Website | Paper | Video | Zenodo
This is the official repository for the ICLR 2024 paper "DiffusionSat: A Generative Foundation Model For Satellite Imagery".
Authors: Samar Khanna 1, Patrick Liu, Linqi (Alex) Zhou, Chenlin Meng, Robin Rombach, Marshall Burke, David B. Lobell, Stefano Ermon.
We use conda to create our environments. You will have to do the following:
cd DiffusionSat
conda create -n diffusionsat python=3.10
# if you want cuda 11.8, run this replace the index url with https://download.pytorch.org/whl/cu118
pip install torch==2.2.2 torchvision==0.17.2 --index-url https://download.pytorch.org/whl/cu121
pip install -e ".[torch]" # install editable diffusers
pip install -r requirements_remaining.txtModel checkpoint files have been uploaded to Zenodo within the DiffusionSat community at this link.
OLD: Model checkpoint files were previously available on Google Drive.
Note that the files on Google Drive may not be forever available, and could be taken down at any moment.
(While the files are on Google Drive, you can use gdown to download them).
This section covers image-generation using single-image DiffusionSat, without control signal inputs.
The relevant jupyter notebook can be found in notebooks/single-image.ipynb.
The relevant model checkpoints can be found here:
| Resolution | Zenodo Page | Download Link |
|---|---|---|
| 512 x 512 | View | Download |
| 256 x 256 | View | Download |
The Jupyter notebook that demonstrates generation with 3D ControlNets is shown
for the Texas housing dataset in notebooks/controlnet_texas_samples.ipynb.
Generating with a ControlNet that accepts a single conditioning image + metadata is similar.
The relevant model checkpoints can be found here:
| Task | Zenodo Page | Download Link |
|---|---|---|
| Texas Super-resolution | View | Download |
| fMoW Sentinel -> RGB Super-resolution | View | Download |
These sections describe how to launch training using accelerate.
In this repository, we provide an example config file to use with accelerate in launch_accelerate_configs.
You can also configure your own file by running accelerate config in your terminal and following the steps.
This will save the config file in the cache location (eg: .cache/huggingface/accelerate/default_config.yaml),
and you can simply copy over the .yaml file to launch_accelerate_configs/ or remove the
--config_file argument from accelerate launch in the bash script.
See this section for more details on how to use webdataset for training.
You will need to specify the dataset shardlist .txt files in ./webdataset_shards.
To train the (text, metadata) -> single_image model, use the following command:
./launch_scripts/launch_256_fmow_satlas_spacenet_img_txt_md.sh launch_accelerate_configs/single_gpu_accelerate_config.yamlTo train the (text, target_metadata, conditioning_metadata, conditioning_images) -> single_image ControlNet model, use the following commands,
detailed below.
As a quick summary, these scripts use a frozen single-image model (see above) as a prior to train a ControlNet (which could be a 3D ControlNet for temporal conditioning images). This ControlNet can then generate a new image for the desired input text and metadata prompt, conditioned on additional metadata and images.
You will also need to provide the path to the single-image model checkpoint
(by specifying this path in the UNET_PATH variable) that will remain frozen throughout training.
./launch_scripts/launch_texas_md_controlnet.sh launch_accelerate_configs/single_gpu_accelerate_config.yamlThis task uses the Texas housing dataset from satellite-pixel-synthesis-pytorch.
The task is: given a low-res and high-res image of a location at time T, and a low-res image of the same location
at time T', generate a high-res image of the location at time T'.
./launch_scripts/launch_fmow_md_superres.sh launch_accelerate_configs/single_gpu_accelerate_config.yamlThe task is: given a multi-spectral low-res image of a location (from fMoW-Sentinel), generate the corresponding high-res RGB image (from fMoW-RGB).
./launch_scripts/launch_fmow_temporal_md_controlnet.sh launch_accelerate_configs/single_gpu_accelerate_config.yamlThis model conditions on a temporal sequence of input RGB images from fMoW-RGB to generate a single new image at a desired timestamp T.
./launch_scripts/launch_xbd_md_controlnet.sh launch_accelerate_configs/single_gpu_accelerate_config.yamlThe task is: given a past (or future) image of a location affected by a natural disaster, generate the future (or past) image after (or before) the natural disaster struck. We use the xBD dataset.
The datasets we use are in webdataset format.
You will need to prepare your datasets in this format to be able to train using the given code,
or you can modify the data-loading to use your own custom dataset formats.
We have provided example shardlists in webdataset_shards. The training code will read the relevant file,
and load data using the data paths in this file. The advantage of using webdataset is that your data
does not need to only be on disk, and you can stream data from buckets in AWS S3 as well.
We also provide a small sample webdataset in webdataset_shards/texas_housing_val_10sample.tar, sourced from
the validation set of the Texas housing super-resolution task.
Example format for each entry in the fMoW webdataset .tar file.
__key__: fmow-{cls_name}-{instance_id} # eg: fmow-airport-airport_0
output.cls: label_idx # eg: 32
input.npy: (h,w,c) numpy array
metadata.json: {'img_filename': ..., 'gsd': ..., 'cloud_cover': ..., 'timestamp': ..., 'country_code': ...}
Note that fMoW also requires a metadata .csv file.
If you find our project helpful, please cite our paper:
@inproceedings{
khanna2024diffusionsat,
title={DiffusionSat: A Generative Foundation Model for Satellite Imagery},
author={Samar Khanna and Patrick Liu and Linqi Zhou and Chenlin Meng and Robin Rombach and Marshall Burke and David B. Lobell and Stefano Ermon},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=I5webNFDgQ}
}