TimeOmni-VL: Unified Models for Time Series Understanding and Generation

Tong Guan; Sheng Pan; Johan Barthelemy; Zhao Li; Yujun Cai; Cesare Alippi; Ming Jin; Shirui Pan

cs.LGFeb 19, 2026

TimeOmni-VL: Unified Models for Time Series Understanding and Generation

Authors: Tong Guan, Sheng Pan, Johan Barthelemy, Zhao Li, Yujun Cai, Cesare Alippi, Ming Jin, Shirui Pan

Organizations: Griffith University · Zhejiang University · NVIDIA · Zhejiang Lab · The University of Queensland · Universit`a della Svizzera Italiana · Politecnico di Milano

Abstract

Recent time series modeling faces a sharp divide between numerical generation and semantic understanding, with research showing that generation models often rely on superficial pattern matching, while understanding-oriented models struggle with high-fidelity numerical output. Although unified multimodal models (UMMs) have bridged this gap in vision, their potential for time series remains untapped. We propose TimeOmni-VL, the first vision-centric framework that unifies time series understanding and generation through two key innovations: (1) Fidelity-preserving bidirectional mapping between time series and images (Bi-TSI), which advances Time Series-to-Image (TS2I) and Image-to-Time Series (I2TS) conversions to ensure near-lossless transformations. (2) Understanding-guided generation. We introduce TSUMM-Suite, a novel dataset consisting of six understanding tasks rooted in time series analytics and coupled with two generation tasks. With a calibrated Chain-of-Thought, TimeOmni-VL is the first to leverage time series understanding as an explicit control signal for high-fidelity generation. Experiments confirm that this unified approach significantly improves semantic understanding and numerical precision, establishing a new frontier for multimodal time series modeling.

Figures & tables

Figure 1: Comparison of architectures for (a) time series understanding model that produces textual answers only, (b) time series generation model that outputs time series only, and (c) unified time series understanding and generation model that supports both answering queries and generating time series.

Figure 2: Overview of the TimeOmni-VL framework. The input time series is first converted into a TS-image $I$ by the (a) TS2I Converter . For understanding tasks, the understanding model directly produces CoT $R$ and the final answer. For generation tasks, the understanding model first generates CoT $R_{\mathrm{gen}}$ as a condition for the generation module to generate the target image $I_{\mathrm{tgt}}$ , which is then converted back to a time series by the (b) I2TS Converter . Detailed pipelines of the TS2I and I2TS converters are shown on the right.

Figure 3: Illustration of improvements in Bi-TSI. (a) Robust fidelity normalization enables near-lossless rendering of high-dynamic-range time series by keeping values within the valid pixel range, whereas the baseline in VisionTS++ ( Shen et al., 2025 ) can overflow this range and fail to represent the spike. (b) Encoding capacity control prevents implicit downsampling when encoding high-dimensional time series , ensuring that the resulting TS-image remains information-preserving, whereas the baseline suffers information loss.

Figure 4: Illustrative examples of the proposed TSUMM-Suite , consisting of six time series understanding tasks and two generation tasks. The generation CoT is directly derived from the understanding tasks, explicitly bridging the two task families.

Figure 5: Performance on TS-image understanding tasks.

Figure 6: Ablation on TS2I strategies. Comparison between our TS2I and the heatmap representation for forecasting (left) and imputation (right). Red arrows indicate the performance gap.

Figure 7: Visual comparison of TS-image construction. Original time series (left). Our TS2I strategy (middle), which aligns periodic cycles explicitly. Standard heatmap representation (right).

Figure 8: Ablation on the understanding model. Comparison between generation-only and understanding-guided generation for forecasting (left) and imputation (right).

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

	Forecasting	Imputation	Understanding	Reasoning
Training Set	40,000	40,000	9,409	2,339
Testbed	685	855	685	2,448

Appendix

Table 3: Detailed quantitative statistics for the four time series tasks in TSUMM-Suite across training sets and testbeds.

	Forecasting	Imputation	Understanding	Reasoning
MAX length	2,592	2,592	2,592	792
MIN length	8	8	8	10
AVG length	957	957	947	109

Appendix

Table 4: Maximum, minimum, and average time series lengths across four tasks.

	Forecasting	Imputation	Understanding	Reasoning
AVG tokens of time series $\mathbf{X}$	7,236	7,236	4,096	860
AVG tokens of context $C$	116	130	479	246
AVG total tokens	7,352	7,366	4,575	1,106

Appendix

Table 5: Average token budgets computed using the tokenizer of our base model Bagel ( Deng et al., 2025 ) .

