BootsofLagrangian/danbooru-202603-webp90-noresize

Danbooru 202603 WebP90 no-resize

This public, manually gated repository contains 10,767,652 verified images in 1,080 WebDataset tar files. Images were retained at their native dimensions and encoded as WebP quality 90. It combines a crawler materialization with the nyanko7/danbooru2023 source surface; crawler records take precedence on overlapping Danbooru post IDs.

This materialization was acquired for research reproducibility because a complete directly usable release was not available for the work that required it.

The frozen source manifest records 6,373,734 crawler images, 4,388,039 source-unique images, and 2,326,707 overlap IDs removed during the merge. It also records 5,927 internally recovered images (0.0550% of the final payload), 783 unavailable source items (0.0073%), 476 deleted posts (0.0044%), and 78 corrupt source items (0.0007%). These states explain why the materialized payload differs from upstream inventories; they are not silently represented as successful images.

Every published tar passed a full-stream SHA-256 audit and exact numeric post-ID member validation. The audit found 29 historical shards with duplicate member names. Twenty-eight contained only byte-identical repeats. One contained 11 alternate WebP encodings, resolved against authoritative source dimensions and decoded-image similarity. Exactly 29/1,080 shards (2.685%) were repacked; the final release has 10,767,652 rows and the same number of distinct post IDs. Authoritative metadata was unavailable for 631 payload rows (0.0059%); their post identity and tar/member bindings remain available. The resolution and repair ledgers are included under audit/.

The metadata Parquet preserves the available authoritative Danbooru metadata and adds asset_instance_id, source page/file URLs, and tar/member bindings. This materialization itself was not produced with img2dataset. For a fresh URL-based acquisition, use the source metadata or Danbooru API and an auditable downloader such as img2dataset.

Loading metadata or images

import os
from datasets import load_dataset

repo = "BootsofLagrangian/danbooru-202603-webp90-noresize"
read_token = os.environ["HF_READ_TOKEN"]

metadata = load_dataset(repo, "metadata", split="train", streaming=True, token=read_token)
metadata_row = next(iter(metadata))
print(metadata_row["source_page_url"], metadata_row["image_shard"], metadata_row["image_member"])

images = load_dataset(
    "webdataset",
    data_files={"train": [f"hf://datasets/{repo}@main/images/*.tar"]},
    split="train",
    streaming=True,
    token=read_token,
    storage_options={"token": read_token},
)
sample = next(iter(images))
image_field = next(name for name in ("webp", "jpg", "jpeg", "png") if name in sample)
image = sample[image_field]
post_id = int(sample["__key__"])  # WebDataset member stems are zero-padded

Danbooru TARs are image-only; convert the zero-padded sample["__key__"] to an integer and join it to metadata post_id/id instead of expecting a JSON sidecar. For rows without authoritative metadata, use source_key as the always-present post identity. Replace main with a commit hash for a frozen run. The metadata config reads the Parquet index; the WebDataset example streams gated image samples.

Each image remains the property of the owner represented by its source URL. Downloading or using an image means acknowledging those third-party rights and accepting responsibility for determining whether the intended research use is permitted. Access and use are limited to research reproducibility. Repository gating does not grant image rights.