KG-GAN: Knowledge Graph-Constrained GAN for Culturally Faithful Virtual Scene Synthesis
Informatica · 18.03.2026 · 50 (10)
DOI: 10.31449/inf.v50i10.12258 ↗Kurz gefasst
Fachartikel in der Zeitschrift Informatica, online erschienen am 18.03.2026. Verfasst von Peishu Song. Die Arbeit wurde bisher 1-mal zitiert (Crossref).
Abstract (Originalsprache)
This paper presents a knowledge-guided generative adversarial network (KG-GAN) for culturally faithful virtual scene construction and immersive experience optimization. A domain knowledge graph (KG) encodes traditional cultural entities and relations as graph embeddings, which condition the generator together with textual prompts. The semantic consistency discriminator calculates text-image cosine similarity based on CLIP cross-modal embedding, jointly scoring lexical semantic consistency and symbol reconstruction accuracy. The generator employs AdaIN for knowledge-conditional modulation, and multi-scale training and PWC-Net optical flow regularization collaboratively optimize local details and inter-frame stability. To balance global realism and local cultural details., we adopt a multi-scale training schedule and an attention controller that dynamically allocates textures and color palettes to culture-critical regions. Temporal stability is further improved with a lightweight consistency loss.Experiments on a curated cultural corpus and human-corrected pairs report cultural semantic fidelity of 89.7%–92.1%, relation compliance of 86.9%–88.6%, inter-frame SSIM under fast transitions of 0.782, and optical-flow smoothness of 0.751. A vocational Chinese-language classroom case study shows higher cultural expressiveness, learner engagement, and motion comfort compared with diffusion-based and CLIP-guided baselines, with ablations confirming the contributions of KG constraints, the semantic discriminator, and attention control. The framework is plug-and-play for XR/education/museums, and we provide KG schemas, code, and evaluation scripts to support reproducible deployment.
Zitieren
APA
Song, P. (2026). KG-GAN: Knowledge Graph-Constrained GAN for Culturally Faithful Virtual Scene Synthesis. Informatica, 50(10). https://doi.org/10.31449/inf.v50i10.12258
Vancouver
Song P. KG-GAN: Knowledge Graph-Constrained GAN for Culturally Faithful Virtual Scene Synthesis. Informatica. 2026;50(10). doi:10.31449/inf.v50i10.12258
BibTeX
@article{Song2026,
author = {Song, Peishu},
title = {KG-GAN: Knowledge Graph-Constrained GAN for Culturally Faithful Virtual Scene Synthesis},
journal = {Informatica},
year = {2026},
volume = {50},
number = {10},
doi = {10.31449/inf.v50i10.12258}
}
EUVIMED