T2S
Summary
T2S is a text-to-time-series generation model that combines a length-adaptive VAE with a text-conditioned Diffusion Transformer trained through flow matching.
Role In The Wiki
T2S anchors the text-to-series generation branch. It is useful for comparing flow-matching generation interfaces: unlike Sundial, which conditions on numeric history for probabilistic forecasting, T2S conditions on natural-language captions to generate synthetic time-series instances.
It also strengthens the synthetic-data thread because it introduces TSFragment-600K, a fragment-level text-time-series dataset produced by captioning local temporal morphology in existing time-series datasets.
T2S should not be treated as forecasting or action-conditioned world-model evidence: it has no observed-history input and no explicit action, control-input, or intervention channel.
Official Artifacts
- Official code: https://github.com/WinfredGe/T2S
- Official dataset: https://huggingface.co/datasets/WinfredGe/TSFragment-600K
- Official checkpoint: https://huggingface.co/WinfredGe/T2S-pretrained_LA-VAE
- Official checkpoint: https://huggingface.co/WinfredGe/T2S-DiT
Evidence
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in t2s-2025 rather than duplicating verdict rows here.
At the entity level, T2S anchors the text-to-series generation branch. It is useful for comparing flow-matching generation interfaces: unlike Sundial, which conditions on numeric history for probabilistic forecasting, T2S conditions on natural-language captions to generate synthetic time-series instances. This page should stay as the object card; source pages carry slot-level verdicts, evidence, and missing pieces.