data-archetype/pexels_hq_diverse_40k_2048

Pexels HQ Diverse 40k 2048

40,000 Pexels photographs selected for image quality and visual diversity, exported into 29 aspect-ratio buckets at a base resolution of 2048 pixels. Every image has two English captions: a natural-language description and a JSON layout of the image with bounding boxes.

Contents

• 40,000 images; 53.7 GB across 61 TAR shards.
• Each sample contains a JPEG image, its natural caption (.txt, UTF-8) and JSON metadata holding the layout caption, keyed by the original image's SHA256 hash.
• manifest.json lists bucket dimensions, shard counts, checksums and the metadata captions (sample_captions).
• source_metadata.csv provides Pexels photo IDs, photographers and selection scores.

Bucket names use 1024-base dimensions: p1216x832, for example, contains 2432 × 1664 images. Actual image dimensions are listed in the manifest.

Captions

• natural — a dense description of the photograph (subject, composition, setting, lighting, colours, visible text quoted), stored as the sample's .txt.
• layout — a JSON layout in the sample's .json under captions.layout (format: layout_json), generated with Gemma 4 31B: {"summary": {"meta": [...], "canvas": "single|multi_panel|collage", "caption": ...}, "frame": {"caption", "tags", "palette"}, "items": [{"label": "character|object|text|panel", "caption", "text"?, "tags", "palette", "box_2d"}]}. summary describes the whole image, frame the photograph as a composition, and items the people, animals, objects, regions and legible text an image generator would need to place, most important first (at most 16). Text items hold the visible text exactly as written. box_2d = [y_min, x_min, y_max, x_max], normalized to 0–1000 over the stored image.

Selection and processing

DINOv3-S embeddings were used to group quality-qualified candidates into 2,048 visual clusters. The original 20,000 images are retained, with 20,000 additional photographs selected against the combined set. Sampling quotas were proportional to the square root of each cluster's size, giving larger clusters more images while increasing the relative representation of smaller clusters. Selection within clusters considered TOPIQ-NR quality scores and visual diversity, with global duplicate filtering using embedding similarity and perceptual hashes. The dataset includes 15,799 photographers, with at most 20 images per photographer.

Images are resized with antialiased bicubic interpolation and center-cropped to their target aspect ratio, without upscaling. Cropping removes at most 5% of the source area. Images are encoded as JPEG at quality 95.

Loading

import json, tarfile
from pathlib import Path

for tar_path in sorted(Path("buckets").rglob("shard-*.tar")):
    with tarfile.open(tar_path) as tf:
        for m in tf.getmembers():
            if m.name.endswith(".json"):
                key = m.name[:-5]
                meta = json.loads(tf.extractfile(m).read())
                natural = tf.extractfile(key + ".txt").read().decode()
                layout = json.loads(meta["captions"]["layout"]["text"])
                jpg = tf.extractfile(key + ".jpg").read()

License

Photographs are subject to the Pexels License and Terms of Service. This dataset is not affiliated with or endorsed by Pexels.