X)

∗Corresponding author

is a research area that meets human needs and has potential at the present time. Driven by the success of diffusion models, Text-to-X generation has made remarkable strides in domains such as image generation [Rombach et al., 2022; Esser et al., 2024], video generation [Girdhar et al., 2023], and speech processing [Le et al., 2024]. Specifically, diffusion models refine noisy data progressively through a learned process, ultimately producing high-fidelity outputs. While Text-to-X generation using diffusion models has been extensively explored in vision and audio data, their application to time series generation is still in its early stages.

Existing studies [Yang et al., 2024] on TSG with diffusion models can be classified into three types based on the conditioning information. i) Label-based condition. A class-conditioned diffusion method for generating synthetic EEG signals is introduced in [Sharma et al., 2023]. ii) Temporal-based condition. Sensor data synthesis using statistical adjustments is explored in [Zuo et al., 2023]. Although class- and temporal-conditioned generation is well discussed [Yuan and Qiao, 2024; Liu et al., 2024b], they are less flexible than text-based approaches [Gao et al., 2024]. iii) Text-based condition. [Fu et al., 2024] uses domainspecific metadata (e.g., location, weather) to generate synthetic energy data, while [Wang, 2024] and [Lai et al., 2025; Alcaraz and Strodthoff, 2023] apply textual conditioning to synthesize sales and clinical ECG data, respectively. However, these methods are often domain-specific, their metadata captions can not support high resolution general alignment between time series and fine-grained captions.

Despite the progress made in applying diffusion models to time series modeling, two significant challenges remain in this domain. First, the scarcity of high-resolution generalpropose text-time series caption datasets limits progress in Text-to-Time Series (T2S) generation, while existing datasets are domain-specific (e.g., healthcare [Johnson et al., 2023], economics [Cortis et al., 2017]). Second, existing TSG models [Yuan and Qiao, 2024; Desai et al., 2021] typically require separate training for datasets of different lengths within each domain, making it challenging to develop length-arbitrary T2S models. Most approaches rely on predefined sequence lengths tied to the training data, limiting their ability to generalize. Since real-world time series exhibit inherent variability in length due to factors such as data collection frequency or system-specific temporal dynamics, the need for lengthspecific training significantly hinders the scalability and practicality of these models.

To address these challenges, we introduce a new fragmentlevel dataset, TSFragment-600K, containing over 600,000 high-resolution fragment-level text-time series pairs, which serves as a foundation for exploring T2S generation. T2S, a diffusion-based model is presented that bridges the gap between natural language and time series data in a domainagnostic manner. Specifically, T2S utilizes a length-adaptive variational autoencoder to encode time series of varying lengths into consistent latent embeddings. The model then aligns textual representations with latent embeddings using flow matching and employs a diffusion transformer as the denoiser. By training T2S in an interleaved manner across diverse datasets, the model is able to generate high-quality and semantic aligned time series of arbitrary lengths during inference, overcoming the fixed-length limitations of prior approaches. All resources have been made available 1 . This work marks three key contributions:

• We systematically explore the existing T2S datasets, and introduce a novel, high-resolution fragment-level multimodal dataset for text-to-time series generation tasks.
• We propose the first domain-agnostic model for text-totime series generation, which integrates flow matching and the diffusion transformer, and is capable of generating semantically aligned time series of arbitrary lengths.
• T2S sets a new state-of-the-art performance across 13 datasets from 12 domains in time series generation, consistently outperforming both diffusion-based models and those based on large language models.

2 Definition and Dataset

2.1 ProblemDefinition and Notation

Let x ∈ RL denote a univariate time series of length L. Textual captions, represented as T, provide semantic guidance across varying levels of granularity.

Definition 1 (Point-Level Description). A point-level description Tp provides semantic annotations for individual time points within the time series x. Each point-level description T(j) p corresponds to the j-th time point, where j ∈ [1, . . . ,L], providing a fine-grained guidance for each point in time series x.

