Dataset Card for Civistash

Dataset Summary

Civistash is a daily snapshot of the top images and videos from CivitAI, the largest community platform for AI-generated media. Each day it fetches the most-reacted-to content on the platform, downloads the media files, and stores them alongside full metadata sidecars in timestamped daily WebDataset shards.

The dataset is designed as a rolling archive — each .tar.gz shard is a self-contained WebDataset partition of one day's popular content, making it easy for downstream projects to consume a specific date range without processing the entire repository.

Format

This is a WebDataset — every .tar.gz shard contains paired files sharing the same stem (the CivitAI image ID). A sample is the tuple of all files with a given stem:

2026-06-08.tar.gz
├── 12345678.png        # media file
├── 12345678.json       # full CivitAI metadata
├── 98765432.jpg        # media file
├── 98765432.json       # full CivitAI metadata
├── 55555555.mp4        # video file
└── 55555555.json       # full CivitAI metadata

The dataset viewer groups files by stem and decodes them per extension:

Extension	Decoded as
.jpg, .png, .webp	Image (preview)
.mp4	Video
.json	Json (full sidecar, see schema below)

One row in the viewer = one image/video. A shard is one day.

Sidecar JSON schema

Each <id>.json sidecar contains the full CivitAI API response for that item: model info, generation parameters, base model, dimensions, creator username, tags, stats — plus a _civistash provenance block with the download timestamp, source URL, on-disk path, and archive date.

{
  "id": 12345678,
  "url": "https://image.civitai.com/…",
  "type": "image",
  "nsfw": "None",
  "width": 1024,
  "height": 1536,
  "hash": "abc123...",
  "meta": {
    "prompt": "a beautiful landscape...",
    "negativePrompt": "blurry, low quality...",
    "cfgScale": 7,
    "sampler": "Euler a",
    "seed": 1234567890,
    "steps": 20
  },
  "modelVersionId": 98765,
  "modelId": 5432,
  "username": "some_creator",
  "createdAt": "2026-06-08T10:30:00.000Z",
  "stats": {
    "reactionCount": 1420,
    "commentCount": 89,
    "cryCount": 3,
    "likeCount": 1420
  },
  "tags": [
    { "id": 1, "name": "landscape" },
    { "id": 2, "name": "digital painting" }
  ],
  "_civistash": {
    "downloaded_at": "2026-06-08T14:30:00Z",
    "source_url": "https://image.civitai.com/…",
    "stored_as": "2026-06-08/12345678.png",
    "archive_date": "2026-06-08"
  }
}

Supported Tasks

• Text-to-image research — study real-world prompt patterns, CFG scale distributions, sampler preferences, and step counts from a large community of AI image generators.
• Aesthetic analysis — correlate community engagement (reactions, comments) with generation parameters and model choice.
• Trend analysis — track which models, styles, and tags dominate the CivitAI platform over time.
• Dataset augmentation — use metadata (prompts, tags) as weak captions for image-captioning or CLIP-style training.
• Model benchmarking — compare outputs of different base models and fine-tunes against community-voted favorites.

Languages

Metadata and prompts are primarily in English. Tags use a controlled vocabulary from the CivitAI platform. Some prompts may contain fragments of other languages (Japanese, Chinese, etc.) when creators use multilingual descriptions.

Dataset Structure

Data Splits

There is no train/validation/test split — this is a raw archive. One shard per day, named YYYY-MM-DD.tar.gz.

