SwayStar123/laion_aesthetic_5.5_HQ

~10 million images from LAION Aesthetics v2, with aesthetic score > 5.5, probability of watermark < 0.95, and image megapixels > 1.05

Gathered using the following script

from datasets import load_dataset
import os

def filt(row):
    w = row["WIDTH"]
    h = row["HEIGHT"]
    p = row["pwatermark"]
    a = row["AESTHETIC_SCORE"]

    if w is None or h is None or p is None:
        return False

    return (w * h >= 1024*1024) and p <= 0.95 and a > 5.5


if __name__ == "__main__":
    os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
    ds = load_dataset("laion/aesthetics_v2_4.75", split="train", num_proc=48, cache_dir="data")

    new_ds = ds.filter(lambda x: filt(x), num_proc=48)

    print("New length: ", len(new_ds))

    new_ds.save_to_disk("filtered")
    new_ds.push_to_hub("SwayStar123/laion_aesthetic_5.5_HQ")