LAION

LAION is a family of open image–text datasets built by LAION (Large-scale Artificial Intelligence Open Network), a non-profit association registered in Germany. Each dataset is a list of image URLs found in the Common Crawl web archive, each paired with the alt-text that accompanied the image on its web page and kept only if a CLIP model judged image and text similar enough. LAION-400M was announced in August 2021, and LAION-5B, with 5.85 billion pairs, in March 2022; the LAION-5B paper by Christoph Schuhmann and colleagues appeared in the NeurIPS 2022 Datasets and Benchmarks Track. LAION distributes metadata, embeddings and nearest-neighbour indices, not images, so every user downloads the pictures from the original sites. The datasets made web-scale contrastive pretraining and text-to-image training reproducible outside industrial labs: OpenCLIP models and Stable Diffusion were trained on them. They have also been audited for misogynistic, pornographic and hateful content, and in December 2023 LAION withdrew LAION-5B after a Stanford Internet Observatory report found links to suspected child sexual abuse material (CSAM) in it. A revised version, Re-LAION-5B, followed in August 2024.

Purpose

CLIP (trained on 400 million pairs) and DALL·E had shown in early 2021 what models trained on web-scale image–text pairs could do, but their training sets were not released. The LAION-400M paper states its aim as filling that gap with a public dataset large enough to train such models from scratch (Schuhmann et al., 2021, §1). The LAION-5B paper repeats the argument at billion scale: ALIGN, BASIC and other successors had been trained on private datasets of up to 6.6 billion pairs, which concentrated the research in a few industrial labs (Schuhmann et al., 2022, §1, Table 1). An open dataset also makes the data itself auditable, a point the paper makes in its safety discussion (§7).

Both releases are presented as research artefacts. The LAION-5B paper recommends academic research use only and advises against deployed systems built on models trained on it without study of their behaviour and biases (§1, §7), and its datasheet says that models should not be put into production using the dataset as it is (arXiv v1, Appendix A, Q43).

Versions

Version	Released	Pairs	Filter	Source
LAION-400M	20 August 2021	413M unique samples	OpenAI CLIP ViT-B/32, cosine $\geq 0.3$, English	LAION blog (2021); Schuhmann et al. (2021), Table 1
LAION-5B	31 March 2022	5.85 billion	CLIP ViT-B/32, $\geq 0.28$ (English), multilingual CLIP ViT-B/32, $\geq 0.26$ (other languages and none detected)	LAION blog (2022); Schuhmann et al. (2022), §3.1, §4
Re-LAION-5B research	30 August 2024	LAION-5B minus the removals below	LAION-5B’s, plus link and hash removal and an unsafe-score filter	LAION blog (2024)
Re-LAION-5B research-safe	30 August 2024	a subset of research	as research, with a stricter unsafe-score threshold	LAION blog (2024)

LAION-400M was built from Common Crawl pages crawled between 2014 and 2021 (LAION blog, 2021); the LAION-5B datasheet says the data were filtered from September 2021 to January 2022 (Q26). The LAION-5B paper calls LAION-400M a previously released subset of LAION-5B (§1, Figure 1). The Re-LAION-5B announcement says instead that LAION-5B was assembled from Common Crawl data up to September 2022 and in summer 2022, dates later than the dataset’s release on 31 March 2022.

LAION also published subsets of LAION-5B selected by other scores: LAION-High-Resolution (170 million pairs, for super-resolution models) and LAION-Aesthetics (Schuhmann et al., 2022, §5.1). LAION-Aesthetics V1 took the 8 million and 120 million English-caption images with the highest scores from a linear model on CLIP ViT-L/14 image embeddings trained on 5,000 human ratings. V2 used a linear predictor trained on 176,000 ratings from the Simulacra Aesthetic Captions dataset, 15,000 rated logos and the AVA aesthetics photographs, and kept nested subsets of English-caption pairs from 1.2 billion (score at least 4.5) down to 625,000 (at least 6.5) (LAION-Aesthetics blog, 2022).

