IL-2023-000007 - [2025] EWHC 2863 (Ch)

FACTUAL BACKGROUND

I do not understand the following description of Getty Images and their business, much of which is set forth in Getty Images’ Opening Submissions, to be controversial.

The Getty Images business was founded in 1995 through the incorporation of Getty Communications plc and, since then, has grown through a series of mergers and acquisitions, including the acquisition of the Canadian company iStockphoto, Inc. The First to Fifth Claimants are members of the Getty Images group (“the Group”). The First Claimant was incorporated under the laws of New York, the Second Claimant was incorporated under the laws of Ireland, the Third and Fourth Claimants were incorporated under the laws of England and Wales and the Fifth Claimant was incorporated under the laws of Canada.

The Group is now a pre-eminent global visual content creator and market-place. Its business involves the licensing of millions of Visual Assets including photographs, video footage and illustrations, as well as audio assets (collectively “Content”), to individuals and business users, such as newspapers, magazines, production companies, advertising agencies, banks, airlines, insurance companies and pharmaceutical companies, in more than 200 countries worldwide. Getty Images licenses the Content in a variety of ways to end users via their standard licensing agreements and, in some cases, bespoke customer licensing deals.

The Copyright Works are said to make up a substantial proportion of the Content. Getty Images assert copying during the training and development of the Model by Stability in respect of millions of Copyright Works of which the First Claimant is either the owner or exclusive licensee. As I have said, it is now accepted that the training and development of the Model did not take place in the United Kingdom and so this is not an allegation that I need to address.

The Content exists in a sophisticated curated database (“the Getty Images Database”) with a plethora of associated metadata which includes, amongst other things, the content type, the date the content was captured or created, the pixel and file size, information relating to the creator of the Content, relevant keywords and, importantly, the relevant caption. Captions can vary from being a lengthy detailed description of the Content to a short generic description. There are two main categories of Content: (i) editorial content comprising Content that is newsworthy or of public interest and depicts real life people, places and events; and (ii) creative content, comprising pre-shot stock photography, illustrations and video used for a variety of purposes.

The Content is said by Getty Images to be highly desirable for use in connection with AI and machine learning because of its high quality and because it is accompanied by content specific, detailed captions and rich metadata.

Getty Images make the Content available through websites at gettyimages.com (launched in 2001), gettyimages.co.uk, and istockphoto.com (acquired in 2006 (the “iStock Website”) (collectively the “Getty Images Websites”). The Getty Images Websites comprise hundreds of millions of Visual Assets, together with associated captions covering a broad range of subject matter. Customers are able to browse the Getty Images Websites using keywords and filters to locate their desired Content, including the associated metadata and captions. The iStock Website is searchable and contains a library of pre-shot creative content with accompanying captions, including “vector files”, namely a resolution independent depiction of a visual image that can be scaled up or down and is often used for illustrations. Each Visual Asset that is available through the Getty Images Websites has an associated page that contains a unique uniform resource locator (“URL”) pointing to a location where the image is stored, together with an “alt text” tag containing a caption for the image. The Getty Images Websites are hosted by servers and are available via both iOS and Android apps, in 23 different languages.

The Content is acquired by Getty Images in a variety of different ways: (i) by acquiring the Content outright, for example by way of an assignment from a rightsholder or the acquisition of a company with a substantial portfolio of Content; (ii) by entering into licence agreements with third party photographers and copyright owners; (iii) via photographers and videographers who are employed by a member of the Group in various jurisdictions; and (iv) through contracts with ‘stringer’ photographers who are hired to cover a specific event and paid a day rate, with the terms of the contract assigning copyright to Getty Images or granting it a perpetual exclusive licence to the Content. Many of these contracts and agreements take the form of a Unified Contributor Agreement (“Contributor Agreement”) or an iStock Artists Supply Agreement (“iStock ASA”). These are template agreements which have been updated and varied over time and which govern the terms on which the Content is licensed to Getty Images.

The Sixth Claimant is an example of a creative Content contributor. It is a US company founded in June 2005 and has an exclusive arrangement with Getty Images whereby it produces commercial imagery for Getty Images to license via the Getty Images Websites. The Sixth Claimant is the trading vehicle of the photographer and full-time employee and director, Thomas Barwick (“Mr Barwick”). It has approximately 35,000 still assets and 15,000 video assets available on the Getty Images Websites. The Sixth Claimant has entered into various iterations of both the Contributor Agreement and the iStock ASA with the First Claimant, the most recent being a Contributor Agreement dated 3 October 2023.

