btan2/pixmo_images

PixMo Images

The raw images backing the PixMo datasets used to train Molmo, packaged as Parquet shards with embedded image bytes so they can be browsed in the dataset viewer and loaded directly with datasets.

The PixMo annotation datasets (allenai/pixmo-*) ship image_urls rather than image bytes. This repository is a content cache of those images, keyed by the SHA-256 of the source URL.

Contents

• 1,073,189 images across 525 Parquet shards (data/train-*.parquet), ~266 GB total.
• Image bytes are stored verbatim (no re-encoding). Format breakdown (by magic bytes): JPEG 904,695 · PNG 160,046 · WebP 7,676 · GIF 1,073 · BMP 22 · TIFF 1 · unrecognized 1.
• Every image_id is unique; the set exactly covers the source cache.

Schema

column	type	description
image_id	string	The image identifier: sha256(image_url) (lowercase hex, 64 chars). This is the on-disk cache key in Molmo's pixmo_images store and the join key to the allenai/pixmo-* datasets. Note: it is the hash of the source URL, not of the stored image bytes.
ext	string	Detected image format from magic bytes (jpg, png, webp, gif, ...).
image	Image	The raw image bytes (HF datasets Image feature).

Usage

from datasets import load_dataset

ds = load_dataset("UWGZQ/pixmo_images", split="train")  # or streaming=True
ex = ds[0]
ex["image"].save("example.jpg")
print(ex["image_id"], ex["ext"], ex["image"].size)

Joining with a PixMo annotation dataset

import hashlib
from datasets import load_dataset

def image_id(url: str) -> str:
    return hashlib.sha256(url.encode()).hexdigest()

points = load_dataset("allenai/pixmo-points", split="train")
images = load_dataset("UWGZQ/pixmo_images", split="train")

# build an index from image_id -> row, then look up by sha256(image_url)
by_id = {r["image_id"]: i for i, r in enumerate(images)}
row = points[0]
img = images[by_id[image_id(row["image_url"])]]["image"]

Provenance & license

These images were collected from public web URLs for the PixMo/Molmo project. Each image retains the copyright and license of its original source; they are redistributed here as a research convenience cache. If you are a rights holder and want an image removed, please open a discussion on this repository. The image_id (URL hash) and the metadata in the allenai/pixmo-* datasets let you trace any image back to its source URL.

How this was built

Images were read read-only from Molmo's on-disk pixmo_images cache (files named by sha256(image_url), no extension) and packed verbatim into Parquet shards of ~2048 images each, with the image column written as the HF datasets Image feature (byte-faithful; no decoding/re-encoding). The sha256(image_url) keying was verified against allenai/pixmo-points (40/40 sampled URLs resolved to on-disk files).