Felix Midjourney Archive

A deduplicated, checksum-addressed preservation dataset of AI-generated images created by Felix / waffles13 with Midjourney. Images are stored in deterministic WebDataset TAR shards with searchable Parquet, JSONL, CSV, and SQLite catalogs.

Contents

• Unique images: 32,477
• Exact duplicate source copies excluded: 12,560
• Images with full embedded prompts: 13,914
• Unique Midjourney Job IDs represented: 22,859
• Total image bytes before TAR overhead: 46.83 GiB

Each sample uses a stable content-derived ID (mj- plus the first 24 hexadecimal characters of its SHA-256). Every image has a same-key .json sidecar inside its WebDataset shard. metadata.parquet is the primary fast-search catalog.

Provenance

The sources are local Midjourney downloads, historical Midjourney ZIP exports, and older organized folders belonging to the creator. Exact duplicates are identified by full-file SHA-256, not by filenames or perceptual similarity. Distinct encodings, crops, resolutions, variants, and upscales remain distinct.

Prompts are recovered in this order:

1. embedded PNG Description metadata;
2. creator-maintained catalog metadata;
3. cleaned historical filename text.

The catalog records prompt_source so users can distinguish complete embedded prompts from potentially truncated filename-derived prompts. Original filenames and every duplicate source alias are retained in metadata; published filenames are cleaned and stable.

Metadata

Important fields include image_id, sha256, prompt, prompt_source, created_at, midjourney_job_id, variant_index, model_version, parameters, tags, dimensions, rating fields, source aliases, and a normalized search_text field.

Tags are deterministic prompt-derived retrieval terms plus structural tags such as orientation, resolution, and Midjourney version. They are search aids, not verified depictions or safety labels.

AI-generation disclosure

The images were generated with Midjourney. They are synthetic media and may contain visual errors, biased representations, unsettling imagery, imitations of artistic conventions, text artifacts, or content that resembles real people, brands, copyrighted characters, or protected designs.

License

The creator dedicates their rights in this dataset and its AI-generated images to the public domain under CC0 1.0 Universal. Anyone may copy, modify, distribute, and use the material for any purpose, including commercially, without asking permission.

CC0 can only waive rights held by the person making the dedication. It does not grant rights owned by third parties, including trademark, publicity/privacy, or rights in recognizable pre-existing works. Users are responsible for evaluating their own use.

Limitations

• Filename-derived prompts may be shortened by historical filename limits.
• Midjourney Job IDs describe generation lineage but do not guarantee that every historical account job is present locally.
• Exact-file deduplication intentionally preserves visually identical images when their bytes differ.
• Automated prompt tags can be incomplete, ambiguous, or inaccurate.
• No claim is made that the dataset is suitable for safety-critical use or that every image is appropriate for every audience.

Integrity and mirrors

SHA256SUMS covers catalogs and WebDataset shards. The same stable image IDs and shard checksums are used by the Hugging Face dataset and Internet Archive volumes. Internet Archive item identifiers and shard-to-volume mappings are listed in internet_archive_volumes.json.