data-archetype/LAION_Aesthetics_1024_bucketed_1024

LAION Aesthetics 1024 Bucketed 1024 Captioned

This is a captioned bucketed-shards export of images from limingcv/LAION_Aesthetics_1024.

Images were filtered and resized/cropped into SDXL-style aspect-ratio buckets, with a 1024-first bucket waterfall and 512 fallback for images that did not fit a 1024 bucket without upsampling. The export contains 381,954 images across 398 uncompressed WebDataset-style tar shards.

The .txt files now contain model-generated captions, not the original LAION web-scrape alt text or surrounding page text. Captions were generated from the bucketed images using OpenRouter models in this priority order:

1. google/gemini-2.5-flash-lite
2. google/gemini-2.0-flash-001
3. mistralai/mistral-medium-3.1

Each sample's .json metadata records caption_variant, caption_selector_index, and caption_source_id. manifest.json records the caption source table and prompt hashes.

One placeholder image containing "IMAGE TEMPORARILY UNAVAILABLE" was removed during caption QA.

Technical details

This repository contains a bucketed-shards export (uncompressed TAR shards).

Format

• Format: bucketed_shards_v1
• Created: 2026-02-25T02:33:00.262325+00:00
• Export ID: fcb3cebdc07a498b
• Manifest: manifest.json
• Image mode: reencode_jpeg

Directory layout:

• manifest.json (global metadata + per-bucket shard listing)
• buckets/<bucket_id>/shard-*.tar

Each TAR shard contains 3 files per sample:

• <key>.jpg (JPEG bytes; either re-encoded RGB JPEG or source JPEG passthrough depending on image_mode)
• <key>.txt (caption text, UTF-8, newline-terminated)
• <key>.json (per-sample metadata: w, h, jpeg, image_mode, caption_variant, caption_selector_index, caption_source_id)

Image preprocessing

If image_mode=reencode_jpeg, images are processed deterministically per-sample:

• EXIF transpose, convert to RGB
• Cover-resize using torch CPU bicubic interpolation with antialiasing (mode=bicubic, antialias=True)
• Never upsample: samples that would require upscaling are skipped (too_small_policy=drop)
• Crop to the bucket target size (crop_strategy=corner, allowed corners [2, 3])

Resize/crop details:

• Cover scale is scale = max(target_w / src_w, target_h / src_h); if scale > 1, the sample is skipped.
• After resize, a crop box is chosen deterministically from the sample key (sha256 of image_id).
  • Corner strategy chooses a corner from allowed_corners where 0=TL, 1=TR, 2=BL, 3=BR (optional small jitter for corner_jitter).

JPEG encoding:

• quality 95
• subsampling policy adaptive_scale (adaptive threshold 0.85)

If image_mode=passthrough_jpeg, the exporter stores the source file bytes as-is (no EXIF transpose / resize / crop / re-encode). Buckets and w/h metadata still refer to the planned target size for that bucket (not necessarily the encoded JPEG dimensions). Loaders should decode the JPEG bytes, apply EXIF orientation if desired, then do resize/crop at load time.

Buckets / resolutions

• Buckets follow SDXL-style proto buckets defined at a 1024×1024 base.
• Base resolution(s): [1024, 512]
• In single-res exports, bucket_id is the proto (1024-base) bucket, e.g. p1024x1024.
• In multi-res exports, buckets are namespaced by base resolution: r<base>_<proto>, e.g. r512_p1024x1024.
• The actual target resolution for each bucket (scaled by the per-bucket base resolution and divisible=32) is stored in:
  • manifest.json → buckets[<bucket_id>].scaled.w/h (and base_resolution)
  • each sample’s <key>.json → w/h

Bucket IDs (preview): r1024_p1024x1024, r1024_p1024x960, r1024_p1088x896, r1024_p1088x960, r1024_p1152x832, r1024_p1152x896, r1024_p1216x832, r1024_p1280x768, r1024_p1344x704, r1024_p1344x768, r1024_p1408x704, r1024_p1472x704, r1024_p1536x640, r1024_p1600x640, r1024_p1664x576, r1024_p1728x576, r1024_p1792x576, r1024_p1856x512, r1024_p1920x512, r1024_p1984x512, … (+20 more)