Definition 2 (Fragment-Level Description). A fragmentlevel description Tf provides semantic annotations for nonoverlapping and contiguous fragments of the time series x. Each fragment-level description T(j) f corresponds to the j-th fragment, where the length |T(j) f | is arbitrary and determined by the specific structure of time series x.

1 https://github.com/WinfredGe/T2S

Definition 3 (Instance-Level Description). An instance-level description Ti provides a global semantic annotation for the entire time series x, providing high-level guidance encompassing all time points.

We can use these captions to guide time series generation:

Definition 4 (T2S Generation). Given a text-time series dataset D = {(x(i) ,T(i) ∗ )}Ni=1 with N samples, the task of T2S generation aims to learn a generative model G that maximizes the conditional probability PG(x | T∗), where T∗ is the textual guidance provided at one of the following levels (Tp, Tf , Ti). The generated time series x must semantically align with the textual guidance T∗.

2.2 TSFragment-600K

The proposed T2S generation task leverages textual captions at three granularity levels: point, fragment, and instance. While datasets for point-level [Liu et al., 2024a] and instancelevel [Kawaguchi et al., 2025] descriptions are readily available, fragment-level descriptions remain an underexplored area. Fragment-level descriptions strike a balance between the granularity of point-level annotations, which may overlook broader temporal patterns, and the holistic nature of instance-level descriptions, which might obscure local dependencies. By encapsulating local temporal trends while preserving meaningful contextual relationships, fragment-level descriptions provide an ideal framework for evaluating the generative capabilities of T2S models.

To this end, we introduce TSFragment-600K, a novel dataset comprising over 600,000 fragment-level text-time series pairs. Each captions captures fine-grained temporal morphological characteristics, offering a rich and nuanced representation of the underlying trends. As illustrated in Figure 1, we employ GPT-4o-mini to generate high-quality natural language descriptions for each time series fragment, focusing on local trends and variations. Unlike prior approaches that rely on predefined dictionaries of time-series changes [Imani et al., 2019], our captions are expressed in natural language, enhancing their interpretability and applicability.

Specifically, we propose a novel generation pipeline to construct fragment-level captions for time series data. First, a univariate time series x is segmented into k non-overlapping fragments,with each fragment x(j) ∈ Rlj as a contiguous temporal segment for which textual captions are generated. Second, a seed-based prompting strategy is designed to capture high-quality captions of fragments. To accurately capture temporal dynamics, human experts curate high-quality GPT-4o descriptions for a subset of fragments, which serve as representative seed prompts. These prompts guide GPT4o-mini in generating concise, consistent, and semantically rich captions for all fragments. A token limit ⟨M⟩ is applied to ensure a balance between informativeness and brevity.

Finally, five candidate captions are generated for each time series sample, and their embeddings are computed using textembedding-3-small. We ensure the quality of generated textual captions by leveraging cosine similarity among their embeddings. Each caption is scored based on the average similarity of its embedding with others, and the one with the highest score is selected as the optimal text-time series pair, ensuring semantic alignment and coherence. Using this pipeline, we generate reliable fragment-level descriptions for eight classical time series datasets across diverse domains, including energy consumption, financial, exchange rates, traffic, air quality, and meteorological variables. The resulting fragment-level dataset, TSFragment-600K, comprises over 600, 000 samples with corresponding captions.

3 Methodology

In this section, we introduce the T2S architecture, as depicted in Figure 2. The architecture consists of two key components:

• T2S Diffusion Transformer (T2S-DiT): T2S-DiT facilitates high-resolution alignment between captions and the temporal latent space. It employs flowmatching [Liu et al., 2022] as the diffusion backbone, the diffusion transformer module [Peebles and Xie, 2023] as the denoiser. Within this denoiser, textual information is integrated with the input features through adaptive layer normalization
• Pretrained Length-Adaptive Variational Autoencoder (LA-VAE): LA-VAE encodes variable-length time series into a unified latent feature space and decodes them back to their original temporal dimensions. An interleaved training strategy is adopted to enable effective handling of varying input lengths during training.