Construction

The pipeline is the same in outline for both versions (Schuhmann et al., 2021, §2; 2022, §3.1):

1. Parsing. Workers read Common Crawl’s WAT files, which hold the metadata of crawled pages, and extracted every HTML IMG tag that had an alt attribute. For LAION-5B the alt-text was passed through the CLD3 language detector, with three outcomes: English, another language, or no language detected with confidence.
2. Downloading. The images were fetched from their URLs with asynchronous requests; for LAION-5B about 300 small cloud workers each processed batches of 10,000 links.
3. Basic filters. Pairs with alt-text shorter than 5 characters or images smaller than 5 KB were dropped; for LAION-5B, so were potentially malicious or very large images (§3.1).
4. Deduplication. For LAION-400M, Bloom filters on the concatenated URL and alt-text dropped repeats during the crawl, and a later PySpark job deduplicated on URL and caption (LAION blog, 2021). The LAION-5B announcement says duplicates were removed with a Bloom filter on the URL (LAION blog, 2022); the paper says only that “redundant images” were removed (§3.1). Neither compares image content, so the same picture at two URLs stays in twice.
5. CLIP filtering. Image and alt-text were embedded with CLIP ViT-B/32 and the pair was dropped if the cosine similarity of the two embeddings was below the threshold. For LAION-400M the threshold was 0.3, chosen by human inspection; the announcement says 0.28 and 0.29 also looked acceptable but 0.3 was kept as the conservative value (LAION blog, 2021). For LAION-5B it was 0.28 for English pairs (OpenAI’s CLIP ViT-B/32) and 0.26 for the others (a multilingual CLIP ViT-B/32), and this step removed about 90% of some 50 billion candidates (Schuhmann et al., 2022, §3.1). The paper notes that larger CLIP models were not available when the work began and that a ViT-L/14 filter would probably have produced a less noisy dataset (§6).
6. Content filtering. For LAION-400M the announcement describes the steps taken against illegal content: images were tagged NSFW, UNSURE or UNLIKELY by comparing their CLIP embeddings with category prompts, and tagged samples were dropped if their alt-text contained words related to minors or their image embedding was close to categories describing minors (LAION blog, 2021, filtering steps 6 to 9). The LAION-5B paper says CLIP embeddings were used to filter illegal content but does not describe how (§3.2). Other potentially offensive samples were tagged rather than removed, a choice the paper defends as leaving the decision to users and enabling research on curation (§3.2).

After filtering, LAION computed CLIP ViT-L/14 image embeddings for LAION-5B, built nearest-neighbour indices over them with autofaiss, and derived the safety and watermark tags from the embeddings (arXiv v1, Appendices B and C).

The similarity threshold

The CLIP filter is a binary classifier on one score, and its threshold trades the precision of the kept set against the recall of good pairs. The example below makes the trade-off visible with pairs whose correctness is known. Each image is a rendered coloured shape on a grey background, sometimes faded (its colour pulled toward grey, like a washed-out photograph) and sometimes with a second, smaller shape as clutter. Each caption is either exact (“a red circle”), padded with search-engine words (“red circle free download hd png”), wrong in colour or in shape, or unrelated spam. Two hand-built encoders stand in for CLIP’s: the image encoder segments the largest region and scores its colour and fill ratio, with its colour evidence weakened when contrast is low and Gaussian noise added for encoder error; the text encoder counts colour and shape words and hashes all other words into eight shared slots. Only exact and padded captions count as good. The scores are not on CLIP’s scale, so the thresholds are not comparable with 0.28 or 0.3.

import numpy as np
from scipy import ndimage

COLORS = {"red": (0.85, 0.15, 0.15), "green": (0.15, 0.7, 0.2),
          "blue": (0.15, 0.25, 0.85), "yellow": (0.9, 0.85, 0.1)}
SHAPES = ["circle", "square", "triangle"]
J = 8                                     # shared slots for every other word
D = len(COLORS) + len(SHAPES) + J
PAD = "free download hd wallpaper best price clipart png".split()
SPAM = "buy cheap shoes sale online shipping deal logo".split()

def draw(img, shape, rgb, r, rng):
    n = img.shape[0]
    cy, cx = rng.uniform(r, n - r, size=2)
    y, x = np.mgrid[:n, :n] + 0.5
    if shape == "circle":
        m = (y - cy) ** 2 + (x - cx) ** 2 <= r ** 2
    elif shape == "square":
        m = (abs(y - cy) <= 0.85 * r) & (abs(x - cx) <= 0.85 * r)
    else:                                 # upright triangle
        m = (y <= cy + r) & (abs(x - cx) <= (y - cy + r) / 2)
    img[m] = rgb + 0.03 * rng.standard_normal((m.sum(), 3))