Regime	Std-based	MAD-based	RFN (Ours)
Gaussian signal	Ideal	Ideal	Ideal
Heavy outliers	Signal washout	Ideal	Ideal
Step / flat	Ideal	Noise amplification	Ideal

Appendix

Table 6: Qualitative behavior of different normalization methods under representative challenging regimes. Ideal indicates faithful visual preservation of the underlying signal structure.

Understanding Task 1: Variable Counting
User	You are given a 896 $\times$ 896 image that encodes one or more time series variables as horizontal bands stacked from top to bottom. How many variables (time series channels) are encoded in this image?

TimeOmni-VL	<think> Each continuous horizontal color band corresponds to one time series channel. Counting the bands from top to bottom gives 3. </think> Answer: The image encodes 3 variables (time series channels).

Appendix

Table 15: Example of Understanding Task 1: Variable Counting.

Understanding Task 2: Variable Y-Range
User	You are given a 896 $\times$ 896 image that encodes one or more time series variables as horizontal bands stacked from top to bottom. - Variables are indexed from 1 (top band) to nvars (bottom band). - First, determine nvars by counting the number of horizontal bands in the image. - Then compute the per-variable height using integer (floor) division: image_size_per_var = max(1, image_size // max(1, nvars)) Using y=0 at the top row and inclusive pixel indices, what is the y-range [y_start, y_end] for variable 1?

TimeOmni-VL	<think> 1) Visual Analysis. Scanning the image reveals 3 distinct horizontal bands. Thus, $n_{\text{vars}}=3$ . 2) Height Calculation. Image height $H=896$ . Per-variable height $h=\lfloor 896/\max(1,3)\rfloor=298$ pixels. 3) Range Derivation. Target: Variable 1 (Topmost, internal index $i=0$ ). • $y_{\text{start}}=i\times h=0\times 298=0$ . • $y_{\text{end}}=(i+1)\times h-1=1\times 298-1=297$ . </think> Answer: Variable 1 occupies y-range: [0, 297].

Appendix

Table 16: Example of Understanding Task 2: Variable Y-Range.

Understanding Task 3: Cycle Bounding Box
User	You are given a 896 $\times$ 896 image that encodes one or more time series variables as horizontal bands stacked from top to bottom. Given metadata: $L=240$ , periodicity = 24. Each cycle corresponds to a vertical strip in the image. Questions: (1) How many pixels wide is one cycle? (2) For variable 2 and cycle 3, give the bounding box [(x1, y1), (x2, y2)] of that cycle region.

TimeOmni-VL	<think> 1) Geometry & Layout. Image size $896\times 896$ . Scanning bands $\rightarrow n_{\text{vars}}=3$ . Band height $h=\lfloor 896/3\rfloor=298$ . 2) Cycle Dimensions. Metadata: $L=240$ , Periodicity $P=24$ . Total cycles $N=\lceil L/P\rceil=\lceil 240/24\rceil=10$ . Width per cycle $w=896/10=89.6$ pixels. 3) Coordinates Calculation. Target: Variable 2 , Cycle 3 . • Y-range: $[(2-1)h,2h-1]=[298,595]$ . • X-range: $[\lfloor(3-1)w\rfloor,\lfloor 3w\rfloor-1]$ . • $x_{1}=\lfloor 2\times 89.6\rfloor=179$ . • $x_{2}=\lfloor 3\times 89.6\rfloor-1=267$ . </think> Answer: (1) Each cycle is 89.6 pixels wide. (2) The bounding box for variable 2, cycle 3 is: [(179, 298), (267, 595)].

Appendix

Table 17: Example of Understanding Task 3: Cycle Bounding Box.

Understanding Task 4: Mean Comparison
User	You are given a 896 $\times$ 896 image that encodes one or more time series variables as horizontal bands stacked from top to bottom. For variable 2, compare cycle 7 and cycle 9 based on overall brightness (0-255 pixel values). These two cycles show visually distinct brightness levels. Given metadata: $L=240$ , periodicity = 24. Which cycle has higher values? Note: Variables are indexed from 1 (top band) to nvars (bottom band).

