LAION Opens 80 Million Video Archive to Break the Multimodal Training Data Bottleneck

LAION, the German non-profit that helped democratize image-model training with its five-billion-image dataset, published a new resource today that attempts the same feat for video: the Big Video Dataset, or BVD, an openly accessible corpus containing 80 million downloaded videos with a combined runtime of 10 million hours. The release lands at a moment when frontier video-generation and video-understanding models increasingly depend on massive private training pipelines that academic researchers simply cannot replicate — and BVD directly targets that gap.

The dataset, built from 1.3 billion video URLs extracted from the CommonCrawl web archive, is accompanied by an arXiv preprint from a twelve-person team drawn from the Tübingen AI Center, the Max Planck Institute for Intelligent Systems, the Jülich Supercomputing Centre, the ELLIS Institute Tübingen, and LAION itself. Authors include Bernhard Schölkopf and Matthias Bethge, two prominent figures in European machine-learning research. The dataset and accompanying processing code are available now under a research-only license that explicitly bars commercial use.

Why Open Video Data Has Been the Binding Constraint

For multimodal AI — models that jointly understand video, audio, and text — the most consequential scarcity has not been compute. Cloud GPU access has become increasingly affordable, as demonstrated by the Open-Sora 2.0 project, which trained a commercial-quality video-generation model for roughly $200,000. The harder problem has been data. Labs producing systems like Sora or Runway Gen-3 draw on massive proprietary video corpora built through licensing agreements or internal data pipelines that external researchers cannot access. The result has been a structural asymmetry: researchers outside a small group of well-funded organizations have had to work with open video datasets that are an order of magnitude smaller than what frontier models use internally.

The previous open benchmark in this space was InternVid, published by the OpenGVLab team at Shanghai AI Lab in 2023 and accepted as a spotlight at ICLR 2024. InternVid covered approximately 7 million videos and 760,000 hours of content, yielding 234 million annotated clips. Those figures were a substantial advance over what existed before. But 760,000 hours is still roughly 13 times smaller than BVD's 10 million hours — and InternVid covers only video and text, not audio. The InternVid paper established the prior benchmark that BVD now significantly surpasses in scale.

How LAION Assembled 10 Million Hours

The BVD pipeline moves through four main stages. Starting from 1.3 billion platform-specific video URLs harvested from CommonCrawl snapshots, the team successfully downloaded 80 million videos — a completion rate that reflects both availability and filtering. From those videos, content-aware scene detection identifies visual and temporal boundaries to extract 55 million clips. This approach matters: splitting video at fixed time intervals cuts across scene transitions, producing clips where the visual content and any associated description belong to two different scenes. Scene-aware segmentation produces clips where the content is internally coherent, which is the right unit for training video-text contrastive models.

Those 55 million clips then receive two distinct layers of synthetic annotation. Video captions are generated automatically using a vision-language model to describe the visual content of each clip. Audio captions are generated separately from the audio track, producing descriptions compatible with CLAP — the Contrastive Language-Audio Pretraining framework used to train audio-text retrieval models. This dual-caption approach is what allows BVD to serve as training data for both video-language and audio-language models from the same underlying corpus.

The third modality comes from frame extraction. The team extracts 300 million still images at scene-change boundaries, producing image-text pairs that can train CLIP-class image-text models. These frames come from real-world video rather than the curated photography that dominates web image datasets, giving them a distinct visual distribution that the authors say complements existing image pre-training sources.

No comparable open corpus currently provides all three modalities — video, audio, and images — from a unified source at this scale. InternVid provides video and text. Panda-70M, another large open dataset, provides roughly 70 million clips but no audio-text pairs. BVD is the first open resource to span all three training targets simultaneously.

Benchmark Results and What They Do and Do Not Show

According to the LAION-BVD paper, ViCLIP models trained on BVD outperform those trained on InternVid's FLT subset by up to 2.1 percentage points on standard video-text benchmarks, with consistent gains when moving from 10 million to 50 million training clips. CLAP models trained on BVD's audio data perform competitively against other large-scale uncurated audio datasets, with favorable scaling trends as model and dataset size increase. CLIP models trained on the extracted frames achieve strong image-text retrieval scores on standard benchmarks.

