EUVIMEDEUVIMED entdecken
Crossref

Geometry-Aware Joint Attention for Efficient Native 3D Editing

Weiwei Cai, Shuangkang Fang, Weicai Ye, Xuanyang Zhang, Yanpei Cao, Gang Yu, Tao Chen

IEEE Signal Processing Letters · 2026 · 33 · 2615-2619

DOI: 10.1109/lsp.2026.3705718 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Weiwei Cai und 6 weiteren Autor:innen. Die Arbeit nennt 26 Literaturangaben. Gefördert von: National Key R&D Program of China, Natural Science Foundation of Shanghai Municipality.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Instruction-guided 3D editing aims to modify a 3D object according to textual instructions while preserving its overall geometric structure. However, existing methods either rely on optimization-based pipelines, which are slow at inference, or perform per-view 2D editing followed by reconstruction, often leading to geometric inconsistency. To address these limitations, we propose UniEdit3D, a native feed-forward 3D editing model that directly operates on structured 3D latent representations within a diffusion transformer. To enable effective feed-forward editing, we introduce Geometry-Aware Structured Attention Editing, a unified attention mechanism that enables joint self-attention reasoning over unified source and target latent tokens, explicitly modeling the spatial cross-correlation between the preserved geometry and the edited structure. Furthermore, to address the lack of training data, we develop an automated data construction pipeline to produce large-scale instruction-driven editing pairs. Experimental results demonstrate that our method reduces FID by 27.2% compared with the strongest prior baseline while generating edited 3D assets within seconds.

Förderung

  • National Key R&D Program of China
  • Natural Science Foundation of Shanghai Municipality

Zitieren

APA

Cai, W., Fang, S., Ye, W., Zhang, X., Cao, Y., Yu, G., & Chen, T. (2026). Geometry-Aware Joint Attention for Efficient Native 3D Editing. IEEE Signal Processing Letters, 33, 2615-2619. https://doi.org/10.1109/lsp.2026.3705718

Vancouver

Cai W, Fang S, Ye W, Zhang X, Cao Y, Yu G, et al. Geometry-Aware Joint Attention for Efficient Native 3D Editing. IEEE Signal Processing Letters. 2026;33:2615-2619. doi:10.1109/lsp.2026.3705718

BibTeX
@article{Cai2026,
  author = {Cai, Weiwei and Fang, Shuangkang and Ye, Weicai and Zhang, Xuanyang and Cao, Yanpei and Yu, Gang and Chen, Tao},
  title = {Geometry-Aware Joint Attention for Efficient Native 3D Editing},
  journal = {IEEE Signal Processing Letters},
  year = {2026},
  volume = {33},
  pages = {2615--2619},
  doi = {10.1109/lsp.2026.3705718}
}
Kostenlos starten →