TopoDiT: A Topology-Constrained Diffusion Transformer for Narrative-Consistent Landscape Image-Sequence Generation of China’s Ancient Trail Landscapes
IEEE Access · 2026 · 14 · 61993-62010
DOI: 10.1109/access.2026.3686219 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Jun Zheng und 4 weiteren Autor:innen. Die Arbeit nennt 42 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Key Project of Philosophy and Social Sciences of Zhejiang Province, China.
Abstract (Originalsprache)
The digital interpretation of Linear Cultural Heritage (LCH) faces a persistent difficulty: translating abstract historical narratives into coherent visual representations of landscape experience. Standard diffusion models often suffer from semantic entanglement and sequence inconsistency, which lead to material bleeding, unstable scene transitions, and perceptual discontinuity across generated images. This study introduces Topology-constrained Diffusion Transformer (TopoDiT), a topology-constrained Diffusion Transformer for narrative-consistent landscape image-sequence generation along the 235-km Zhejiang–Anhui Ancient Trail. The framework leverages Persistent Homology to encode sequence transitions into differentiable structural embeddings and integrates Adjacent Semantic Consistency ( $\mathcal {L}_{adj}$ ) and Transition-Type Consistency ( $\mathcal {L}_{trans}$ ) losses. It was trained on a multimodal dataset of 8,420 curated text-image pairs. Results show that TopoDiT improves single-image quality and cross-frame stability (FID = 18.24; LPIPS Var. = 0.068) relative to unconstrained baselines. We report an internal Topo-Consistency Rate (iTCR) of 79.3% as a training-aligned diagnostic. To mitigate self-validation risk caused by reusing the same feature space and transition logic as the training objectives, we introduce independent external evaluations: an external topology-consistency metric (eTCR) computed with a held-out encoder/classifier not used in training, and a blinded expert assessment of sequence plausibility and transition readability. These external metrics demonstrate that TopoDiT improves narrative-consistent 2D image-sequence generation without claiming geometrically validated landscape space. It therefore offers a scalable workflow for the digital interpretation of cultural landscape narratives, while not claiming terrain-faithful or construction-ready spatial reconstruction.
Förderung
- Key Project of Philosophy and Social Sciences of Zhejiang Province, China
Zitieren
APA
Zheng, J., Wang, B., Zhou, X., Wu, X., & Zhao, S. (2026). TopoDiT: A Topology-Constrained Diffusion Transformer for Narrative-Consistent Landscape Image-Sequence Generation of China’s Ancient Trail Landscapes. IEEE Access, 14, 61993-62010. https://doi.org/10.1109/access.2026.3686219
Vancouver
Zheng J, Wang B, Zhou X, Wu X, Zhao S. TopoDiT: A Topology-Constrained Diffusion Transformer for Narrative-Consistent Landscape Image-Sequence Generation of China’s Ancient Trail Landscapes. IEEE Access. 2026;14:61993-62010. doi:10.1109/access.2026.3686219
BibTeX
@article{Zheng2026,
author = {Zheng, Jun and Wang, Bingjiang and Zhou, Xiaofei and Wu, Xue and Zhao, Shiru},
title = {TopoDiT: A Topology-Constrained Diffusion Transformer for Narrative-Consistent Landscape Image-Sequence Generation of China’s Ancient Trail Landscapes},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {61993--62010},
doi = {10.1109/access.2026.3686219}
}
EUVIMED