def render(shape, color, faded, clutter, rng, n=32):
    """Grey background, one shape; faded = washed-out colour; clutter = 2nd shape."""
    img = 0.5 + 0.03 * rng.standard_normal((n, n, 3))
    if clutter:
        s2 = rng.choice([s for s in SHAPES if s != shape])
        c2 = rng.choice([c for c in COLORS if c != color])
        draw(img, s2, np.array(COLORS[c2]), rng.uniform(4, 7), rng)
    rgb = np.array(COLORS[color])
    draw(img, shape, 0.35 * rgb + 0.325 if faded else rgb, rng.uniform(7, 12), rng)
    return np.clip(img, 0, 1)

def image_embedding(img, rng):
    """Stand-in image encoder: colour and shape evidence of the largest region."""
    border = np.concatenate([img[0], img[-1], img[:, 0], img[:, -1]])
    bg = np.median(border, axis=0)
    lab, k = ndimage.label(np.linalg.norm(img - bg, axis=2) > 0.12)
    e = np.full(D, 0.05)
    if k:
        fg = lab == np.argmax(np.bincount(lab.ravel())[1:]) + 1
        mean = img[fg].mean(axis=0)
        d2 = ((np.array(list(COLORS.values())) - mean) ** 2).sum(axis=1)
        pc = np.exp(-d2 / 0.02); pc /= pc.sum()
        ys, xs = np.nonzero(fg)
        fill = fg.sum() / ((np.ptp(ys) + 1) * (np.ptp(xs) + 1))
        ps = np.exp(-(fill - np.array([0.79, 1.0, 0.52])) ** 2 / 0.0072)
        contrast = min(1.0, np.linalg.norm(mean - bg) / 0.5)
        e[:7] = np.concatenate([contrast * pc, ps / ps.sum()])
    return np.clip(e + rng.normal(0, 0.08, D), 0, None)   # encoder error

def text_embedding(caption):
    """Stand-in text encoder: colour and shape words; other words hashed to J slots."""
    v = np.zeros(D)
    for w in caption.split():
        if w in COLORS:
            v[list(COLORS).index(w)] += 1
        elif w in SHAPES:
            v[len(COLORS) + SHAPES.index(w)] += 1
        elif w != "a":
            v[7 + sum(map(ord, w)) % J] += 0.5
    return v

KINDS = ["exact", "padded", "wrong colour", "wrong shape", "spam"]

def make_pairs(n, rng):
    """Image-caption pairs whose correctness is known; 'good' = exact or padded."""
    sims, kind = [], rng.choice(5, size=n, p=[0.35, 0.2, 0.15, 0.15, 0.15])
    faded, clutter = rng.random(n) < 0.3, rng.random(n) < 0.3
    for k, f, cl in zip(kind, faded, clutter):
        s, c = rng.choice(SHAPES), rng.choice(list(COLORS))
        img = render(s, c, f, cl, rng)
        cap = {0: f"a {c} {s}",
               1: f"{c} {s} " + " ".join(rng.choice(PAD, rng.integers(1, 7), replace=False)),
               2: f"a {rng.choice([o for o in COLORS if o != c])} {s}",
               3: f"a {c} {rng.choice([o for o in SHAPES if o != s])}",
               4: " ".join(rng.choice(SPAM, 4, replace=False))}[k]
        a, b = image_embedding(img, rng), text_embedding(cap)
        sims.append(a @ b / (np.linalg.norm(a) * np.linalg.norm(b)))
    return np.array(sims), kind <= 1, faded, kind

def scores(sim, good, faded, t):
    keep = sim >= t
    rec = lambda g: (keep & g).sum() / g.sum()
    return keep.mean(), (keep & good).sum() / keep.sum(), rec(good), \
        rec(good & ~faded), rec(good & faded)

sim, good, faded, kind = make_pairs(2000, np.random.default_rng(0))
print("mean cosine (seed 0)   vivid  faded")
for k, name in enumerate(KINDS):
    print(f"  {name:19s} {sim[(kind == k) & ~faded].mean():.2f}   "
          f"{sim[(kind == k) & faded].mean():.2f}")
print("threshold  kept  precision  recall  recall(vivid)  recall(faded)")
for t in [0.3, 0.4, 0.5, 0.6, 0.7, 0.8]:
    kept, p, r, rv, rf = scores(sim, good, faded, t)
    print(f"   {t:.1f}     {kept:.2f}    {p:.3f}    {r:.3f}      {rv:.3f}         {rf:.3f}")

