EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Query-Adaptive Diffusion Transformer for Efficient Any-Subset Conditional Sampling on Tabular Data and Beyond

Juno Hwang, Junghyo Jo

IEEE Access · 2026 · 14 · 170007-170026

DOI: 10.1109/access.2026.3748904 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Juno Hwang und 1 weiteren Autor:innen. Die Arbeit nennt 48 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Research Foundation of Korea, 2021 Cultural Heritage Smart Preservation and Utilization Research and Development Program from the Cultural Heritage Administration National Research Institut….

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Diffusion models provide a flexible framework for conditional generation across tabular and multimodal domains. Many scientific and statistical inference tasks, however, require arbitrary conditional sampling, such as estimating posterior distributions, conditional expectations, or uncertainty under any subset of observed variables. Existing diffusion methods that support such flexibility typically denoise and sample over the full variable set, leading to computation that scales with data dimensionality rather than the query itself. We propose Query-Adaptive Diffusion Transformer (QuADiT), a diffusion framework that formulates conditional generation as query-specific inference. By representing variables as variable-length tokens, QuADiT restricts diffusion to only the conditioning and target variables specified by a query, eliminating redundant computation while preserving conditional flexibility. We further study practical design and implementation choices, identifying practical defaults on four tabular benchmarks. Finally, we demonstrate that the approach extends beyond tabular data to multimodal tasks, including tabular-conditioned image generation.

Förderung

  • National Research Foundation of Korea
  • 2021 Cultural Heritage Smart Preservation and Utilization Research and Development Program from the Cultural Heritage Administration National Research Institut…

Zitieren

APA

Hwang, J., & Jo, J. (2026). Query-Adaptive Diffusion Transformer for Efficient Any-Subset Conditional Sampling on Tabular Data and Beyond. IEEE Access, 14, 170007-170026. https://doi.org/10.1109/access.2026.3748904

Vancouver

Hwang J, Jo J. Query-Adaptive Diffusion Transformer for Efficient Any-Subset Conditional Sampling on Tabular Data and Beyond. IEEE Access. 2026;14:170007-170026. doi:10.1109/access.2026.3748904

BibTeX
@article{Hwang2026,
  author = {Hwang, Juno and Jo, Junghyo},
  title = {Query-Adaptive Diffusion Transformer for Efficient Any-Subset Conditional Sampling on Tabular Data and Beyond},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {170007--170026},
  doi = {10.1109/access.2026.3748904}
}
Kostenlos starten →