Getty Images v. Stability AI is a claim in the English High Court by Getty Images, which licenses photographs and other visual content, against Stability AI Limited, the company behind the Stable Diffusion image-generation model. Getty alleged that Stability scraped millions of its images from Getty websites to train Stable Diffusion. The judgment handed down on 2025-11-04 held that Getty’s secondary copyright infringement claim failed and that its trade mark claims succeeded only in a few limited respects.[1]
Background
Stable Diffusion is a latent diffusion model trained on large datasets of images and captions created by crawling and scraping the internet. The court noted that inference does not require training data and that the model does not store training data.[1] Version 1.x was released through CompVis GitHub and Hugging Face pages, and Version 2.0, from November 2022, was released with Stability named on the CreativeML Open RAIL++-M licence as copyright owner. Stability also offered the models through DreamStudio and a developer API governed by its own terms of service.[1] The court found that, on the Hugging Face and GitHub pages and in the API terms, there was no express requirement that a user read or accept the CreativeML licence terms.[1]
The dispute
Getty pleaded primary and secondary copyright infringement, database right infringement, trade mark infringement and passing off. Shortly before closing submissions it abandoned the training and development claim, because there was no evidence that training took place in the United Kingdom, the outputs claim, because the relevant prompts had been blocked, and the database right claim.[1] What remained was the secondary infringement claim, under which Getty said the model is an infringing copy because making its weights in the UK would have infringed, the trade mark claims over synthetic images bearing Getty and iStock watermarks, and passing off.[1] The trial also tested, by reference to sample agreements and New York law, whether licences relied on by Getty were exclusive licences under section 92 of the Copyright, Designs and Patents Act 1988.[1]
Decision or outcome
Mrs Justice Joanna Smith dismissed the secondary infringement claim: although an “article” may be intangible, a model that does not store or reproduce any copyright works, and never has, is not an “infringing copy”.[1] On trade marks, Getty succeeded under section 10(1) and 10(2) of the Trade Marks Act 1994 for iStock watermarks generated by users of v1.x through DreamStudio and the developer platform, and under section 10(2) for a Getty watermark generated by v2.1. The court described these findings as “historic and extremely limited in scope”. The section 10(3) claim was dismissed and the court declined to address passing off.[1] On the licensing issue, it held that some sample licences were not exclusive licences under section 92 and others were. It made no finding on the number of works used in training and found additional damages unjustified.[1] The cited sources do not cover any appeal.
Significance for software licensing and SAM practice
This is a first-instance decision on specific facts, and the court itself said its findings were narrow. For organisations that use generative AI services, it illustrates that the legal position of a model’s weights, the licence under which they are released, and the terms of the service through which they are accessed are separate questions. The case is relevant to generative AI licensing. For United States decisions on AI training, see Kadrey v. Meta Platforms, Bartz v. Anthropic and Thomson Reuters v. Ross Intelligence.
Lessons learned
- Where training happens matters. The training claim was abandoned for lack of evidence that training took place in the UK.[1]
- A model is not necessarily a copy. The court held that Stable Diffusion’s weights do not store or reproduce copyright works and are not an infringing copy.[1]
- Exclusive licence status turns on the wording. Sample licences were analysed one by one under New York law and reached different outcomes under section 92.[1]
- Open model licences may not be accepted at the point of access. On the pages and terms examined, the court found no express requirement to read or accept the CreativeML licence.[1]