Getty Images own various trade marks (together the “Marks”) which relate to the Getty Images and iStock name and logo:

It is common ground that the Marks are inherently distinctive and have acquired a reputation by virtue of the extensive use made of them by Getty Images. The Getty Images logo and brand have remained largely unchanged and in constant use for the last 30 years, since the launch of Getty Images in 1995. The ISTOCK Marks have existed since at least 2003. The Marks are used in all aspects of Getty Images’ business such as trading names, websites, social media, company letterheads, building signage, email signatures and all marketing materials and merchandise. Unchallenged evidence from Getty Images shows a high volume of followers, interactions, impressions and reach on the Getty Images social media accounts during the period 2020-2022. Furthermore, editorial content is used in much of the world’s media, and includes a photo credit on any image that is used, which refers to Getty Images (and sometimes the name of the photographer). Getty Images undertakes extensive marketing, spending very significant sums every year. Consistent with the breadth and depth of its operations, in each of the years 2017-2022, the First, Second, Third and Fifth Claimants’ UK revenue generated under the Getty and ISTOCK brands has run into many millions of pounds.

Each of the Visual Assets that appears on the Getty Images Websites displays a watermark that contains one or other of the Marks: either a Getty Images watermark containing a Getty Mark or, on the iStock Website, an iStock watermark containing an ISTOCK Mark. It is Getty Images’ case that these watermarks have become iconic in their own right. If a Visual Asset is downloaded from the Getty Images Websites, it will feature the Getty Images or iStock watermark with the Marks appearing within a grey translucent banner which is overlaid on the image. In the case of still photographs, the name of the photographer will appear beneath the Mark in the watermark, as can be seen in the following example, taken by Mr Barwick:

It is only by properly licensing the Visual Asset from Getty Images that a version is made available to the customer without the Getty Images or iStock watermark on it.

In 2023, Getty Images launched its own AI software tools, developed in conjunction with NVIDIA, “Generative AI by Getty Images” and “Generative AI by iStock”, which are available to Getty Images’ customers via subscription through an application program interface (“API”) on the Getty Images Websites (“the GAI”). The GAI was trained on Getty Images’ creative pre-shot content library (which is all licensed content).

Stability was incorporated on 4 November 2019 in England and Wales. It carries on business in the field of machine learning software, including Deep Learning models for image and music generation, and large language models (“LLMs”) for the generation of text output.

Stable Diffusion is based on independent research work undertaken by academic researchers (including Professor Bjӧrn Ommer (“Professor Ommer”) and Mr Robin Rombach (“Mr Rombach”) at the Computer Vision and Learning Group (“CompVis”) at Ludwig Maximilian University of Munich (“LMU”) and IWR Heidelberg University, Germany). Mr Rombach co-authored an academic paper entitled “High Resolution Image Synthesis with Latent Diffusion Models” (“the Latent Diffusion Paper”), first published on 20 December 2021 and subsequently published in revised form on 13 April 2022. Mr Patrick Esser (“Mr Esser”) of Runway ML (“Runway”) was also involved in the publication of the Latent Diffusion Paper. This paper proposed a new method of generating images using a diffusion model which was trained on images transformed using existing trained autoencoders to a reduced-definition latent representational space. Such a latent diffusion model offered technical advantages over existing (i.e. pixel-based) diffusion models including being less resource intensive and able to generate higher resolution output images due to its efficiency.

On 21 December 2021, CompVis published source code and pre-trained model weights for its own Latent Diffusion model via a CompVis public web portal called Github, which allows users to upload and share code and data. In April 2022, CompVis published an updated version of this model on the CompVis Github page, trained by members of CompVis, including Mr Rombach. It is common ground that Stability had no responsibility for this model. It is, however, the unchallenged evidence of Stability’s technical expert in these proceedings that the Latent Diffusion Paper, together with the materials made available on the CompVis Github page, provided the structure and code for the “underlying model architecture” of the first iteration of Stable Diffusion.

Stable Diffusion was originally released further to an agreement between Mr Rombach and Mr Emad Mostaque (“Mr Mostaque”), founder and CEO of Stability, pursuant to which Stability gave Mr Rombach and another CompVis researcher access (via the internet) to cloud hosting and processing resources (“the AWS Cluster”) made available to Stability by Amazon Web Services Inc (“AWS”). The AWS Cluster was located outside the United Kingdom. Stability says that it utilised the AWS Cluster with the aim of offering such services to academic and other non-profit researchers and so to promote the development and growth of open source machine learning models. An article on the LMU website dated 1 September 2022 explains that “[i]n their project, the LMU scientists had the support of the start up Stability.Ai, on whose servers the AI model was trained”. Professor Ommer describes how “[t]his additional computing power and the extra training examples turned our AI model into one of the most powerful image synthesis algorithms”.

On 10 August 2022, Stability announced the first stage of the release of Stable Diffusion to researchers on its website, describing it as “a text-to-image model empowering billions of people to create stunning art within seconds”. The announcement went on to say that Stable Diffusion “runs on under 10GB of VRAM on consumer GPUs, generating images at 512x512 pixels in a few seconds” (Footnote: ²). It was intended that Stable Diffusion would shortly be made available to the public (i.e. on an open source basis), thereby “democratizing image generation”.

