[!WARNING] we do not have a full checkpoint conversion validation, if you encounter pipeline loading failure and unsidered output, please contact me via bili_sakura@zju.edu.cn

BiliSakura/ddpm-cd

Consolidated DDPM-CD change detection — Single repo with shared UNet backbone and multiple cd_head variants (trained on different datasets and timestep configs).

Model Structure

• Backbone: Shared SR3-style UNet (same across all variants)
• cd_head: Dataset-specific change detection heads in cd_head/{variant}/

Available cd_head Variants

Usage

Load with explicit custom_pipeline (pipeline.py is in the repo, use relative path) and cd_head_subfolder:

from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "BiliSakura/ddpm-cd",
    custom_pipeline="pipeline",
    trust_remote_code=True,
    cd_head_subfolder="levir-50-100",
).to("cuda")

# Images in [-1, 1], shape (B, 3, H, W)
change_map = pipe(image_A, image_B, timesteps=[50, 100])
pred = change_map.argmax(1)  # (B, H, W), 0=no-change, 1=change

Important: Pass the same timesteps used during training for each variant (see table above).

Switching cd_head at Runtime

pipe = DiffusionPipeline.from_pretrained(
    "BiliSakura/ddpm-cd",
    custom_pipeline="pipeline",
    trust_remote_code=True,
    cd_head_subfolder="levir-50-100",
).to("cuda")
# Load different cd_head
pipe.load_cd_head(subfolder="whu-50-100-400")
change_map = pipe(image_A, image_B, timesteps=[50, 100, 400])

Citation

@inproceedings{bandaraDDPMCDDenoisingDiffusion2025,
  title = {{{DDPM-CD}}: {{Denoising Diffusion Probabilistic Models}} as {{Feature Extractors}} for {{Remote Sensing Change Detection}}},
  shorttitle = {{{DDPM-CD}}},
  booktitle = {Proceedings of the {{Winter Conference}} on {{Applications}} of {{Computer Vision}}},
  author = {Bandara, Wele Gedara Chaminda and Nair, Nithin Gopalakrishnan and Patel, Vishal},
  year = 2025,
  pages = {5250--5262},
  urldate = {2025-12-28}
}