What Image-to-Video Models Were Trained On: Known Datasets, Disputed Sources, and Live Litigation
The known datasets, disputed sources, and live litigation concerning the training data used by image-to-video generation models, with named cases and jurisdictions.
Image-to-Video Model Training Data Provenance and Live Litigation
The legal risk lands on the person who publishes the output, not the company that built the tool. Training data provenance is the single fact that determines whether your clip rests on a settled licence, a disputed scrape, or a source a court has already ruled infringing. This documents what each major developer has disclosed, what third parties have found, and which lawsuits are in progress as of September 2026.
Every image-to-video model learns motion patterns from millions of video clips. If those clips were used without permission, the model may be a derivative work, and every clip it generates inherits that legal uncertainty. Some developers paid for their data. Some relied on public compilations with contested copyright status. Some published model cards that state clearly what they used. Some published nothing at all.
The EU AI Act Article 53, which takes effect in stages through 2026, requires deployers of certain high-impact AI models to disclose a sufficiently detailed summary of the training data used. For image-to-video models, that means the developer must publish which sources and which licensing terms apply. The UK IPO consultation on computer-generated works is separately examining whether training on publicly available data without explicit permission constitutes fair dealing. Neither regulation has been tested in court, but both set a floor for what counts as acceptable disclosure.
Stable Video Diffusion Training Data: What Is Known
The Named Sources And Their Limits
The model card for Stable Video Diffusion names two primary training sources: LAION-5B and WebVid-10M. LAION-5B is a collection of image-text pairs scraped from the public web. WebVid-10M contains 10 million video clips with captions, sourced from stock footage sites and publicly available video. Neither collection required explicit permission from the copyright holders of the individual images or clips at the time of collection.
Stability AI also announced a partnership with Shutterstock in July 2023, under which Shutterstock contributed licensed images and video to the training pipeline. That partnership provided a royalty-paid data stream for later model versions, but it did not retroactively cover the LAION-5B or WebVid-10M data used in the initial release. The model card states that the model is released under the Stability AI Community Licence, which grants non-commercial use only. Commercial use requires a separate licence.
The Lawsuits That Hang Over LAION-5B
The training data transparency offered here is better than most competitors, because the model card names specific collections rather than a vague phrase like 'publicly available data'. But those collections remain legally contested. LAION-5B is the subject of ongoing discovery in Andersen v Stability AI, filed in the Northern District of California in January 2023. The plaintiffs argue that the collection contains their copyrighted images, collected without consent, and that models trained on it are infringing derivative works. The case has not reached trial as of September 2026.
Getty Images v Stability AI, filed in the same court in February 2023, challenges Stability AI's use of Getty's watermark-stripped images in the LAION collection. That case is further along, with a ruling on the fair use defence expected in late 2026 or early 2027.
Runway Gen-3 Training Data Copyright Dispute
Runway has not published a training data statement for Gen-3 Alpha. The company's public materials describe the model as trained on 'a large dataset of videos and images', but they do not name specific sources, collections, or licensing arrangements. This disclosure gap is unusual for a commercial product at Gen-3's user base. The training data question has become a central one for any legal or compliance person clearing output for broadcast or advertising.
What The Watermarks Suggest
Third-party investigations have not confirmed what Runway used. The model's output sometimes reproduces watermarks from stock footage providers, which suggests those clips were present in the training set. A watermark from a specific stock agency appearing in a Gen-3 generated clip is not definitive proof of unauthorised use, but it is strong circumstantial evidence that the model memorised that agency's content. No lawsuit has been filed against Runway as of September 2026, but the lack of a training data statement makes the model a higher-risk choice for any use that requires a clear chain of provenance.
Runway's partnership with Adobe Stock, announced in 2024, provides a licensed data stream for future model versions. The agreement allows Runway to train on Adobe Stock's video library in exchange for revenue sharing with contributors. That partnership covers models trained after the agreement date, but it does not retroactively cover whatever data was used for Gen-3 Alpha's initial training. The company has not stated whether Gen-3 Alpha was retrained on the Adobe Stock data or whether the partnership applies only to Gen-4 or later releases.
| Model | Training Data Statement | Known Source | Active Lawsuit | Licensed Data Stream |
|---|---|---|---|---|
| Stable Video Diffusion | Published model card | LAION-5B, WebVid-10M | Andersen v Stability AI (ND Cal, Jan 2023); Getty Images v Stability AI (ND Cal, Feb 2023) | Shutterstock partnership (Jul 2023) |
| Runway Gen-3 Alpha | None published | Undisclosed | None | Adobe Stock partnership (2024) |
| Wan 2.1 | Not disclosed by Alibaba | Undisclosed | None | None disclosed |
| LTX-Video | Model card states Apache 2.0 licence | Undisclosed training data | None | None disclosed |
| Kling 1.6 | Not disclosed by Kuaishou | Undisclosed | None | None disclosed |
| Luma Dream Machine | Not disclosed by Luma Labs | Undisclosed | None | None disclosed |
| Pika 2.0 | Not disclosed by Pika Labs | Undisclosed | None | None disclosed |
| OpenAI Sora | Not disclosed | Undisclosed | None | None disclosed |
| Google Veo 2 | Not disclosed by DeepMind | Undisclosed | None | None disclosed |
Open-Weights Models: Wan 2.1 and LTX-Video
Transparency varies even within the category of models that let you run inference locally. Wan 2.1, released by Alibaba, provides multiple parameter sizes (1.3 billion and 14 billion) and runs on consumer hardware via ComfyUI and Hugging Face Diffusers. The model card does not name the training sources. Alibaba has not published a data processing pipeline, a list of inputs, or a licence for the training material itself. The model weights are available for download. The training data is not. This is the gap between open-weights and open-source: you can use the model, but you cannot reproduce it or audit its training.
