Open by default
Datasets, models, and tools built for the open machine-learning research community.
A large-scale open video dataset for multimodal learning, containing 1.3B platform-specific video URLs collected from CommonCrawl, 80M downloaded videos with a total duration of 10 million hours, 55 million captioned video clips, and 300 million captioned video frames.
The updated LAION-5B: 5.5 billion text–image-link pairs with additional safety fixes, released in research and research-safe versions.
An effort to create a large-scale dataset consisting of 3D models and descriptor pairs.
An audio dataset for training CLAP and other models, containing a raw and processed dataset, the latter containing .flac files with captions, labels, and other metadata.
A repository containing datasets to train a watermark classifier.
An open source implementation of OpenAI's CLIP (Contrastive Language-Image Pre-training).
An open vision–language model trained on LAION-2B with OpenCLIP.
An implementation of DALL-E 2, OpenAI's text-to-image synthesis neural network, in Pytorch.
Generate text from embedding, using pretrained encoder and language models.
Contrastive Language-Audio Pretraining: open audio–text models, pretrained checkpoints, and evaluation code.
A contrastive language pretraining model for videos.
An implementation of OpenAI's CLIP text encoders for any language.
A detector for not safe for work content within images using CLIP.
Train Contrastive and generative models on all modalities.
Download, resize, and package images from URL lists for reproducible dataset processing.
Compute CLIP embeddings and build a search system for image–text datasets.
GPU controlled Hetzner Cloud workers swarm for Crawling@Home project.
Evaluate vision–language models on zero-shot classification and retrieval tasks.
A 10-million-hour open video dataset for multimodal pre-training.
Earlier releases and subsets, preserved for reference. For the updated LAION-5B dataset, use Re-LAION-5B.
Formerly known as crawling@home (C@H), an openly accessible 400M image-text-pair dataset.
The original 2022 image–text dataset. Re-LAION-5B is the updated release with additional safety fixes.
600M captions generated using BLIP from Laion2B-en.
3B translated samples from Laion5B.
A subset of the LAION5B database, with high resolution images over 1024x1024, containing 170 million samples.
Subsets of the original LAION-5B selected using models that estimate visual aesthetics.