We first employ LA-VAE to map time series of varying lengths into the latent space. The T2S-DiT module then denoises this latent space conditioned on the caption, aligning the textual and temporal features. During inference, a noise sequence of arbitrary length, encoded by LA-VAE, generates aligned time series based on the given caption.

3.1 T2S DiffusionModel

FlowMatching Framework. Inspired by [Esser et al., 2024; Polyak et al., 2024] and experimental comparisons with DDPM, we adopt the flow matching framework [Lipman et al., 2022], specifically adopting the rectified flow approach [Liu et al., 2022] using optimal transport paths. This framework offers superior generation quality and a more stable inference process compared to DDPM [Ho et al., 2020], with reduced training costs.

Flow matching consists of forward and reverse processes. During training, given a time series sample in latent space z1, a noisy sample z0 ∼ N(0, I), and a time step t ∈ [0, 1], the forward path zt is defined as:

where zt evolves along an optimal transport path defined by:

and the ground truth velocity of this transition is described by: vt = dzt dt = z1 - z0. In the reverse process, the denoiser model aims to predict the velocity uθ(zt, t, C) , where C represents the text prompt embedding generated by T∗ in Definition 4. The model minimizes the mean squared error (MSE) between the ground truth and predicted velocities:

During sampling, pure noise z0 ∼ N(0, 1) is iteratively denoised to obtain ˆz1 by solving the ordinary differential equation (ODE) with the well-trained denoiser uθ (zt, t,C). T2S adopts a classifier-free guidance framework, which does not rely on explicit class labels for conditioning. Unlike traditional class-based conditional generation methods, which use fixed condition categories to guide the generation process, classifier-free guidance provides a more flexible and effective way to balance unconditional and conditional generation. During training, the method randomly sets the condition to zero by a random ratio. During inference, the model first performs conditional generation, followed by unconditional generation, and then combines the results using a guidance scale δ. The formula is:

where uθ(zt, t,C) is the noise estimater with input zt and condition C and uθ(zt, t) is the conditioned noise estimater . Diffusion Transformer. Building on the superior performance of the DiT in computer vision [Esser et al., 2024] and recent advances in time series analysis [Chen et al., 2024], we developed a patchified DiT denoiser that leverages DiT's fine-grained visual feature extraction and its ability to capture subtle latent temporal patterns, aligning these patterns with corresponding textual descriptions through the 2D latent representations encoded via LA-VAE. The latent space representation of time series can be conceptualized as single-channel, two-dimensional grayscale images zt. We first patchify zt into a sequence of tokens. Then the input tokens are obtained by summing the sequence of tokens with two-dimensional positional embeddings. ct and zt represent the conditioning associated with time step t and input tokens. To achieve alignment between the conditioning ct and the input tokens zt, we employ an adaptive layer normalization (AdaLN), a form of modulated layer normalization [Huang and Belongie, 2017]:

where scaling parameter αt,1, αt,2, γt,1, γt,2 and shifting parameter βt,1, βt,2 are chunked outputs, dynamically adjusted based on the textual information, as shown in Equation 6.

The overall procedure can be formulated as follows:

where µ(i) t and σ(i) t denote the mean and variance of the input zt for the i-th layer, respectively, and MH-A represents the multi-head attention. The entire process integrates adaptive layer norm into the transformer architecture, enabling it to dynamically adapt to textual conditioning. By leveraging this mechanism, the diffusion transformer aligns the input time series tokens with the contextual information, improving the generative performance of the model.

