Best AI Datasets

The Best Alternatives to LAION for Image-Text Datasets in 2026

Answer in brief: LAION's Re-LAION-5B (5.5 billion text–image-link pairs) and its new LAION-BVD video set are still the biggest open multimodal resources. Depending on your goal, though, these are better choices. DataComp-1B is a 1.39-billion-row set built around benchmark-driven filtering, with its URL-text pairs and metadata under CC-BY-4.0. COYO-700M has 747 million image-text pairs with rich quality signals on every row, listed as CC-BY-4.0. PD12M is 12.4 million public-domain and CC0 images with licenses recorded per row. COCO is the standard for detection, segmentation and captioning. This is independent research from each dataset's official card or site.

Top pick: DataComp-1B. If you want web-scale image-text pairs with CLIP scores and face bounding boxes already attached, plus a public benchmark to measure your filtering against, DataComp is the strongest LAION alternative.

LAION alternatives compared

DatasetScaleWhat is in each rowLicensingBest for
Re-LAION-5B (baseline)5,526,641,167 pairsURL, caption, metadata; research and research-safe versionsApache-2.0 metadata; images are third-party linksLargest open CLIP-style pretraining
DataComp-1B1.39B rowsURL, text, CLIP B32/L14 similarity, face bounding boxes, sha256CC-BY-4.0 for URL-text pairs and metadata; each image keeps its own copyrightFiltered CLIP training, data-centric research
COYO-700M747M pairspHash, CLIP similarity, two NSFW scores, watermark and aesthetic scoresCC-BY-4.0 (per dataset card); images are third-party linksBuilding your own quality filters
PD12M12.4M imagesCaption, per-row license (e.g. CC0), source institutionPublic domain or CC0, recorded per rowTraining where clear image rights matter
COCO330K images, 1.5M object instances80 object and 91 stuff categories, 5 captions per image, keypoints for 250K peopleSee COCO termsDetection, segmentation, captioning benchmarks

DataComp-1B: filtering you can measure

DataComp turned dataset building into a benchmark: fix the training recipe and compete on the data. DataComp-1B is the filtered pool that came out of that work. Every row includes CLIP B/32 and L/14 similarity scores and face bounding boxes, so you can refilter or blur faces without recomputing everything. The URL-text samples and metadata are distributed under CC-BY-4.0, and the card notes that the individual images remain under their own copyrights. DataComp.ai now runs tracks for CLIP, language models, reasoning and VLMs.

COYO-700M: quality signals on every row

COYO's strength is its metadata. Its 747 million alt-text and image pairs, collected from HTML documents, each carry a perceptual hash for deduplication, image-text CLIP similarity, two NSFW scores, a watermark score and an aesthetic score. That makes it a strong base if you want to design your own quality threshold. The dataset card lists CC-BY-4.0, and the dataset is not gated. As with any link-based set, the images themselves belong to third parties.

PD12M: images you can actually account for

PD12M is small by web standards but unusually clean on rights. Each row records its license and its source institution, such as the National Museum in Warsaw or the Smithsonian NMNH. The images are hosted in the dataset's own S3 bucket, so you are not chasing broken links. If provenance is your priority, start here.

COCO: the benchmark everyone reports

COCO is not a pretraining corpus. It is the dataset that vision models are evaluated and fine-tuned on: 330K images, more than 200K of them labeled, with instance segmentation, captions and keypoints. You can browse and load it through FiftyOne.

When to stay with LAION

Stay with LAION if you need maximum scale, video (LAION-BVD has 80M downloaded videos totalling 10M hours and 55M captioned clips) or LAION's tools: img2dataset, CLIP Retrieval, CLIP Benchmark and OpenCLIP.

Pricing model

All the datasets here are free to download. For link-based sets (Re-LAION, DataComp, COYO) you pay for the bandwidth and storage of fetching the images yourself, which is often the biggest cost. PD12M and COCO host their images. If you need licensed or custom-annotated imagery, see Best AI training data providers.

FAQ

Is the original LAION-5B still available?

LAION has superseded it with Re-LAION-5B (August 2024). That release removed 2,236 links identified with partners including IWF, C3P and the Stanford Internet Observatory.

Which LAION alternative has the clearest image licensing?

PD12M. Each row records a public-domain or CC0 license and its source institution. DataComp-1B and COYO-700M license their metadata under CC-BY-4.0, but the linked images keep their own copyrights.

Do I have to download the images myself?

For Re-LAION-5B, DataComp-1B and COYO-700M, yes: they distribute URLs and metadata. Tools such as img2dataset do the fetching.

Keep exploring

Sources

Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.