res = np.array([scores(*make_pairs(2000, np.random.default_rng(s))[:3], 0.8)
                for s in range(10)])
print("threshold 0.8, seeds 0-9: precision %.3f-%.3f, recall(vivid) %.3f-%.3f, "
      "recall(faded) %.3f-%.3f" % (res[:, 1].min(), res[:, 1].max(), res[:, 3].min(),
                                   res[:, 3].max(), res[:, 4].min(), res[:, 4].max()))

mean cosine (seed 0)   vivid  faded
  exact               0.94   0.83
  padded              0.82   0.73
  wrong colour        0.48   0.57
  wrong shape         0.52   0.30
  spam                0.08   0.10
threshold  kept  precision  recall  recall(vivid)  recall(faded)
   0.3     0.80    0.671    0.980      0.996         0.943
   0.4     0.78    0.678    0.969      0.987         0.928
   0.5     0.66    0.788    0.951      0.963         0.925
   0.6     0.55    0.924    0.928      0.937         0.910
   0.7     0.51    0.976    0.909      0.925         0.873
   0.8     0.44    0.985    0.803      0.832         0.738
threshold 0.8, seeds 0-9: precision 0.980-0.991, recall(vivid) 0.829-0.863, recall(faded) 0.651-0.738

Unrelated spam is removed at every threshold; the threshold matters for captions that are partly right. A caption with the right shape and the wrong colour, or the reverse, shares half of its evidence with the image and scores around 0.5, so a low threshold keeps most of them and precision stays near 0.67. Raising the threshold past that band lifts precision to 0.92 at 0.6, at little cost in recall, because good pairs mostly score higher. Above 0.7 the cost appears: padded captions spread their weight over search-engine words, and their scores fall toward the threshold and below it. This resembles the alt-text problem the LAION-5B paper notes: captions that are search-engine spam or keyword lists (§6).

The precision and recall of the whole set hide a second effect. On faded images the encoder’s colour evidence is weak, so good pairs score lower and are the first to go when the threshold rises (at 0.8 the faded and vivid recall ranges over seeds 0 to 9 do not overlap, and seed 0 is the best of the ten for faded images). The weak colour evidence also lets wrong-colour captions on faded images score higher than on vivid ones, so the filter is less able to judge exactly the images it is least sure of. A single global threshold therefore changes the composition of what is kept as well as its quality. In our reading this is the mechanism behind the LAION-5B paper’s warning that filtering with CLIP brings in CLIP’s own flaws and biases (§6; arXiv v1, Appendix G.2): the subjects and styles the filtering model represents poorly are kept or dropped at different rates from the rest.

The printed precision and recall come from known labels. LAION had none; the thresholds were set by inspecting samples, and the LAION-400M announcement reports a hand check of 3,456 samples that found 3 bad captions (0.09%) among them (LAION blog, 2021).

Dataset Size

Subset	Pairs	Language	Source
LAION-400M	413M unique samples (413,871,335 in the September 2021 files, as reported by Birhane et al.)	English	Schuhmann et al. (2021), Table 1; Birhane et al. (2021), §2
LAION-2B-en	2.32 billion	English	Schuhmann et al. (2022), §4
LAION-2B-multi	2.26 billion	more than 100 other languages; Russian 10.6%, French 7.4%, German 6.6%, Spanish 6.6%, Chinese 6.3%	Schuhmann et al. (2022), §4
LAION-1B-nolang	1.27 billion	no language detected with confidence	Schuhmann et al. (2022), §4
LAION-5B (total)	5.85 billion		Schuhmann et al. (2022), abstract
Re-LAION-5B	5,526,641,167		LAION blog (2024)

The datasheet and the LAION-5B announcement round the three subsets to 2.3, 2.2 and 1.2 billion, and the LAION-Aesthetics announcement speaks of 2.37 billion LAION-5B images with English captions. The paper attributes the no-language subset mostly to product names, places and other short text (§3.1, §4). LAION-400M’s image sizes are tabulated in the paper: 211 million samples have both sides at least 256 pixels and 9.6 million both sides at least 1,024 (Table 1). For LAION-2B-en, the announcement counts 1,324 million images with both sides above 256 pixels and 76 million above 1,024.

