morphingthroughtime

Morphing Through Time — Pretrained Weights

Weights for Morphing Through Time: Diffusion-Based Bridging of Temporal Gaps for Robust Alignment in Change Detection (Madani & Patel).

• 📄 Paper: https://arxiv.org/abs/2511.07976
• 💻 Code: https://github.com/Anita-Madani/Morphing-through-time-

<p align="center"> <img src="pipeline.png" width="100%" alt="Morphing Through Time pipeline"> </p>

Given a bi-temporal pair (I_A, I_B), DiffMorpher synthesizes K=5 intermediate frames; RoMa estimates the flow between consecutive frames, composed into F_{A→B}; a residual flow-refinement U-Net corrects it to F̂_{A→B}, which warps I_B onto I_A before the (frozen) change-detection backbone.

Contents

This repository hosts the trained Stage-3 residual-refiner checkpoints (<dataset>/refiner.pth for LEVIR / WHU / DSIFN). The diffusion backbone (Stable Diffusion 2.1) and RoMa weights download automatically on first use, so they are not stored here.

pip install -U huggingface_hub
bash scripts/download_weights.sh    # from the code repo; pulls checkpoints into ./checkpoints/

License

Non-commercial research use (CC BY-NC 4.0). The morphing stage is derived from DiffMorpher under the S-Lab License 1.0; see the code repository for details.

Citation

@article{madani2025morphing,
  title   = {Morphing Through Time: Diffusion-Based Bridging of Temporal Gaps for Robust Alignment in Change Detection},
  author  = {Madani, Seyedehanita and Patel, Vishal M.},
  journal = {arXiv preprint arXiv:2511.07976},
  year    = {2025}
}