3.2 Length-adaptive VAE

T2S generation should support arbitrary-length generation to meet real-world application demands. For example, website user activity analysis and medical monitoring. However, previous works [Lai et al., 2025] typically assume a fixed length, limiting their applicability. To address this limitation, we propose a pretrained LA-VAE, enabling the modeling and generation of time series with arbitrary lengths. In this study, we mixed data with lengths of 24, 48, and 96, then trained them in a unified framework. During sampling, arbitrarylength data can be generated within a specified range. Given an input time series x, the LA-VAE encoder transforms x into a latent representation h. This latent representation h is subsequently upsampled to a fixed-size latent embedding z = Up-Samp(ht), which serves as the input to the diffusion model. Through the diffusion process, a refined latent embedding ˆzt is generated. A downsampling operation is then applied to ˆzt, yielding the latent vector ˆh = Down-Samp(ˆzt). Finally, the VAE decoder reconstructs ˆh into a time series with the original length. This process enables the handling of variable-length time series within a unified framework.

Consistency loss. Linear interpolation during upsampling and downsampling introduces blurriness and artifacts, as it fails to capture nonlinear features and signal curvature. To mitigate this, we introduce a latent space consistency loss termMSE(h, ˆh) to enhance reconstruction quality:

where the first term enforces fidelity to the original time series, and the second ensures consistency in the latent space.

Algorithm 1 Interleaved Training for Mixed Datasets

Require: Datasets {D1, . . . ,Dk}, LA-VAE ϕ(·), T2S-DiT gθ(·), batch size, number of iterations N. 1: n0 = 0, ni ← |Di|, i = 1, . . . , k 2: DatasetSampling(): 3: n← ∑ki=1 ni 4: j ∼ Uniform(1, n) 5: Find m ∈ Z+ such that j ∈ (∑mi=1 ni, ∑m+1 i=1 ni ] 6: Return Dm[j - ∑m-1 i=1 ni] 7: for iter = 1 to N do 8: D← {}, loss = 0 9: for i = 1 to batch size do 10: D ← D ∪ DatasetSampling() 11: end for 12: for length i = 1 to k do 13: batchi = {s ∈ D | len(s) = i} 14: ˆS ← gθ (ϕ(batchi)) 15: loss = loss + L in equation 10 16: end for 17: Update ϕ 18: end for

3.3 Interleaved Training

Traditional sequential training paradigm often lead to catastrophic forgetting. To effectively train models on datasets of varying lengths within a unified framework, we propose a novel interleaved training paradigm, outlined in Algorithm 1.

Assume a domain contains d datasets with different lengths, denoted as {l1, l2, . . . , ld}. During training, we shuffle all samples across these datasets and randomly sample batch size samples for each batch. Within each iteration, training is interleaved across these batches, which improves the model's generalization ability.

4 Experiments

We conduct an extensive evaluation across 13 datasets spanning 12 domains to assess the performance of the T2S, aiming to address the following key research questions:

• RQ1: How does T2S compare in performance to existing state-of-the-art methods given fragment-level captions?
• RQ2: How does T2S compare in performance to existing methods given point-level and instance-level captions?
• RQ3: How do the different components within the T2S affect its overall generation performance?
• RQ4: How sensitive is the T2S's performance to key hyperparameters, and does it require additional fine-tuning?
• RQ5: How effective is T2S when trained on limited data?

4.1 Experimental Settings

Datasets. We evaluated our model on three distinct datasets: Point-Level, Fragment-Level, and Instance-Level.

• Point-Level Dataset: The Time-MMD dataset [Liu et al., 2024a] links individual time series points with corresponding textual news, consisting of 23,618 data points across six domains, including Climate, Economy, and Social Goods. We adapted the dataset by concatenating each time series point with its associated text.
• Fragment-Level Dataset: TSFragment-600K pairs time series data with captions across seven domain, including Electricity, ETT, Exchange, and Traffic [Wu et al., 2021]. It consists of over 600,000 samples.
• Instance-Level Dataset: SUSHI, a simulated dataset [Kawaguchi et al., 2025], comprises 2,800 samples generated from 15 pre-defined functions.

Evaluation Metrics. We evaluated performance using Mean Squared Error (MSE), Weighted Absolute Percentage Error (WAPE) [Shao et al., 2024], and Mean Reciprocal Rank at 10 (MRR@10)[Craswell, 2009].