TimeOmni-VL	<think> 1) Geometry & Layout. $n_{\text{vars}}=3$ , band height $h=298$ . Total cycles $N=10$ , cycle width $w=89.6$ . Target Band: Variable 2 (Middle, Green channel). Y-range: $[298,595]$ . 2) Region Localization. • Cycle 7: X-range $[\lfloor 6\times 89.6\rfloor,\lfloor 7\times 89.6\rfloor-1]=[537,626]$ . • Cycle 9: X-range $[\lfloor 8\times 89.6\rfloor,\lfloor 9\times 89.6\rfloor-1]=[716,805]$ . 3) Statistical Comparison. Calculating mean pixel values (Green channel): • $\mu_{\text{cyc7}}=192.8$ (High brightness). • $\mu_{\text{cyc9}}=120.9$ (Low brightness). Difference: $\Delta=\|192.8-120.9\|=71.9$ . Cycle 7 is significantly brighter than Cycle 9. </think> Answer: 7 is brighter (difference: 54.6%).

Appendix

Table 18: Example of Understanding Task 4: Mean Comparison.

Understanding Task 5: Anomaly Detection
User	You are given a 896 $\times$ 896 image that encodes one or more time series variables as horizontal bands stacked from top to bottom. For variable 3, identify cycles that show significantly abnormal brightness compared to the typical level. Given metadata: $L=240$ , periodicity = 24. Definition: “Significantly abnormal” means cycles where the average pixel value (0-255 scale) is either brighter (mean + 18) or darker (mean - 18) than the overall mean for this variable. Provide the count and bounding boxes (x1, y1, x2, y2) for each anomalous cycle. Note: Variables are indexed from 1 (top band) to nvars (bottom band).

TimeOmni-VL	<think> 1) Geometry & Layout. Target: Variable 3 (Blue channel). Band height $h=298$ . Y-range $[596,893]$ . Total cycles $N=10$ . Cycle width $w=89.6$ . 2) Statistical Analysis. Computed global mean of 10 cycles: $\mu=126.6$ . Thresholds: • Bright: $>126.6+18=144.6$ . • Dark: $<126.6-18=108.6$ . 3) Anomaly Identification. Scanning individual cycle averages: • Cycle 2: Mean $67.5<108.6$ $\rightarrow$ Dark Anomaly. Bbox: $[(89,596),(178,893)]$ . • Cycle 5: Mean $185.7>144.6$ $\rightarrow$ Bright Anomaly. Bbox: $[(358,596),(447,893)]$ . </think> Answer: For variable 3, there are 2 anomalous cycles (1 bright, 1 dark). Bounding boxes: [(89, 596), (178, 893)], [(358, 596), (447, 893)].

Appendix

Table 19: Example of Understanding Task 5: Anomaly Detection.

Understanding Task 6: Trend Analysis
User	You are given a 896 $\times$ 896 image that encodes one or more time series variables as horizontal bands stacked from top to bottom. For variable 3, cycle 10, analyze the time series within this cycle: 1. What color channel is used for this variable? 2. What is the bounding box (x1, y1, x2, y2) of this cycle region? 3. After recovering to original time series values, describe the pattern and trend of this cycle in 2-3 sentences. Given metadata: $L=240$ , periodicity = 24. Recovery information: - Denorm: $\text{val}=\tanh(\text{norm})\cdot 4.0\cdot\sigma+\mu$ - Mean ( $\mu$ ): 1.52, Std ( $\sigma$ ): 2.77 Note: Variables are indexed from 1 (top band) to nvars (bottom band).

TimeOmni-VL	<think> 1) Geometry & Layout. Image size $896\times 896$ , $n_{\text{vars}}=3$ . Band height $h=\lfloor 896/3\rfloor=298$ . Total cycles $N=\lceil 240/24\rceil=10$ . Cycle width $w=896/10=89.6$ pixels. 2) Localization. Target: Variable 3 (Bottom), Cycle 10 (Rightmost). • Y-range: $[(3-1)\times 298,3\times 298-1]=[596,893]$ . • X-range: $[\lfloor 9\times 89.6\rfloor,\lfloor 10\times 89.6\rfloor-1]=[806,895]$ . 3) Recovery & Analysis. Channel: Blue (associated with Var 3). Applying the denormalization formula reveals the raw values. The series starts at 1.57, rises to a peak, and then fluctuates downwards. </think> Answer: 1. Color channel: Blue . 2. Bounding box: [(806, 596), (895, 893)] . 3. Trend analysis: The series initially rises to a prominent peak of 3.22 at $t=5$ , followed by a substantial and fluctuating decline to its minimum of -0.01 at $t=16$ . It then recovers moderately towards the end. Overall, the series shows a slight net decrease.

Appendix

Table 20: Example of Understanding Task 6: Trend Analysis.

Generation Task 1: Time Series Forecasting
User	You are given a 896 $\times$ 896 image that encodes 2 time series variable(s) as horizontal bands stacked from top to bottom. Each series contains 2 cycles, where each cycle has 288 time steps (totaling $2\times 288=576$ observations). The right side is masked and needs to be predicted for the next 1 cycle (totaling $1\times 288=288$ observations). Based on the observable parts in the time series image, please restore the masked right side.

