dingshizhe/mid_512

Midjourney 23M — 512×512

This repository contains a 512×512 JPEG conversion of the VGT Midjourney dataset. The complete conversion contains 23,167,456 images in 5,000 logical shards.

Image conversion

Each source image was processed as follows:

1. Apply the EXIF orientation.
2. Convert to RGB.
3. Center-crop the short side to obtain a square (no aspect-ratio stretching).
4. Resize to 512×512 with Lanczos resampling.
5. Encode as JPEG quality 85 with 4:2:0 chroma subsampling.

File layout

Images are packaged into uncompressed tar archives because the full dataset contains more than 23 million individual files. JPEG data is already compressed, so applying gzip would add substantial processing time with little space saving.

data/{group}/{shard}.tar
manifests/{group}/{shard}.jsonl

Each archive normally contains 5,000 JPEG files. The corresponding JSONL manifest records the source filename, converted image path, width, and height.

The dataset was processed on two nodes over disjoint shard ranges. Uploads are also performed over those same disjoint ranges so paths do not overlap.