Note: images are sorted by their likelihood. That's why images with smaller idx are much more noisy. We will release a filtered version soon.
We collected 90 000 high-resolution landscape images from Unsplash (64,669 images) and Flickr (25,331 images).
| Path | Size | Number of files | Format | Description |
|---|---|---|---|---|
| Landscapes HQ | 283G | 90,000 | PNG | The root directory with all the files |
| ├ LHQ | 155G | 90,000 | PNG | The complete dataset. Split into 4 zip archives. |
| ├ LHQ1024 | 107G | 90,000 | PNG | LHQ images, resized to min-side=1024 and center-cropped to 1024x1024. Split into 3 zip archives. |
| ├ LHQ1024_jpg | 12G | 90,000 | JPG | LHQ1024 converted to JPG format with quality=95 (with Pillow)* |
| ├ LHQ256 | 8.7G | 90,000 | PNG | LHQ1024 resized to 256x256 with Lanczos interpolation |
| └ metadata.json | 27M | 1 | JSON | Dataset metadata (author names, licenses, descriptions, etc.) |
*quality=95 in Pillow for JPG images (the default one is 75) provides images almost indistinguishable from PNG ones both visually and in terms of FID.
Downloading files:
python download_lhq.py [DATASET_NAME]
https://creativecommons.org/publicdomain/zero/1.0/ http://www.usa.gov/copyright.shtml
The individual images from LHQ dataset have one of the following licenses:
- Unsplash License
- Creative Commons Attribution License
- Creative Commons Attribution-NonCommercial License
- Creative Commons Public Domain Mark
- Creative Commons Public Domain Dedication (CC0)
- United States Government Work
To see, which image has which license, please see the corresponding metadata.
The dataset itself is published under Creative Commons Attribution 2.0 Generic (CC BY 2.0) License: https://creativecommons.org/licenses/by/2.0/. This means, that you can use it however you like, but you should attribute the source (i.e. give a link to this repo or cite the paper).
Images were obtained by downloading 450k images from Unsplash and Flickr using a set of 400 manually constructed search queries and preprocessing it with a pretrained Mask R-CNN to filter out images that likely contained objects and Inception V3 statistics to remove "too out-of-distribution" samples. For more information, see Section 4 of the paper: https://arxiv.org/abs/2104.06954
@article{ALIS,
title={Aligning Latent and Image Spaces to Connect the Unconnectable},
author={Skorokhodov, Ivan and Sotnikov, Grigorii and Elhoseiny, Mohamed},
journal={arXiv preprint arXiv:2104.06954},
year={2021}
}