TimeOmni-VL	<think> 1) Variable Counting. The image contains 2 independent time series encoded as horizontal bands. Brighter pixels indicate larger values, and darker pixels indicate smaller ones. 2) Variable Y-Range. Total size is $896\times 896$ . With $n_{\text{vars}}=2$ , each band height is $h=896/2=448$ pixels. Total cycles $=2\text{ (history)}+1\text{ (pred)}=3$ . Cycle width $\approx 896/3$ pixels. 3) Cycle Bounding Box. The task is to predict the rightmost 3rd cycle. • Var 1 (Top): Pred bbox $\approx[(597,0),(895,447)]$ . • Var 2 (Bottom): Pred bbox $\approx[(597,448),(895,895)]$ . 4) Mean Comparison & Anomaly Detection & Trend Analysis. First, analyze the historical pattern (576 steps) for trend and seasonality. Then, extrapolate the next 288 steps for each variable independently. Finally, restore the visual representation maintaining consistent brightness encoding. </think> Special Note: Following the GIFT-Eval evaluation protocol ( Aksu et al., 2024 ) , the required prediction length ( $P$ ) for this case is $48$ . Although the model generates a sequence of $288$ points, the output is truncated to the first $48$ points for quantitative evaluation.

Appendix

Table 21: Example of Generation Task 1: Time Series Forecasting.

Generation Task 2: Time Series Imputation
User	You are given a 896 $\times$ 896 image that encodes 1 time series variable(s) as horizontal bands stacked from top to bottom. Each series contains 6 cycles, where each cycle has 24 time steps (totaling $6\times 24=144$ observations). Some regions in these series are masked (shown as black areas) and need to be imputed. Different series may have masked regions at different positions, and the number of masked cycles can vary across series. Based on the observable parts in the time series image, please restore all the masked black regions for each series.

TimeOmni-VL	<think> 1) Variable Counting. The image contains 1 independent time series encoded as horizontal bands. Brighter pixels indicate larger values, and darker pixels indicate smaller ones. 2) Variable Y-Range. Total size is $896\times 896$ . With $n_{\text{vars}}=1$ , the band height is $h=896$ pixels. Total cycles $=6$ . Cycle width $\approx 896/6\approx 149$ pixels per cycle. 3) Cycle Bounding Box. Each series has specific masked black regions. • Var 1: Missing cycles 2–3. Mask bbox $\approx[(149,0),(447,895)]$ . 4) Mean Comparison & Anomaly Detection & Trend Analysis. First, analyze the observable patterns before (cycle 1) and after (cycles 4–6) the gap to identify trend and seasonality. Then, impute the missing cycles (2–3) using the surrounding context, maintaining consistent brightness encoding. Finally, output the restored image. </think>

Appendix

Table 22: Example of Generation Task 2: Time Series Imputation.

Bad Case I: Time Series Forecasting
User	You are given a 896 $\times$ 896 image that encodes 7 time series variable(s) as horizontal bands stacked from top to bottom. Each series contains 4 cycles, where each cycle has 24 time steps (totaling $4\times 24=96$ observations). The right side is masked and needs to be predicted for the next 2 cycles (totaling $2\times 24=48$ observations). Based on the observable parts in the time series image, please restore the masked right side.

Bagel
TimeOmni-VL

Appendix

Table 23: Bad Case I: A failure case of Bagel in the time series forecasting task.

Bad Case II: Time Series Imputation
User	You are given a 896 $\times$ 896 image that encodes 7 time series variable(s) as horizontal bands stacked from top to bottom. Each series contains 24 cycles, where each cycle has 96 time steps (totaling $24\times 96=2304$ observations). Some regions in these series are masked (shown as black areas) and need to be imputed. Different series may have masked regions at different positions, and the number of masked cycles can vary across series. Based on the observable parts in the time series image, please restore all the masked black regions for each series.

Bagel
TimeOmni-VL

Appendix

Table 24: Bad Case II: A failure case of Bagel in the time series imputation task.