These results should be read carefully. The experiments were run by the paper's authors — the same team that built the dataset — not by independent evaluators. The specific benchmark names and exact evaluation conditions are detailed in the full paper rather than the abstract, and the "+2.1 pp" figure represents the upper end of a range, achieved at the larger clip-count training runs. The consistency of improvement across training scales (from 10M to 50M clips) is a meaningful positive signal — random noise in data quality tends to produce inconsistent trends — but independent reproduction will be necessary before practitioners can rely on these numbers for production decisions.

Read more: Frontier vision-language models still fail key spatial reasoning tasks, benchmark finds

The Legal Architecture Behind the Release

LAION's ability to build and release BVD rests partly on a series of German court rulings that tested the legality of scraping copyrighted material for non-commercial research training data under EU law. In September 2024, the Regional Court of Hamburg dismissed a cease-and-desist lawsuit brought by photographer Robert Kneschke against LAION, finding that LAION's dataset creation activities fell within Germany's §60d UrhG — the national transposition of Article 3 of the EU Digital Single Market Directive, which carves out a text-and-data-mining exception for scientific research. In December 2025, the Hamburg Higher Regional Court (Oberlandesgericht) issued an appellate ruling in the same case (Ref. 5 U 104/24) that provided further guidance on the admissibility of AI training and data mining under German law.

LAION's research-only license for BVD is the mechanism designed to keep the dataset within that legal protection. The text-and-data-mining exception under Article 3 covers scientific research; it does not cover commercial use. By prohibiting commercial applications, LAION is explicitly trying to remain within the exception's scope. Whether downloading and redistributing a dataset compiled under that exception extends the protection to all downstream researchers is a question that legal scholars note remains partially unresolved in EU law — but LAION's approach of a research-only license and explicit terms of use is the currently defensible posture.

Limitations and Risks Practitioners Should Know

Several constraints bound what BVD actually offers. The research-only license is not a technical restriction, but it means commercial AI developers cannot legally incorporate the dataset into production training runs. Any organization that trains a model on BVD and then deploys that model commercially would be operating outside the terms of use.

All captions — both video and audio descriptions — are synthetically generated. The paper does not specify which captioning models were used to produce them, which makes it difficult to independently assess the quality ceiling. Synthetic captions carry known risks: hallucinated visual content in descriptions, failure to capture domain-specific terminology, and linguistic uniformity that may not reflect the diversity of human-authored captions. Models trained on synthetic captions at scale have shown these data to be useful but not equivalent to human annotation for high-precision retrieval tasks.

LAION also explicitly warns that BVD, like other large-scale web datasets, may contain biases, stereotypes, and uneven representation across languages, regions, and topics. This is a known property of CommonCrawl-derived corpora, which reflect the composition of the indexed web rather than any curated balance. LAION's prior flagship dataset, LAION-5B, was temporarily taken offline in 2023 after a Stanford Internet Observatory study identified thousands of suspected child sexual abuse material images within the corpus. BVD faces an analogous risk with video content, which is harder to audit at scale than static images. The LAION team has not publicly described a video-specific content audit methodology for BVD, and practitioners building models on the dataset should account for that uncertainty.

What the Release Changes for Open AI Research

The historical pattern LAION established with LAION-5B is instructive. That dataset's 2022 release enabled the training of Stable Diffusion and catalyzed a broad open image-generation ecosystem, substantially narrowing the gap between what proprietary labs could produce and what academic researchers and independent developers could replicate. The same dynamic has not yet played out for video — frontier video-generation systems like Sora remain well ahead of open alternatives in quality — but the data scarcity that has been one structural reason for that gap just changed.

What shifts next depends on how the research community uses BVD. The scaling trends the authors document — consistent benchmark improvement as training clips increase from 10M to 50M — suggest that simply training on more of this data continues to help, which is the kind of result that makes a dataset genuinely useful rather than a one-time benchmark exercise. The question for 2026 and beyond is whether the combination of BVD-scale open data and increasingly affordable compute produces a generation of open video-understanding and video-generation models comparable to what LAION-5B produced for images. The data constraint has been removed. The compute and architecture work remains.

Read more: DeepSeek adds vision to its V4-Flash model in latest multimodal expansion