image
imagewidth (px) 1.92k
1.92k
|
---|
A newer version of this dataset is available. Please use that one instead: https://huggingface.co/datasets/Zarxrax/ModernAnimation1080_v3
This dataset, designed for training Single Image Super Resolution models, includes 2200 full size frames taken from Blu-rays of modern animation across a variety of different art styles. Efforts have been taken to try to ensure that all sources were produced at native 1080p or better (or perhaps, it would be better to say "Full HD").
The only processing that has been done was a conversion to RGB, and for letterboxed sources, cropping. Sol Levante was produced at 4k, and has been downscaled to 1080p to match the other sources.
All frames are at a size of 1920x1080, with the exception of Wolfwalkers and Wish, which have had their letterboxes cropped off.
The files are named according to the source, making it easy to exclude any sources you don't want. It's designed to be modular, just use what you want.
You may wish to do further processing such as generating tiles prior to training. You are also expected to generate any LR images yourself.
All images have been saved as lossless webp to conserve space. My tests show that webp typically decodes just slightly slower than png, so there should not be any significant performance difference if you train from these directly. Of course, feel free to convert to png if you like.
23 additional validation images have also been included, one from each source.
To download, click on "Files and Versions" and download the file "ModernAnimation1080_v2.zip"
- Downloads last month
- 48