The dataset is made of anime screencaps.

Using tags "anime screencap", or just "screencap" might help with matching the likeliness, but will most probably reduce quality by quite a bit.

I'm always open for feedback, so I can improve my work in the future.