Why A Permissive Licence Is Not Enough
LTX-Video, from Lightricks, is distributed under the Apache 2.0 licence, which permits commercial use, modification, and redistribution of the model code and weights. The model card states the training data is undisclosed. Apache 2.0 covers the software, not the data. A user who trains a derivative model on the LTX-Video architecture still needs their own legally cleared collection. The advantage of LTX-Video is the licence for the model itself, not for the material used to create it. Inference is fast enough for interactive use on a single consumer GPU, but the data provenance gap remains.
For any compliance workflow, the distinction matters. A model with a permissive licence but undisclosed training data does not provide a legal safe harbour. The EU AI Act Article 53 requires a summary of the training data from the developer, not from the user. If the developer does not provide it, the deployer may be in non-compliance even if the user did not train the model. The UK IPO consultation on computer-generated works is examining whether training on unlicensed web data constitutes fair dealing, and the outcome will affect how these models are treated under UK law.
Dataset Litigation 2026: Key Cases and Jurisdictions
Litigation over training data for video models is concentrated in three jurisdictions. The United States has the most active docket, with Andersen v Stability AI and Getty Images v Stability AI both proceeding through discovery and motions practice. The Andersen case, filed in the Northern District of California on January 13, 2023, is a class action on behalf of artists whose work appears in the LAION-5B collection. The Getty Images case, filed in the same court on February 3, 2023, focuses on Getty's watermark-stripped images. Both cases raise the question of whether training a generative model on copyrighted images constitutes fair use, and both are likely to be appealed regardless of the trial outcome.
Europe And China
The European Union does not yet have a direct training data lawsuit against a video model developer, but the EU AI Act creates a regulatory enforcement pathway. National data protection authorities can investigate whether a model's training data complied with the GDPR at the time of collection. A developer who scraped personal images from social media without consent may face a GDPR fine even if no copyright claim succeeds. The UK IPO consultation, launched in March 2025, is examining whether to create a specific exception for text and data mining for AI training. The outcome could create a safe harbour for models trained in the UK on publicly available data, but the consultation is still open and no legislation has been introduced.
China has its own regulatory framework under the 2023 Generative AI Administrative Measures, which require content controls and watermarking for models released in China. The measures do not address training data copyright directly, but they impose an obligation to respect intellectual property rights. Kling, Wan, and Seedance are all subject to these measures. No training data lawsuit has been filed in a Chinese court against any of these developers as of September 2026.
What the EU AI Act Article 53 Requires
EU AI Act Article 53 applies to providers of general-purpose AI models that are classified as having systemic risk. An image-to-video model with more than 10^25 floating point operations during training triggers the systemic risk classification automatically. Below that threshold, the European Commission can designate a model as systemic risk based on its capabilities. The article requires the provider to publish a sufficiently detailed summary of the training data used. The summary must include the sources of the data, the selection procedures, and the licensing terms where applicable.
What Non-Compliance Means For You
The requirement applies to models placed on the market after the relevant application date. For models already on the market before the date, the provider must comply within a transition period. Stable Video Diffusion, released before the transparency provisions took effect, must now be accompanied by a data summary from Stability AI. Runway Gen-3 Alpha, released after the relevant date, was required to have a data summary at the point of release. It did not. That non-compliance could lead to an investigation by the national market surveillance authority in the EU member state where Runway first made the model available.
The practical consequence for a user in the EU is that publishing a clip generated by a model whose training data is not disclosed may itself be a compliance issue under the AI Act. The deployer, not the provider, is responsible for the output. If the deployer cannot demonstrate that the model's training data was lawfully obtained, the deployer may be liable for the infringement. Start your compliance workflow with a model that has a published training data statement, not with one that does not.
What To Do Next: Audit Your Model Before You Publish
Check the model card first. If the developer has not published a training data statement, that is a red flag you cannot resolve by reading the card harder. Move to a model that has one. For commercial use, Stable Video Diffusion with the Shutterstock partnership data stream is the best documented option, but the Getty Images litigation still casts a shadow over the entire LAION-5B lineage. For EU-based publishing, any model released without a data summary after the EU AI Act application date is non-compliant on its face. The single most common mistake is assuming that a permissive licence on the weights means the training data was lawfully obtained. Apache 2.0 does not clear the data. The CC licence on a collection does not clear the images in it. Verify the chain of provenance yourself, or accept the litigation risk on every clip you publish.
Common Questions
What does the Getty Images v Stability AI case mean for image-to-video models?
The case challenges Stability AI's use of Getty's watermark-stripped images in the LAION-5B training collection. A ruling against Stability AI would mean that any model trained on LAION-5B, including Stable Video Diffusion, may be an infringing derivative work. The case is in the Northern District of California, filed February 2023.
Does the EU AI Act require me to disclose my training data if I use an open-weights model?
No. The obligation is on the provider who trained the model, not on you as the user. But if you fine-tune an open-weights model on your own data, you become the provider and must comply with Article 53. If you use the model as released, the provider's disclosure obligations apply.
Is Wan 2.1 safe to use for commercial video output?
The Apache 2.0 licence on Wan 2.1 covers the software and weights. The training data is undisclosed. If a court later finds that the training data was infringing, commercial output from Wan 2.1 could be challenged. There is no lawsuit against Alibaba as of September 2026, but the risk is not zero.
What is the difference between open-weights and open-source for training data?
Open-weights gives you the model file only. Open-source also gives you the training code, the data processing pipeline, and the collection itself. No current image-to-video model is open-source in the full sense. All of them are open-weights at best, meaning the training data provenance remains undisclosed.