This dataset contains sexually explicit (NSFW) AI-generated images and prompts. It is intended only for research on NSFW detection / content guarding and AI-generated-image detection. By requesting access you confirm you are an adult and will not redistribute the content or use it to train generative models.

AIGC Dataset — NovelAI Generations

⚠️ Contains NSFW content. Images and prompts have not been filtered. Many are sexually explicit.

1,562,685 AI-generated images (≈332 GB) made with NovelAI's image models (V3 → V4 → V4.5 → V5) between 2023-11 and 2026-09 through a private Discord bot (Kohaku-NAI). Every image has its full generation metadata (model, sampler, steps, CFG, seed, prompt, …) and an exact generation timestamp. That makes the set usable for:

AI-generated image detection. Every sample is synthetic, and the model and revision are known, so you can study per-generator and cross-generator generalization and drift over time.
NSFW detection / content guard. It covers real-world user prompts for an anime-style generator, with prompts included.

Structure

nai/
  <model>_<rev>/                           e.g. nai-diffusion-4-5-full_4BDE2A90/
    NNNNN_<first-day>_<last-day>.tar       uncompressed tar, <= 5 GB, samples in time order
    NNNNN_<first-day>_<last-day>.idx.json  random-access byte offsets for the tar
    NNNNN_<first-day>_<last-day>.parquet   per-shard metadata table
  manifest.parquet                         all shards, one row per sample
  stats.json                               sample / byte / month counts per shard
  README.md                                technical notes
  tools/                                   scan.py / pack.py / verify.py / model_map.json

rev is the model hash from NovelAI's Source tag. The same API model name can have several revisions.

Each sample <key> is stored as two tar members (WebDataset convention):

member	content
`<key>.png` / `<key>.webp`	original file bytes, unmodified; the NovelAI metadata is still embedded (PNG tEXt / WebP EXIF UserComment)
`<key>.json`	the same metadata, stored separately (see below)

key = <snowflake>_<sha3-256[:8]>, where snowflake >> 22 is the unix-ms generation time.

Metadata (`<key>.json`)

field	description
`model`, `model_rev`	API model name + NovelAI model hash
`model_rev_inferred`	`true` when the rev was inferred from the date (no `Source` tag; 128,088 samples from 2024-01/02)
`source`	raw NovelAI `Source` string
`api_model`	requested model name (from request payload, when present)
`timestamp_ms`, `datetime`	generation time (UTC); `null` for the 2,712 `undated` samples
`ext`, `width`, `height`, `size`, `sha256`	file info
`meta_kind`	`png` (tEXt) / `exif-nai` (NovelAI fields in WebP EXIF) / `exif-api` (request payload in WebP EXIF)
`embedded`	parsed embedded metadata (`Comment` = generation parameters incl. prompt / negative prompt)
`sidecar`	separate request JSON, when the bot saved one (9,911 samples)
`collection`, `n_duplicates`	internal source bucket, number of dropped byte-identical copies

manifest.parquet holds all of these plus shard, img_offset, img_size, json_offset, json_size, and flattened metadata columns emb.* / side.* (e.g. emb.comment.sampler, emb.comment.steps, emb.comment.v4_prompt.caption.base_caption). NovelAI's metadata schema changed over time, so the manifest's columns are the union across all samples. Fields a sample doesn't have are null, and columns whose type changed over time are stored as strings.

Statistics

model	rev	samples	size	date range	format
nai-diffusion-3	C1E1DE52	885,383	256.5 GB	2023-11-25 → 2024-09-18 (+2,712 undated)	140,004 png / 745,379 webp
nai-diffusion-3	7BCCAA2C	534,935	51.3 GB	2024-09-18 → 2026-09-22	webp
nai-diffusion-3	None	68	0.03 GB	2023-12-26 → 2024-09-13	png / webp
nai-diffusion-4-curated-preview	F6E18726, 4F49EC75, 79F47848, C1CCBA86	13,664	1.1 GB	2025-02-22 → 2025-04-11	webp
nai-diffusion-4-full	37442FCA	929	0.08 GB	2025-04-16 → 2026-09-11	webp
nai-diffusion-4-5-full	4BDE2A90	124,642	14.0 GB	2025-07-04 → 2026-09-25	webp
nai-diffusion-5-full	0ADF9AB7	1,471	0.17 GB	2026-08-21 → 2026-09-26	webp
nai-diffusion-4-5-curated / 5-curated	C02D4F98 / DB276663	1 / 3	–	2026-08 → 2026-09	webp
unmapped (V4.5 B9F340FD, V4 7ABFFA2A, SD F1022D28, SDXL 9CC2F394)		1,589	0.18 GB	2024-04 → 2025-07	webp

Formats: 1,422,659 WebP and 140,026 PNG.
Resolutions: mostly NovelAI's standard sizes: 832×1216 (54%), 1216×832 (18%), 1024×1024 (17%).
Duplicates: deduplicated by sha256 over the whole collection (none found).

Loading

Random access, without extracting anything:

import json, pyarrow.parquet as pq

root = "nai/"
m = pq.read_table(root + "manifest.parquet",
                  columns=["key", "model", "shard", "img_offset", "img_size", "json_offset", "json_size"])
r = m.slice(0, 1).to_pylist()[0]
with open(root + r["shard"], "rb") as f:
    f.seek(r["img_offset"]); img_bytes = f.read(r["img_size"])
    f.seek(r["json_offset"]); meta = json.loads(f.read(r["json_size"]))

Streaming with webdataset:

import webdataset as wds
ds = wds.WebDataset("nai/nai-diffusion-4-5-full_4BDE2A90/{00000..00003}_*.tar")  # or list shards from manifest
for s in ds:
    img = s.get("webp") or s.get("png"); meta = json.loads(s["json"])

Caveats

WebP files are lossy re-encodes. The bot converted most outputs to WebP (quality 75), and only 140k early V3 samples are original PNGs. For AIGC detection, format and compression are a strong shortcut. Control for them, e.g. re-encode real and fake images the same way, or use the PNG subset.
Embedded metadata. The image bytes still contain the generation metadata. Detectors should decode pixels only and never read the raw file bytes or EXIF.
Single generator family. Everything comes from NovelAI's anime-style models, and there are no real images. Pair the set with a real-image corpus for detection training.
No NSFW labels. The dataset has no human NSFW annotations. Prompts and parameters are available as weak signals.
Model mapping. model names come from the bot's request logs and the dates each model was available (see nai/tools/model_map.json). The *-unmapped groups and the tiny *-curated groups are uncertain.
Privacy. User identifiers (Discord IDs, usernames, per-user folders) were removed. Prompts are user-written and may contain arbitrary text.

License

CC BY 4.0.

Under the NovelAI Terms of Service (effective 2025-12-08), generated images ("Content" — "the generated material that you specifically create using our Services") belong to the user: "You … retain all rights and ownership of your Content" and "We do not claim any ownership rights to your Content." All images in this dataset were generated on the dataset author's own NovelAI account (the bot ran entirely on it), so the author holds these rights and releases them under CC BY 4.0. The ToS places no restriction on redistributing outputs or using them for training. Note that AI-generated images may not be eligible for copyright protection in every jurisdiction.

Downloads last month: 23

Total file size:

332 GB

You need to agree to share your Hugging Face information to access this dataset