2026-10-14T14:24:31Z https://www.scidb.cn/oai

10.57760/sciencedb.j00240.00256 2026-07-03T11:44:16Z

2026-07-03 Dongba Painting Portrait Generation Dataset (DongbaPGD) doi:10.57760/sciencedb.j00240.00256 en To achieve high-quality generation of character images in Dongba painting style, this article constructs a Dongba painting portrait image and text dataset. Collect high-definition scanned images of physical objects and related literature from the collection of Lijiang Museum. In response to the issue of overlapping characters and environments, this article uses manual annotation segmentation to extract character regions, and obtains a total of 93 high-quality, unobstructed portrait samples of Dongba paintings. Experts familiar with Naxi Dongba culture write prompt words for each image, mainly describing key information such as actions, clothing, and colors. Finally, a Dongba painting portrait image text dataset that can be used for model training and evaluation is constructed.  This paper constructs a functional dataset for the suppression of human body structure artifacts and style constraints: (1) Dongba painting generated portrait artifact segmentation dataset. To train DGASNet, pixel level annotation of human artifact regions was performed using SAM (segment anything model) (Kirillov et al., 2023) to construct a dataset containing 1000 generated portrait samples from Dongba paintings, covering both normal and multi class structural artifacts. （2） Style reference dataset. To construct a DSCR style prototype library, 62 representative Dongba portraits were selected as style reference samples, covering typical Dongba painting color features and line strokes.  dongba; dongba painting; text-to-image generation Bai Yufan PUBLIC https://creativecommons.org/licenses/by-nc-nd/4.0/ dataset Science Data Bank