Getty sues Stability AI over 12 million training images and the watermarks that came out of the model

2023-02-03Copyright

Getty Images (US), Inc. v. Stability AI, Inc., No. 1:23-cv-00135 (D. Del., complaint filed Feb. 3, 2023); Getty Images statement of January 17, 2023 announcing proceedings against Stability AI in the High Court of Justice in London

Getty Images (US), Inc. filed a complaint in the United States District Court for the District of Delaware against Stability AI, Inc., developer of the Stable Diffusion model and the DreamStudio interface. It alleges that Stability AI "has copied more than 12 million photographs from Getty Images' collection, along with the associated captions and metadata," by following links in the LAION datasets to Getty's websites, and that Getty has identified more than 12 million such links. It pleads 7,216 registered works in Exhibit A and a registered database compilation, Registration No. TXu002346096. It alleges that the model "sometimes memorizes and regenerates specific images," citing two research papers. The watermark allegations run both directions: Stability removed Getty watermarks from some images in the course of copying them, and the model "frequently generates output bearing a modified version of the Getty Images watermark," including on "bizarre or grotesque" images. Eight claims follow: copyright infringement; providing false copyright management information under Section 1202(a) of Title 17 of the United States Code; removing it under Section 1202(b); trademark infringement, unfair competition and dilution under the Lanham Act; and Delaware deceptive trade practices and dilution. The prayer asks for destruction of "all versions of Stable Diffusion trained using Getty Images' content without permission." Getty announced on January 17, 2023 that it had commenced proceedings in the High Court of Justice in London.

What it changesThe pleading treats a watermark in the output as proof of what the model memorized, and that move cuts into the patent program on two sides. On the defensive side, a company that cannot say which datasets trained which model version has no answer to a complaint built from the training set's own link list, so training-data provenance records (source, license, date, version, preprocessing) belong in the engineering record now. On the drafting side, those same records are the factual basis for the training description a specification needs under Section 112 of Title 35 of the United States Code; an application that says a model was trained on image-text pairs without saying which, how many or how they were filtered supports neither a claim to the trained model nor a claim to the training method. For the trade secret program, the same record is the inventory of what the company owns. None of this is a ruling; it is a complaint, and the fair use question it raises is unlitigated.

More on Copyright and training data · Litigation, damages and PTAB challenges

← All recent developments · Search all developments

Burbage Law is the practice of Jon Burbage, a registered patent attorney. 342 United States patents have issued from applications he wrote; the full list is on the Patents page. More about Jon Burbage.

This is general information about a published decision, agency guidance or piece of legislation. It is not legal advice and does not create an attorney-client relationship. Items concerning offices outside the United States are reported for information only and are not counsel on the law of those countries.