• Weighted Absolute Percentage Error (WAPE):

where Ω and ˆΩ represent truth space and generative space, while yi and ˆ yi denote the corresponding i-th sample.

• Mean Reciprocal Rank at 10 (MRR@10):

where ranki = argmin(n | cos(ˆ yi,n, yi) > threshold).

Baselines. To evaluate T2S, we compared its performance against state-of-the-art fully trained models and zero-shot large language models, encompassing diverse paradigms to robustly assess its effectiveness. Among fully trained models, DiffusionTS [Yuan and Qiao, 2024] uses diffusion with text and time embeddings for time-series generation and adapts to text-conditional generation by injecting text embeddings into its encoder and decoder using AdaLayerNorm. Meanwhile, TimeVAE [Desai et al., 2021] utilizes a variational autoencoder framework with caption embeddings, using dense layers with ReLU to fuse input and text for conditioning. For the zero-shot baselines, GPT-4o [OpenAI, 2023] and Llama3.1-8b [Dubey et al., 2024] are employed. To curb hallucination and ensure reliable outputs, Llama-3 is guided by domain-specific prompts, a repeat generation loop, and targeted post-processing.

4.2 Performance Comparison on Fragment-Level Descriptions (RQ1)

Table 1 presents the fragment-level performance comparison across six datasets, evaluated using WAPE, MSE, and MRR@10. T2S achieves top performance across all metrics, securing 14 out of 18 entries (77.8%) for MSE, significantly outperforming DiffusionTS (5.6%) and TimeVAE (11.1%). On the exchange rate dataset, T2S achieves an average MSE of 0.039, representing a 56.0% improvement over DiffusionTS and a 68.9% improvement over TimeVAE. These results demonstrate T2S's superior ability to align textual and temporal features across diverse datasets and fragment lengths. Moreover, T2S's interleaved training strategy enables cross-length training within each dataset, removing the need for length-specific training required by baseline models, thereby enhancing scalability and generalization.

Table 1: Performance comparison on fragment level. All results are evaluated on three different metrics WAPE, MSE, and MRR@10. For interleaved training, arbitrary lengths of {24, 48, 96} were selected, with evaluations performed separately for each length.

4.3 Performance Comparison on Point and Instance-Level Descriptions (RQ2)

Table 2 presents the performance comparison at the point and instance levels across seven datasets. These evaluations assess models' abilities to capture fine-grained temporal annotations and generate coherent global patterns. T2S outperforms all baselines across three metrics. It consistently achieves the lowest MSE values in 17 out of 18 entries, surpassing DiffusionTS and TimeVAE, which struggle to capture fine-grained temporal variations. Similarly, T2S secures the best WAPE scores in 16 out of 18 entries, demonstrating its robustness. T2S achieves the top MRR@10 score of 0.314 on the instance-level SUSHI dataset, its WAPE and MSE results show an advantage over zero-shot models, only a slight underperformance compared to DiffusionTS in certain scenarios. In summary, T2S demonstrates state-of-the-art performance by effectively balancing fine-grained precision with high-level semantic understanding.

4.4 Ablation Study (RQ3)

As shown in Figure 3 ,we conducted three ablation experiments on three datasets: Exchange Rate, Air Quality, and Traffic, evaluating nine different settings. First, replacing the flow matching backbone with DDPM resulted in consistent performance drops, with an average error increase of 311.00% across all datasets. Second, substituting the DiT denoiser with an MLP drastically degraded results, with errors rising by 877.67% on the Exchange Rate dataset. Lastly, removing text guidance severely impacted high-resolution generation, with average error increase of 495.13%, 327.10%, 205.23% across datasets with lengths {24, 48, 96}, respectively, highlighting the critical role of text in guiding generation quality. These findings clearly highlight the critical role of each component, demonstrating that the full model is essential for achieving high-quality generation.