The Re-LAION-5B total is the figure LAION gives; the post does not say which of the two versions it counts, and it does not equal 5.85 billion minus either version’s stated removals, so the arithmetic cannot be checked from the post.

Annotation Types

LAION adds no human annotation. Each LAION-5B pair comes as a row of metadata in Apache Parquet files (Schuhmann et al., 2022, §4; LAION blog, 2022):

• a 64-bit identifier, the image URL and the alt-text;
• the image’s width and height;
• the language, for LAION-2B-multi, from CLD3;
• similarity, the CLIP ViT-B/32 cosine similarity that passed the filter (OpenAI CLIP for English, multilingual CLIP otherwise);
• punsafe, the probability of an unsafe image from a classifier on CLIP embeddings;
• pwatermark, the probability of a watermark from a watermark detector.

LAION-400M’s rows have a coarser tag, NSFW, UNSURE or UNLIKELY, and a LICENSE field holding a Creative Commons licence URL when one was found on the page, otherwise a question mark (LAION blog, 2021).

The announcement reports the share of images above its tag thresholds: 2.9% unsafe and 6.1% watermarked in LAION-2B-en at $p > 0.5$ and $p > 0.8$, and similar values for the other subsets; the paper gives 3% NSFW for LAION-5B (§4). The safety classifier’s accuracy on a balanced, manually annotated set of 3,000 samples was 0.960 on CLIP ViT-B/32 embeddings and 0.961 on ViT-L/14 embeddings. The watermark detector was trained on 90,000 images, half of them watermarked; the announcement says a reliable test set proved hard to build because people disagreed about what counts as a watermark (LAION blog, 2022).

Beside the metadata, LAION released CLIP embeddings (image and text embeddings for LAION-400M, about 1 TB; ViT-L/14 image embeddings for LAION-5B, about 9 TB) and k-nearest-neighbour indices built with autofaiss, which powered a public search interface searchable by text or image (LAION blog, 2021; Schuhmann et al., 2022, Appendix C).

Splits

There are none. The LAION-5B datasheet answers “No” to recommended training, validation or test splits (Q12); models trained on it are evaluated on other benchmarks.

Tasks

LAION supplies pretraining data rather than a benchmark task. Its creators list contrastive image–text pretraining, text-to-image generation, image captioning and image–text retrieval, with multilingual and low-resource language work and safety research as further uses (Schuhmann et al., 2022, §5.1; datasheet Q39, Q41). The embeddings and indices also make it a source for building narrower datasets by search, which the paper reports for face representation learning and other domains (§5.1).

Evaluation

LAION has no evaluation protocol of its own. The LAION-5B paper validated the dataset by training CLIP models on it with OpenCLIP and evaluating them zero-shot on ImageNet, four ImageNet distribution shifts and VTAB+, a set of 35 tasks the authors assembled, with the class-prompt averaging used by CLIP (§5.2, Table 2).

The paper raises the obvious risk of this arrangement: test images of downstream benchmarks may sit in Common Crawl and so in LAION, which would inflate zero-shot results. The authors judge the risk small, citing OpenAI’s overlap analysis for CLIP and the construction of ObjectNet from new photographs, and leave measuring the overlap to future work (§6). The benchmark overfitting article discusses the general problem.

License

As of October 2026, the LAION-400M and LAION-5B announcements state that LAION distributes the metadata (the Parquet files) under CC BY 4.0 and that “the images are under their copyright” (LAION blogs, 2021 and 2022). The LAION-5B datasheet adds that LAION owns the metadata and releases it under CC BY 4.0 and does not own the copyright of the images or the text (Q48, Q49). The alt-text is part of the metadata LAION licenses but, by the datasheet’s own answer, not text whose copyright LAION holds; in our reading, what the CC BY 4.0 licence grants for the alt-text is therefore unclear.

The terms of Re-LAION-5B are unclear. The Re-LAION-5B announcement says both versions are released under the Apache 2.0 licence (LAION blog, 2024). As of October 2026 the Hugging Face page of the English research subset shows the licence CC BY 4.0, and the page of the English research-safe subset has no dataset card and states no licence. Both are gated: the announcement says access requires affiliation information and agreement to use data that may contain explicit and disturbing content for research purposes. The research-only use that LAION recommends for both releases is advice, not a licence term.