Split	Description
default (train)	Every image from every daily shard, ordered by date

Data Fields

The sidecar JSON is exposed as the json column. The media file is exposed as jpg / png / webp / mp4 depending on its type. Other fields:

Sidecar key	Type	Description
id	integer	CivitAI image ID
url	string	Direct media URL on CivitAI CDN
type	string	image or video
nsfw	string	NSFW level: None, Soft, Mature, X
width	integer	Image width in pixels
height	integer	Image height in pixels
hash	string	Perceptual hash
meta	object	Generation parameters (prompt, negative prompt, CFG scale, sampler, seed, steps, etc.)
modelVersionId	integer	Specific model version used
modelId	integer	Base model ID
username	string	CivitAI creator username
createdAt	datetime	When the image was posted
stats	object	Reaction count, comment count, cry count, like count
tags	array	Tag objects with id and name
_civistash.downloaded_at	datetime	Civitash fetch timestamp
_civistash.source_url	string	Same as url, kept for provenance
_civistash.stored_as	string	Local on-disk path (relative to stash root)
_civistash.archive_date	string	YYYY-MM-DD — which daily shard this lives in

Dataset Creation

Curation Rationale

CivitAI is the largest public repository of AI-generated images, with millions of uploads and an active community voting system. However, it has no official bulk-export or historical snapshot API. This dataset fills that gap by providing a scheduled, reproducible archive of the platform's most popular daily content.

Source Data

All data originates from the CivitAI public API (GET /api/v1/images). Media files are downloaded from the CivitAI CDN. No scraping of the website HTML is performed.

Collection Process

1. Query the API for the top images of the current period (day, week, month, or all-time), sorted by most reactions.
2. Skip any images already present in the local archive (deduplication by ID across all date partitions).
3. Download each image sequentially with retry backoff (1s/2s/4s) on rate limits and transport errors.
4. Write a JSON sidecar containing the full API response plus a _civistash provenance block.
5. Bundle the day's partition into a .tar.gz WebDataset shard (file pairs at the tarball root, grouped by CivitAI image ID) and upload to this Hugging Face dataset repository.

Annotations

The metadata fields (meta, tags, stats, modelVersionId, etc.) are provided directly by the CivitAI API and are not annotated by the Civistash tool. The _civistash provenance block is the only addition.

Personal and Sensitive Information

CivitAI usernames are public by design. No private user data or authentication tokens are included in the archive. Media flagged as NSFW may be present depending on the archive configuration — the nsfw field in each sidecar allows downstream consumers to filter content.

Considerations for Using the Data

Biases

The dataset reflects the popularity bias of the CivitAI platform: only the most-reacted-to images are included, which skews toward content that engages the platform's userbase. Model representation is biased toward popular base models (Stable Diffusion variants, Flux, etc.). This is not a random sample of AI-generated content — it is explicitly a popularity-ranked snapshot.

Licensing

The media files and metadata in this dataset are sourced from CivitAI and are subject to the CivitAI Terms of Service. Individual images may carry additional licenses set by their creators. Consumers of this dataset are responsible for complying with all applicable licenses and terms.

How to self-host / run the archiver

This dataset is produced by Civistash, an open-source Rust CLI tool. You can run your own instance to archive different periods, sort orders, NSFW levels, or upload to your own Hugging Face repo.

Source code: github.com/Hyphonical/Civistash

# One-shot: fetch and bundle today's top 200 images
civistash --period Day --limit 200 --bundle

# Daemon: run continuously with daily cycles, auto-upload to HF
civistash --daemon --period Day --limit 1000 --bundle --upload-hf your-org/your-dataset

# Docker
echo "CIVITAI_TOKEN=eyJ…" > .env
echo "HUGGINGFACE_TOKEN=hf_…" >> .env
docker compose up -d

Full documentation is available in the project README.

Additional Information

Dataset Curators

This dataset is maintained by Hyphonical using the automated Civistash archiver.

Licensing Information

Media and metadata sourced from CivitAI. Refer to the CivitAI Terms of Service and individual content licenses for usage terms.

The Civistash tool itself is licensed under the MIT License.

Citation

If you use this dataset in research, please cite:

@misc{civistash2026,
  author = {Hyphonical},
  title = {Civistash: Daily Top CivitAI Images Archive},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Hyphonical/Civistash}},
  note = {Archived with the Civistash tool: \url{https://github.com/Hyphonical/Civistash}}
}

Contributions

Archiving is fully automated. For issues or feature requests, please open an issue on the GitHub repository.