Explore similar work

May 18, 2026

Chronicle: A Multimodal Foundation Model for Joint Language and Time Series Understanding

Real-world time series come with text: metadata, descriptions, news, reports. Yet time series foundation models process numerical sequences in isolation, and the multimodal text-and-time-series models that attempt to bridge the two all adapt a pretrained language model post hoc, inheriting representations shaped without ever seeing temporal data. These models are also evaluated almost exclusively against other multimodal baselines, not against the strongest unimodal foundation models in either domain, leaving open whether joint training is needed at all. We present Chronicle, a compact 324M-parameter decoder-only transformer trained from scratch on natural language and time series within a single unified architecture. Both modalities share the same transformer blocks, attention mechanism, and residual stream; the bulk of pretraining uses unimodal batches so cross-modal capability emerges purely from shared parameters, with a short alignment stage that interleaves the two. To our knowledge, Chronicle is the first model jointly pretrained on text and time series from scratch, and the first multimodal model evaluated against dedicated foundation models in both domains. It matches Gemma-3-270M-PT on 19 NLU tasks, sets a new bar for frozen-embedding time series classification on 24 UCR/UEA datasets, and produces multimodal forecasts on Time-MMD that beat every supervised fusion baseline, all from a single backbone.

Paul Quinlan, Jeremy Levasseur, Qingguo Li +1

1InertialAI · Department of Electrical and Computer Engineering, Queen’s University · Department of Mechanical and Materials Engineering, Queen’s University

Cross-Modal Learning Multivariate Time Series Forecasting Time Series Classification Unified Multimodal Models Multimodal Pretraining Multimodal Foundation Models+2

Ask PDF

Oct 13, 2026

MMTA: Benchmarking Multimodal Temporal Analysis with Time Series, Text, and Vision

Time-series analysis is increasingly studied in multimodal settings, yet most benchmarks pair temporal signals with text or with visualizations derived from the same signal. This leaves open whether multimodal large language models (MLLMs) can integrate time series with real-world visual evidence. We introduce MMTA (multimodal temporal analysis), a benchmark of tri-modal tasks with diverse cross-modal complementarity patterns among time series, text, and vision. MMTA spans nine domains -- robotics, autonomous driving, human action, manufacturing, AR assistance, healthcare, recommendation, finance, and climate -- under video+TS and image+TS layouts, with a 3,524-sample test split covering classification and forecasting. The benchmark standardizes heterogeneous real-world sources into a shared schema with domain-specific prompts, provenance documentation, and success-weighted metrics that penalize both wrong answers and failed generations. Evaluating eight open-source and proprietary MLLMs reveals a shared bottleneck: representing time series as text creates very long token sequences and weakens fine-grained alignment with visual evidence, especially in long-video settings. We then introduce TimeOmni-v, a multimodal LLM extension with a dedicated time-series encoder, temporal positional alignment, modality-aware token layout, and a forecasting head. TimeOmni-v improves long-video classification and time-series inference efficiency, showing that MMTA not only exposes representation and grounding failures but also motivates targeted model design.

Ziyang Zhang, Shenyi Li, Yilin Wang +4

Tsinghua University · Shanghai Artificial Intelligence Laboratory

Multimodal Large Language Models Multimodal Model Evaluation LLM-Based Time Series Forecasting Temporal Video Understanding

Ask PDF

Aug 5, 2026

Decoupling Perception from Description: Computation-Grounded Representation Alignment between Multivariate Time Series and Language

Training multimodal models to align time series with language runs into a self-supervision trap. The usual recipe asks an LLM to read a series and write a description, so label quality is capped by the perceptual skill the model is supposed to learn. The data can never teach more than the labeler already knows. A second gap makes this worse: most datasets use a single variable, but the patterns that matter (cross-channel correlation, lead-lag structure, co-occurring anomalies) appear only with several variables, right where the labeling LLM's limits are most exposed. These two problems create a trilemma: existing methods are reliable, realistic, or scalable, but none achieves all three. We resolve this by decoupling perception from description. Deterministic code computes a set of statistics from real, open-source multivariate series; the LLM verbalizes those precomputed facts. Perception, which LLMs do poorly, is handled by computation, while the LLM handles expression. This produces CGTime, our 4B-parameter computation-grounded time-series-language model. CGTime outperforms far larger general-purpose models on multivariate understanding tasks: it attains the best multivariate fact score on our held-out benchmark (0.283 vs. 0.173 for GPT-4o-mini and 0.203 for GPT-5.4-nano), a gap that survives Holm-corrected paired significance tests against every baseline. It also states verifiable numerical facts in generated captions more accurately and covers a broader range of statistical properties.

Xinran Feng, Yi Xie, Chao Zhang +4

Thrust of Financial Technology, The Hong Kong University of Science and Technology (Guangzhou) · Technical University of Munich · Heilbronn Data Science Center, Munich Data Science Institute +1

Cross-Modal Representation Learning Multimodal Pretraining

Ask PDF