Limitations and Bias

LAION inherits the biases of the web, of alt-text and of the model that filtered it. The datasheet says the data mirror the biases of society at the time of collection, citing correlations between gender and occupation, violence and nudity, and that the dataset should not be used to make decisions about people (Q42). Alt-text is written for screen readers and search engines, not for training: the paper notes that it may be search-engine spam, keyword lists or otherwise uninformative (§6). Three effects are documented:

• The filter’s own biases. CLIP was trained on an unreleased dataset, so its biases cannot be fully audited, and cosine similarity is a heuristic that can prefer a demeaning description over a neutral one (Schuhmann et al., 2022, arXiv v1, Appendix G.2). Birhane, Prabhu and Kahembwe (2021, §3.3.3) gave an example with the scikit-image astronaut photograph, for which CLIP scored a stereotyped description higher than an accurate one (0.31 against 0.28), so that at a threshold of 0.3 only the stereotyped pair would have been kept.
• Duplicates. Deduplication by URL leaves the same image at different addresses. Webster et al. (2023) compressed CLIP features to search LAION-2B-en for near-duplicates and estimated that roughly 700 million images, about a third, are duplicated, with 91% precision on a checked sample (arXiv preprint, abstract and §1); the LAION-5B datasheet also mentions near-duplicates (Q13). Duplication matters for generative models, which are more likely to reproduce images that occur many times (Thiel, 2023, §3.1, citing memorization studies of diffusion models).
• Link rot. The datasheet offers no guarantee that the linked images still exist (Q14). The Stanford Internet Observatory found about 30% of the URLs it submitted for checking no longer live in late 2023 (Thiel, 2023, §2), so studies that download the images at different times do not train on identical data.

Audits have documented harmful content. Birhane, Prabhu and Kahembwe (2021), an arXiv preprint posted on 5 October 2021 and not peer-reviewed, searched LAION-400M through the public search interface as it stood on 12 September 2021 and string-matched the released metadata. They reported that ordinary queries, including words for women, family relationships, nationalities and ethnic groups, returned pornographic and sexually violent images, that searches returned stereotyped results, and that many images with neutral content carried offensive alt-text (§2). For three search words, they counted the captions that also contained any of a short list of explicit terms, a crude string filter that they present as a demonstration rather than a measurement: 34.1%, 16.4% and 28.2% of the matches, against the announcement’s NSFW share of less than 1%. The CLIP-derived tag marked only 164 to 918 of each word’s 16,766 to 37,769 matches as NSFW (Table 1). In a peer-reviewed follow-up in the NeurIPS 2023 Datasets and Benchmarks Track, Birhane et al. (2023, §3–4) scored the alt-text of 3.2 million LAION-400M and 12.8 million LAION-2B-en samples with the pysentimiento hate-speech classifier. The share of texts flagged as hateful, targeted or aggressive at a probability of 0.5 was 0.298% for LAION-400M and 0.344% for LAION-2B-en, a relative increase of 12.26% even when the 95% confidence intervals are compared at their nearest ends. They found that filtering by the image-based punsafe score did not remove all hateful alt-text, and they suggest the lowered CLIP threshold (0.3 to 0.28) as one possible reason for the increase. The LAION-5B paper cites the 2021 audit as motivating its improved tagging and the watermark filter (§7).

Privacy, Consent and Copyright

The LAION-5B datasheet (arXiv v1, Appendix A, 2022) answers the consent questions directly. Individuals were not notified of the collection (Q30). LAION followed Common Crawl’s practice of respecting each site’s robots.txt file, which the datasheet treats as the site’s consent to crawling, but it states that people shown in the photographs may not have consented to their upload (Q31). People in the images may be identifiable by face recognition or through the alt-text (Q19), the dataset contains sensitive content (Q20), and its links can lead to personal information such as photographs of faces or medical images (Appendix G.1). The datasheet reports that the University of Washington’s Human Subjects Division said the work did not require IRB review, because the data were public and the researchers did not interact with the people depicted, and that the NeurIPS ethics review found no serious ethical issues (Q27). It cites Zheng et al.’s use of LAION-400M for facial representation learning as showing that the data can be misused for identification (Q39).