Bucket distribution:

bucket_id	target_w×h	aspect	count
r1024_p1216x832	1216×832	1.462	98,867
r1024_p1344x768	1344×768	1.750	45,275
r1024_p1024x1024	1024×1024	1.000	42,714
r1024_p832x1216	832×1216	0.684	39,685
r1024_p1152x832	1152×832	1.385	32,157
r1024_p1280x768	1280×768	1.667	26,323
r1024_p832x1152	832×1152	0.722	19,021
r1024_p896x1088	896×1088	0.824	14,921
r1024_p896x1152	896×1152	0.778	13,322
r1024_p1152x896	1152×896	1.286	12,073
r1024_p1088x896	1088×896	1.214	9,649
r1024_p960x1088	960×1088	0.882	4,015
r1024_p1088x960	1088×960	1.133	3,924
r1024_p1344x704	1344×704	1.909	3,230
r1024_p960x1024	960×1024	0.938	2,964
r1024_p1024x960	1024×960	1.067	2,855
r1024_p768x1344	768×1344	0.571	2,846
r1024_p768x1280	768×1280	0.600	2,414
r1024_p1408x704	1408×704	2.000	2,047
r1024_p1472x704	1472×704	2.091	1,009
r1024_p1536x640	1536×640	2.400	981
r1024_p704x1472	704×1472	0.478	578
r1024_p704x1408	704×1408	0.500	347
r1024_p1600x640	1600×640	2.500	308
r1024_p1664x576	1664×576	2.889	106
r1024_p1792x576	1792×576	3.111	80
r1024_p1856x512	1856×512	3.625	63
r1024_p640x1536	640×1536	0.417	39
r1024_p1728x576	1728×576	3.000	36
r1024_p576x1664	576×1664	0.346	24
r1024_p640x1600	640×1600	0.400	21
r1024_p2048x512	2048×512	4.000	20
r1024_p512x2048	512×2048	0.250	14
r1024_p576x1728	576×1728	0.333	8
r1024_p576x1792	576×1792	0.321	6
r1024_p1984x512	1984×512	3.875	4
r1024_p1920x512	1920×512	3.750	2
r1024_p512x1856	512×1856	0.276	2
r1024_p512x1920	512×1920	0.267	2
r1024_p512x1984	512×1984	0.258	2

Caption selection (waterfall)

Captions are selected from dataset.sqlite using the first matching selector (highest priority wins). Within the same selector, the newest caption source is preferred.

Caption provenance:

• Per-sample <key>.json includes caption_source_id (int, from dataset.sqlite).
• manifest.json includes a caption_sources table mapping caption_source_id → backend/model/created_at plus prompt hashes (not prompt text).

Caption sources used:

caption_source_id	backend	model	created_at	system_prompt_sha256	user_prompt_sha256
1	openrouter	google/gemini-2.5-flash-lite	1776529602	15d506325e85…	8a9eef48b245…
2	openrouter	google/gemini-2.0-flash-001	1776531880	15d506325e85…	8a9eef48b245…
3	openrouter	mistralai/mistral-medium-3.1	1776550454	15d506325e85…	8a9eef48b245…

Caption priority (waterfall) + planned usage:

selector_index	variant	backend	model	planned_images
0	caption_gemini_2_5_flash_lite	openrouter	google/gemini-2.5-flash-lite	32,942
1	caption_gemini_2_flash	openrouter	google/gemini-2.0-flash-001	348,865
2	caption_mistral_medium_3_1	openrouter	mistralai/mistral-medium-3.1	147

Available caption variants (top 30):

selected	variant	images_with_ok_caption
✓	caption_gemini_2_flash	349,060
✓	caption_gemini_2_5_flash_lite	32,942
✓	caption_mistral_medium_3_1	147

Missing caption policy: empty

Export summary

• images_seen: 381,954
• images_exported: 381,954
• skipped_no_caption: 0
• skipped_too_small: 0
• decode_errors: 0
• encode_errors: 0

Efficient loading

Recommended

Treat this as a webdataset-style collection of tar shards:

• Prefer sequential reads of tar files for throughput.
• Shuffle at the shard level (and optionally within-shard) for good randomness without expensive random I/O.
• Use manifest.json to list buckets and shards.

Python (webdataset)

import webdataset as wds

urls = "buckets/*/shard-*.tar"  # glob; adjust if you want a single bucket only
ds = (
    wds.WebDataset(urls)
    .decode("pil")            # decodes .jpg to PIL.Image
    .to_tuple("jpg", "txt", "json")
)
for jpg, caption, meta in ds:
    ...

Python (tarfile, no extra deps)

import io, json, tarfile
from pathlib import Path

tar_path = next(Path("buckets").rglob("shard-*.tar"))
with tarfile.open(tar_path, "r") as tf:
    members = tf.getmembers()
    for m in members:
        if not m.name.endswith(".txt"):
            continue
        key = m.name[:-4]
        caption = tf.extractfile(m).read().decode("utf-8").strip()
        meta = json.loads(tf.extractfile(tf.getmember(key + ".json")).read().decode("utf-8"))
        jpg_bytes = tf.extractfile(tf.getmember(key + ".jpg")).read()
        ...