ddpm-cd-pretrained-256

[!WARNING] we do not have a full checkpoint conversion validation, if you encounter pipeline loading failure and unsidered output, please contact me via [email protected]

BiliSakura/ddpm-cd-pretrained-256

Unconditional image generation pipeline — DDPM with SR3 backbone, pre-trained on remote-sensing imagery. Compatible with ddpm-cd-diffusers.

This is an image generation model only, not a change-detection pipeline. For change detection, use this UNet as a feature extractor and add a CD head (see pretrained-cd-models).

Model Details

• Architecture: SR3-style UNet (DDPM backbone, from Image-Super-Resolution-via-Iterative-Refinement)
• Resolution: 256×256
• Training: Unconditional DDPM on remote-sensing images
• Scheduler: DDPMScheduler, 2000 steps, cosine schedule

Usage

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

from diffusers import DDIMScheduler, DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "BiliSakura/ddpm-cd-pretrained-256",
    custom_pipeline="pipeline",
    trust_remote_code=True,
).to("cuda")
pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)

# Control speed vs quality with num_inference_steps (default 2000). Use fewer (e.g. 50–250) for faster inference.
images = pipe.generate(batch_size=4, image_size=256, num_inference_steps=50)

Optional: Use as feature extractor for change detection

This model can serve as a backbone for DDPM-CD. Add a CD head and fine-tune with train_cd.py; pre-trained CD heads are in ddpm-cd.

accelerate launch scripts/train_cd.py \
    --pretrained_model_path BiliSakura/ddpm-cd-pretrained-256 \
    --train_data_dir dataset/LEVIR-CD256 \
    --val_data_dir dataset/LEVIR-CD256 \
    --output_dir experiments/cd-levir \
    --resolution 256 \
    --timesteps 50 100 400 \
    --feat_type dec

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}
}