4.5 Parameter Sensitivity (RQ4)

Recent study [Li et al., 2024] demonstrates the pivotal role of the inference stage in affecting the performance of diffusion models. Building on this, we explored the sensitivity of the flow matching diffusion model to key inference parameters: classifier-free guidance scales (CFG) and generation time steps,evaluated using MRR@10. Figure 4 shows a heatmap illustrating performance impact, with yellow regions yielding superior results and green areas reflecting suboptimal performance. Notably, the model achieves higher MRR@10 scores within the range of CFG scores between 7 and 10 and generation time steps between 20 and 50. This analysis underscores the importance of precise inference-stage parameter selection in optimizing the performance of Flow Matching models.

Table 2: Performance comparison on point and instance levels. All results are evaluated on three different metrics WAPE, MSE, and MRR@10. For the instance-level dataset, SUSHI is used for training and inference with a fixed length of 2048. For point-level training, arbitrary lengths of {24, 48, 96} were selected for interleaved training.

4.6 Data Scarcity (RQ5)

To explore the impact of dataset size on the model's performance, we evaluate the model on the Exchange Rate dataset under varying dataset scales. Specifically, we train and evaluate the model on subsets consisting of different proportions of the full dataset. As shown in figure 4, the model demonstrates consistent improvement with increasing dataset size. Notably, using only 50% of the dataset, the model achieves 93.8% of the full dataset performance. Similarly, for the 48- and 96-length generations, the model reaches 92.4% and 91.3% of the full dataset performance, respectively. These results highlight the model's strong generalization ability in data-scarce scenarios, showing its capacity to generate highquality time-series even with limited training data.

5 RelatedWork

Text-Time Series Datasets. Existing text-time series pair datasets can be categorized into three types: instance-level, point-level, and fragment-level, based on how the caption and time series are temporally aligned. At the point level, each time series point is paired with an event description, such as financial news or clinical notes [Yu et al., 2023; Cortis et al., 2017; Liu et al., 2024a]. At the instance level, TRUCE [Jhamtani and Berg-Kirkpatrick, 2021] and SUSHI [Kawaguchi et al., 2025] utilize time series features, such as upward trends and peaks, as dictionary entries to generate coherent signals. At the fragment level, several researchers have introduced datasets tailored for time series reasoning tasks [Williams et al., 2024; Chow et al., 2024]. A large-scale, fine-grained, general-purpose text-time series dataset for time series generation tasks remains in its early exploration.

Text-Time Series Generation. The general text-to-time series paradigm can be achieved through contrastive learning or generative modeling. Recently, contrastive learning has been employed to facilitate text-to-time series mapping. However, these approaches are primarily focused on retrieval tasks [Ito et al., 2024; Rizhko and Bloom, 2024] and cannot be directly applied to time series generation. In contrast, generative modeling, including variational autoencoders (VAEs) [Desai et al., 2021; Lee et al., 2023], diffusion models [Yuan and Qiao, 2024; Kong et al., 2020; Wen et al., 2023; Narasimhan et al., 2024], and large language models [OpenAI, 2023; Dubey et al., 2024], provides more versatile frameworks for generating time series conditioned on textual descriptions. Among these, conditional diffusion models [Yuan and Qiao, 2024; Cao et al., 2024; Narasimhan et al., 2024] show promise for text-to-time series generation due to their ability to model complex temporal dynamics and generate temporally coherent sequences. For instance, time series generation conditioned on healthcare metadata [Alcaraz and Strodthoff, 2023; Lai et al., 2025] and sensor metadata [Zuo et al., 2023; Fu et al., 2024] has been explored. However, these methods are often domain-specific and fail to address the more general alignment between time series and their corresponding captions, limiting their broader applicability.

6 Conclusion

We proposed TSFragment-600K, a high-resolution fragment-level multimodal dataset for text-to-time series generation tasks, and T2S, the first domain-agnostic model for general text-to-time series generation. Leveraging LAVAE and T2S-DiT, T2S generates semantically aligned time series of arbitrary lengths with high fidelity. Comprehensive validation across 12 diverse domains demonstrates T2S's superior performance, establishing a robust foundation for text-to-time series generation.