.png){kind=icon}
Return
StructDiff: A Training-Free Framework for Precise Structural Control in Remote Sensing Image Synthesis
.png){kind=icon}
PRE
AI
DOI：10.1109/tmm.2026.3742634.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Diffusion models have advanced remote sensing (RS) image synthesis, yet maintaining strict geospatial integrity remains difficult. Existing methods face a practical trade-off: training-based approaches are condition-specific and require costly retraining, whereas conventional training-free techniques often struggle to balance structural preservation with prompt-driven appearance control. In this paper, we present StructDiff, a training-free framework for precise structural control in RS image synthesis without task-specific retraining for new control conditions. Its core component, the Fixed-Timestep Structural Conditioner (FTSC), extracts a geometric prior from a single fixed diffusion timestep and uses it as a stable structural constraint during generation. StructDiff is further equipped with a Multimodal Prompt Modulator (MPM), implemented with a frozen off-the-shelf vision-language model, and a Restart Refinement Strategy (RRS) for late-stage artifact suppression. Experiments on DIOR and DOTA, together with a downstream object detection study under low-data regimes, show that StructDiff provides a strong balance between geospatial accuracy, semantic alignment, and perceptual quality among training-free controllable baselines, supporting its use in controlled augmentation workflows such as on-the-fly data augmentation and rapid virtual prototyping in remote sensing. Code is publicly available at: https://github.com/bxuanz/StructDiff.
Keywords:
Remote Sensing
Diffusion Models
Structural Control
Training-Free
Image Synthesis
Journal
IF:
9.7
Papers:
4.6K
Citations:
2.4W
Organization
Cited Papers
No cited papers available