For removal, the datasheet points to the search tool for finding one’s own images, to the host sites, and to LAION’s contact form and e-mail address for removing links (Q32, Q56). It notes that deleting an image from its host leaves the link leading nowhere, while removing a link from the dataset leaves the image itself on the web. As of October 2026 LAION’s FAQ offers a takedown form for personal data under the GDPR, says verified entries are removed from the repositories LAION controls and from future releases, and says LAION cannot act on past releases that circulate outside its control, for example through torrents.

On copyright, the same FAQ (as of October 2026) states LAION’s position that its datasets are indexes of links and contain no images, that the images it downloaded to compute CLIP similarities were discarded, and that the text and data mining exceptions of Article 3 of the EU Directive on Copyright in the Digital Single Market and §60d of the German Copyright Act cover its research. These are LAION’s statements, not legal findings. The text-to-image generation article summarizes official documents on training generative models with such data.

Withdrawal and Re-LAION-5B

The Stanford Internet Observatory report. David Thiel’s report, Identifying and Eliminating CSAM in Generative ML Training Data and Models, was first released on 20 December 2023, according to its changelog. It examined all three subsets of LAION-5B, starting in September 2023 (Thiel, 2023, §2–4). The method was designed so that no known material was downloaded or stored; candidates for the classifier step were held briefly in memory and deleted (§6): the URLs of all 32,138,129 entries with punsafe of at least 0.995 were submitted to Microsoft’s PhotoDNA service, which matches perceptual hashes of known material, and matches were sent to the Canadian Centre for Child Protection (C3P) for verification. MD5 hashes of LAION images were compared with hash lists of known material from the National Center for Missing and Exploited Children, which also found images whose links had died. The CLIP embeddings of confirmed matches seeded nearest-neighbour searches, whose results went through PhotoDNA and a classifier from Thorn, with high-probability candidates sent to C3P. The report identified 3,226 entries of suspected CSAM, of which C3P validated 1,008 as CSAM or likely CSAM (Tables 2 and 3), reported all detected URLs to NCMEC and C3P, and called its count a significant undercount because hash lists are incomplete, many links were dead, LAION’s reference images were not available to it and the unsafe classifier is imperfect (§6–7). It judged the filters used during collection too narrow, since only images both tagged NSFW and matching the minor-related CLIP filters had been discarded, and noted that images had not been checked against hash lists of known material (§5.2). It recommended that datasets compiled from the web be checked against hash lists of known material through organizations that hold them, and that holders of LAION-5B-derived training sets delete them or clean them through such intermediaries (§5.2, §7).

The withdrawal. On 19 December 2023 LAION posted a statement that it was taking its datasets down temporarily, in an abundance of caution, to ensure they were safe before republishing them, and that it was working with the Internet Watch Foundation (IWF) (LAION, 2023). The statement said that LAION had applied filters for illegal content before each release, pointing to the LAION-400M announcement, and that its datasets contain links, not images. LAION later wrote that it took down all known accessible copies of LAION-5B and its derivatives and urged users to delete their copies or remove the suspected links (LAION blog, 2024, Phase 1). As of October 2026 the Hugging Face addresses of the original LAION-2B-en and LAION-400M metadata redirect to gated repositories named for Re-LAION, and the LAION-400M one has no dataset card.

Re-LAION-5B. LAION announced Re-LAION-5B on 30 August 2024, describing work with the IWF, C3P and the Stanford Internet Observatory (LAION blog, 2024). As LAION describes it, the partners supplied lists of MD5 image hashes and URL hashes of known material, 2.2 million from the IWF and 14 million from C3P as of July 2024, and the report’s author supplied further hashes. LAION matched these against hashes of URLs and images it had computed for LAION-5B in early 2023, and LAION says this let it remove the links without accessing any of the linked content. It reports 1,129 matches from C3P, 18 from the IWF and 1,714 from the Stanford hashes, totalling 2,236 removed links that include the report’s 1,008; the three counts sum to 2,861, and the post does not say how overlaps reduce them to 2,236. LAION calls 2,236 an upper bound on links to suspected CSAM, since many of the links were probably dead. It also removed 399 links reported by Human Rights Watch as potentially containing private information about children. Two versions were then released:

• Re-LAION-5B research additionally removes samples with punsafe above 0.95 whose text matches keyword filters built from child-protection organizations’ recommendations, which LAION reports as 1.121% of the original dataset (about 65 million pairs).
• Re-LAION-5B research-safe additionally removes samples with punsafe above 0.45, the majority of NSFW content, which LAION reports as 3.044% (about 176 million pairs). It is a subset of the research version, and LAION says all further derived datasets use this filter.

