GALA: Generation-Aware Cross-Modal Alignment for Text-to-Time-Series Synthesis

Haochen Zhang; Gengwei Zhang; Laura Yao; Nicholas Konz; Tianlong Chen

cs.CLAug 13, 2026

GALA: Generation-Aware Cross-Modal Alignment for Text-to-Time-Series Synthesis

Authors: Haochen Zhang, Gengwei Zhang, Laura Yao, Nicholas Konz, Tianlong Chen

Organizations: UNITES Lab, University of North Carolina at Chapel Hill

Abstract

Synthesizing time series from natural language is emerging as the most expressive form of controllable time series generation. However, existing text-conditioned generators either take caption embeddings frozen from off-the-shelf text encoders, or adapt the encoder end-to-end, letting the denoising loss shape the embeddings only as a by-product. In either case, the conditioning representation is never deliberately matched to the signal modality, leaving it ill-suited to guide generation. We address this by introducing GALA: Generation-Aware cross-modaL Alignment for text conditional time series generation. GALA is a two-stage approach that first contrastively couples a pretrained text encoder with a time-series foundation model into a shared embedding space with both encoders adapted to generation by an auxiliary generative loss, and then freezes the resulting caption embedding to drive a flow-matching generator. On TSFragment-600K, spanning four domains and three fragment lengths, GALA sets a new state of the art, ranking first in 30 of 36 metric columns and reaching an average rank of 1.08/1.08/1.42 at lengths 24/48/96 against 1.92/2.00/1.75 for the strongest baseline. We further find that generator-internal text encoders force a trade-off between fidelity and caption adherence, whereas conditioning on the aligned embedding breaks it: FID, CTTP, and JFTSD all improve at once. Ablating the auxiliary loss degrades FID, CTTP and JFTSD together, it indicates the generative term is a necessary component of the alignment rather than an add-on.

Figures & tables

Figure 1: GALA synthesizes a time series from a free-form caption by first aligning text and series into a shared, generation-aware embedding space, then conditioning a generator on the frozen text embedding.

Figure 2: Overview of GALA. Stage 1 contrastively aligns a LoRA-adapted text encoder and time-series encoder into a shared space, with a generation-aware auxiliary loss keeping the text embedding sufficient to synthesize its fragment. Stage 2 freezes the aligned text encoder and trains a flow-matching generator on its embedding.

Method	ETTh1s			ETTm1			Electricity			Traffic			Avg.
	FID $\downarrow$	CTTP $\uparrow$	JFTSD $\downarrow$	FID $\downarrow$	CTTP $\uparrow$	JFTSD $\downarrow$	FID $\downarrow$	CTTP $\uparrow$	JFTSD $\downarrow$	FID $\downarrow$	CTTP $\uparrow$	JFTSD $\downarrow$	Rank $\downarrow$
Length $L=24$
DiffuSETS	0.089	0.05	0.56	0.877	-0.04	1.13	0.102	0.10	0.31	0.080	0.14	0.23	5.08
T2S	0.078	0.07	0.54	0.254	0.11	0.53	0.166	0.09	0.37	0.403	0.07	0.56	5.17
BRIDGE	0.028	0.03	0.53	0.050	0.06	0.38	0.057	0.07	0.29	0.117	0.05	0.33	4.67
Text2Motion	0.067	0.22	0.38	0.025	0.36	0.15	0.055	0.24	0.18	0.077	0.26	0.17	3.08

Table 1: Comparison with existing text-to-time-series generation methods on TSFragment-600K across four domains and three lengths (

L\in\{24,48,96\}

). Each entry is the mean over

3

independent sampling seeds. Per-cell standard deviations are omitted for space and are uniformly small: the largest of all

216

0.0165

(DiffuSETS, FID on ETTh1s-

96

), the largest of GALA’s

36

0.0033

, and

98\%

are below

0.01

. Significance test is deferred to Technical Supplement. Bold and underline mark the best and second-best per column, ties broken by the unrounded means. Avg. Rank is the mean rank over that length’s

12

columns.

Table 6

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Value
Encoders
Text encoder	Embedding-Gemma-300M (frozen, LoRA)
Series encoder	Chronos-2 (frozen, LoRA)
Shared embedding dim $d$	$512$
Series normalization	dataset-level global $z$ -score
LoRA

Appendix

Table 5: Stage-1 alignment hyperparameters. In the fine-tuning-method ablation, FT-Full uses learning rate

1\times 10^{-5}

; all other settings are unchanged.

Hyperparameter	Value
Backbone (JiT/DiT-2D)
Normalization / attention	RMSNorm, QK-normalized
Positional encoding	2-D RoPE
Feed-forward / conditioning	SwiGLU, adaLN-Zero
Hidden size	$128$
Depth (layers)	$4$

Appendix

Table 6: Stage-2 generator hyperparameters.

Table 9

Table 10

Table 11

Table 12

	FID		CTTP		JFTSD
Ours vs.	W–L	$p_{\text{Holm}}$	W–L	$p_{\text{Holm}}$	W–L	$p_{\text{Holm}}$
VerbalTS	7–5	$0.133$	12–0	$0.0012$	11–1	$0.0012$
Text2Motion	12–0	$0.0012$	12–0	$0.0012$	12–0	$0.0012$
BRIDGE	11–1	$0.0068$	12–0	$0.0012$	12–0	$0.0012$
DiffuSETS	12–0	$0.0012$	12–0	$0.0012$	12–0	$0.0012$
T2S	12–0	$0.0012$	12–0	$0.0012$	12–0	$0.0012$

Appendix

Table 16: Pairwise one-sided Wilcoxon signed-rank tests of GALA against each baseline, run separately per metric over the

12

settings (four datasets

\times

three lengths), with Holm correction over the five comparisons within a metric. W–L counts the settings won and lost. The only comparison that fails to reach significance is FID against VerbalTS, i.e. the two are indistinguishable on fidelity.

Method	ETTh1s			ETTm1			Electricity			Traffic
	FID	CTTP	JFTSD	FID	CTTP	JFTSD	FID	CTTP	JFTSD	FID	CTTP	JFTSD
Length $L=24$
DiffuSETS	3.4	4.4	5.4	1.6	1.1	2.3	3.3	4.4	3.6	1.7	1.0	2.6
T2S	1.0	6.6	6.3	1.9	0.7	3.2	2.6	4.7	2.8	6.1	1.3	4.3
BRIDGE	0.5	8.2	8.0	1.5	2.3	2.6	1.9	3.2	3.8	9.1	3.7	7.3
Text2Motion	0.9	2.1	1.2	0.6	3.0	2.4	1.5	3.0	2.1	0.4	3.9	1.7

Appendix

Table 17: Sample standard deviation over the

3

sampling seeds for every cell of Table 1 in the main paper. All values are

\times 10^{-3}

, so an entry of

3.4

means

0.0034

. The largest entry in the whole table is

16.5

(

0.0165

; DiffuSETS, FID on ETTh1s-

96

) and the largest for GALA is

3.3

(

0.0033

	$L=24$			$L=48$			$L=96$
Source	PCA 2	PC 90	eff.	PCA 2	PC 90	eff.	PCA 2	PC 90	eff.
ETTh1s	$0.62$	$8$	$4.6$	$0.47$	$10$	$6.4$	$0.56$	$9$	$4.7$
ETTm1	$0.64$	$9$	$3.6$	$0.66$	$11$	$3.8$	$0.62$	$11$	$4.7$
Electricity	$0.71$	$6$	$3.8$	$0.54$	$8$	$5.5$	$0.44$	$11$	$7.2$
Traffic	$\mathbf{0.93}$	$\mathbf{2}$	$\mathbf{2.3}$	$\mathbf{0.90}$	$\mathbf{2}$	$\mathbf{2.5}$	$\mathbf{0.87}$	$\mathbf{4}$	$\mathbf{2.6}$

Appendix

Table 18: Shape diversity of the four sources, measured on

2000

per-series

z

-normalized fragments per configuration. PCA 2 is the share of shape variance in the first two principal components, PC 90 the number of components reaching

90\%

, and eff. the participation ratio of the spectrum. Traffic (bold) needs the fewest dimensions at every length, i.e. it contains the fewest distinct shapes.

Figure 6: Morphological diversity at

L=48

. (a) Cumulative PCA explained variance of the

z

-normalized fragments; markers sit where each source reaches

90\%

(dashed line). Traffic saturates after two components while the others need eight to eleven. (b, c) The six KMeans centroid shapes of Traffic and of ETTh1s on identical axes: Traffic’s clusters are one daily template at six different phases, whereas ETTh1s’s are distinct morphologies.

Explore similar work

Sep 14, 2026

CodeTS: Verifiable Text-to-Time Series Generation via Executable Code

Text-to-Time Series Generation (Text-to-TS) provides a promising paradigm for synthesizing time series from natural language, enabling scenario-specific generation when real observations are scarce or costly to acquire. However, existing methods typically lack an explicit mechanism for deriving generation logic from textual descriptions to guide time series synthesis. In this paper, we propose CodeTS, a verifiable framework that uses code as an intermediate generation interface, reformulating Text-to-TS generation as a Text-to-Code-to-TS process. CodeTS first maps textual temporal descriptions into an explicit code space, where executable code specifies how textual requirements shape target temporal patterns, and then obtains the time series through code execution. To learn this code generation process reliably without real code annotations, CodeTS constructs aligned Text-Code-TS triplets from structured temporal attributes for supervised initialization. More importantly, we further design multi-stage execution-based rewards that verify format validity, code executability, and time series quality, enabling real Text-TS pairs to provide training signals for Reinforcement Learning with Verifiable Rewards (RLVR). Extensive experiments on eight benchmarks across short, medium, and long generation lengths demonstrate that CodeTS provides a strong zero-shot solution for Text-to-TS generation, outperforming LLM-based baselines and achieving better averaged results than supervised generative baselines trained on the target datasets.

Xudong Yuan, Shunyu Liu, Tongya Zheng +3

Zhejiang University, Hangzhou, China · Nanyang Technological University, Singapore · South China University of Technology, Guangzhou, China

Time Series Generation

Ask PDF

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

Jun 6, 2026

Mix, Don't Pick: Why Synthetic Corpus Composition Matters for Time Series Foundation Model Pretraining

Choosing the wrong synthetic generator for time-series foundation model pretraining is costly: under identical training budgets, the best and worst generators produce up to a

2\times

gap in forecasting error, yet the field has no principled way to make this choice. The problem is compounded by the fact that generator rankings are not stable across architectures: across 11 generator families evaluated on Chronos-T5-Mini and Moirai-Small trained from scratch, we find that which generators are useful depends on the model architecture. Rather than solving the generator selection problem, we sidestep it: a simple equal-weight mixture of all generators matches or beats the best individual generator for both architectures, and composing this mixture with real data yields the strongest pretraining corpora overall. Synthetic pretraining is therefore a corpus composition problem, not a generator selection problem, and composition choices should be validated per model family rather than assumed to transfer.

Aaryan Nagpal, Debdeep Sanyal, Murari Mandal +2

Birla AI Labs, Mumbai, India · KIIT, Bhubaneswar, India · BITS Pilani, Pilani, India

Synthetic Data Pretraining Time Series Forecasting Time-Series Foundation Models Synthetic Data Evaluation

Ask PDF