LAION recommends that holders of copies of the original LAION-5B use the Re-LAION metadata to remove the affected entries, and its announcement concludes that classifier-based filtering alone could not guarantee the absence of such material at web scale and that dataset builders should obtain hash lists from expert organizations. That conclusion agrees with the report’s recommendation.

Common Uses

• Open CLIP models. OpenCLIP models trained on LAION-400M and LAION-2B-en came within a few points of OpenAI’s CLIP on zero-shot ImageNet (ViT-L/14 on LAION-2B-en: 75.2% against 75.6%; Schuhmann et al., 2022, Table 2), the first open reproductions of OpenAI’s largest CLIP models (§1). The CLIP and vision–language pretraining articles cover the scaling studies that followed.
• Stable Diffusion. Stable Diffusion v1 was trained on LAION-2B-en, LAION-High-Resolution and an aesthetics-filtered subset (Schuhmann et al., 2022, Appendix F.2); the Stable Diffusion article gives the training stages and the naming differences between sources. The latent diffusion model of Rombach et al. that preceded it trained its text-to-image model on LAION-400M.
• Other models. The paper lists BLIP, VQ-Diffusion, MAGMA and its own GLIDE fine-tuning among early uses of the data (§5.1), and Imagen trained on about 460M internal image-text pairs together with the whole of LAION-400M (Saharia et al., 2022, §4.1).

Historical Milestones

• August 2021: LAION-400M is announced as an open counterpart to CLIP’s private training set (LAION blog, 2021); the paper follows in November and appears at the NeurIPS 2021 Data-Centric AI workshop.
• October 2021: Birhane, Prabhu and Kahembwe post an audit of LAION-400M’s content as an arXiv preprint.
• 31 March 2022: LAION-5B is released: 5.85 billion pairs, 2.32 billion of them English, with ViT-L/14 embeddings, indices and safety and watermark scores (LAION blog, 2022).
• August 2022: LAION-Aesthetics V2 is announced (16 August), and Stable Diffusion v1, trained on LAION-5B subsets, is released (see the Stable Diffusion article).
• November 2022: The LAION-5B paper receives a NeurIPS 2022 Outstanding Datasets and Benchmarks Paper award (NeurIPS Blog, 21 November 2022).
• December 2023: The Stanford Internet Observatory report is released and LAION withdraws its datasets.
• August 2024: Re-LAION-5B is released in research and research-safe versions.

Related Datasets

• Conceptual Captions (CC3M) and Conceptual 12M were also built from web alt-text, with additional cleaning that the second relaxed, at 3 and 12 million pairs (Schuhmann et al., 2022, §2, Table 1).
• CLIP’s WebImageText (WIT), 400 million pairs, and ALIGN’s 1.8 billion noisy alt-text pairs were the private datasets LAION set out to match.
• COCO captions and Flickr30k are small, human-annotated caption datasets; the LAION-5B paper evaluates its CLIP models’ zero-shot retrieval on the Flickr30k 1K and COCO 5K test sets (arXiv v1, Tables 8 and 9). Visual Genome adds region descriptions and relationships.
• YFCC100M offers about 99 million Flickr images and one million videos with user metadata rather than alt-text (Schuhmann et al., 2022, §2).
• DataComp (Gadre et al., 2023) treats the filtering itself as the object of study: participants choose a training subset from a candidate pool of 12.8 billion Common Crawl pairs, and a standardized CLIP training run and 38 downstream test sets score the choice (abstract).

Resources

• The paper: Schuhmann et al. (2022), on how LAION-5B was built, the CLIP reproductions and the paper’s datasheet.
• LAION-400M announcement and LAION-5B announcement: construction, metadata columns, statistics and licence.
• Re-LAION-5B announcement: the safety revision and the two release versions, with links to the gated repositories.
• LAION FAQ and dataset requests: takedown and removal requests.
• img2dataset: LAION’s tool for downloading and resizing images from URL lists.
• clip-retrieval: embedding computation and nearest-neighbour search over the indices.
• CLIP-based NSFW detector and LAION-5B watermark detection: the classifiers behind punsafe and pwatermark.
• OpenCLIP: training code and models trained on LAION data.
• Stanford Internet Observatory report: the December 2023 findings and method.