Thereafter, Stable Diffusion was released to the public in various iterations or ‘checkpoints’, namely v1.1, v1.2, v1.3 and v1.4. These versions (published together as “v1.x”) were made available for download (including in the United Kingdom) by release of the source code and model weights via the CompVis GitHub and CompVis Hugging Face web portals on or around 22 August 2022. A user accessing the Model via these portals is able to download the inference code from GitHub and the model weights from Hugging Face and set up his or her local computer to run the software, thereby running the inference offline (“the Direct Download Method”). Alternatively, a user can access the Hugging Face ‘Diffusers’ library which contains inference code and allows the user to download the model weights through a code interface (“the Diffusers Method”). This requires use of additional software including MiniConda, the Nvidia Cuda Toolkit and PyTorch. Once the Model has been downloaded using either of these methods, expert users have the opportunity to modify the code and to run additional training with their own data.

At the same time, Stability announced the “Stable Diffusion Public Release” on its website, providing a link to the Model Card and weights and also recommending the use of “DreamStudio”, a commercial platform hosted outside the UK which enables users in the UK and elsewhere to access Stable Diffusion v1.4 without the need to download the Model; users are able to run the inference on Stability’s computing system using a web interface. Access to DreamStudio was provided by Stability via a public test version of DreamStudio referred to as DreamStudio Beta, available via Stability’s website and also accessible at beta.dreamstudio.ai. The Model Cards for v1.1, v1.2, v1.3 and v1.4 explain, amongst other things, that (each version of) the Model is a Latent Diffusion Model that uses a fixed pretrained text encoder (CLIP ViT-L/14) and (under the heading “Limitations”) that it has been trained on a “large-scale dataset LAION-5B which contains adult material”.

Stability’s announcement on 22 August 2022 anticipated the continuing release of optimized versions of the Model together with the provision of API access (Footnote: ³). A further, unofficial, version of the Model (v1.5) was released in or around October 2022. Although originally forming part of Getty Images’ pleaded case, references to this version have been excised from the pleadings and (although it is mentioned in some of the evidence) it no longer forms part of that case.

At some time in the Autumn of 2022, after the release of Stable Diffusion v1.x, Mr Rombach joined Stability as an employee and Head of Research. He was based in Germany.

A TechCrunch article published on 17 October 2022 quotes Mr Mostaque as saying that DreamStudio had more than 1.5 million users “who’ve created over 170 million images”. The article also states that, according to a Stability press release, the open source version of Stable Diffusion had been downloaded more than 200,000 times.

On 24 November 2022, Stability launched Stable Diffusion 2.0 (“v2.0”), making available the Python source code for v2.0 on Stability Github and the pre-trained model weights on Stability Hugging Face on an open source basis. Stability AI is named on the licence (CreativeML Open RAIL++-M) as a copyright owner. The Stability Github page explains that v2.0 has been “trained from scratch”. In other words, it did not use any of the model weights obtained from the process of training the version 1.x Models, but represented an entirely fresh start, using a new training dataset and a different text encoder model. Stability released an API Platform for v2.0 on 25 November 2022, thereby enabling users to run the inference on Stability’s computing network using an API without the need to download the Model. Stable Diffusion v2.0 was also made available on DreamStudio from around 1 December 2022 until early 2024.

Version 2.1 (“v2.1”) of Stable Diffusion was released in December 2022 as a further development of the existing v2.0 checkpoint, using additional steps but tweaking the settings of the NSFW Filter. Further checkpoints of v2.0 are referred to together as v2.x. The Model Cards for v2.x Models explain amongst other things, that (each version of) the Model is a Latent Diffusion Model that uses a fixed pretrained text encoder (OpenCLIP-ViT/H) and (under the heading “Limitations”) that the Model was trained “on a subset of the large-scale dataset LAION-5B, which contains adult, violent and sexual content”. In contrast to the v1.x Model Cards, the v2.0 Model Card goes on to say that “[t]o partially mitigate this, we have filtered the dataset using LAION’s NSFW detector…”, which is later said to produce a “p_unsafe score of 0.1 (conservative)”. The acronym NSFW is short for Not Safe For Work (“NSFW”) material.

On 13 April 2023, Stability announced the release of Stable Diffusion XL Beta and thereafter made available further model checkpoints in this series (together “SD XL”), capable of outputting higher resolution images of 1024x1024 pixels. Thus in June, July and November 2023, Stability released Stable Diffusion SD XL 0.9, XL Base 1.0 and XL Turbo. SD XL Beta was made available through an API and the DreamStudio platform, while the model weights and associated source code for versions SD XL 0.9, XL Base 1.0 and XL Turbo were made available for download on Hugging Face (and in one case on Git Hub). SD XL Turbo makes use of a technique called adversarial diffusion distillation (“ADD”) and combines aspects of both generative adversarial networks (“GANS”) and diffusion models to create high quality images quickly.

The SD XL Model Cards explain amongst other things, that (each version of) the Model is a Latent Diffusion Model that uses two fixed pretrained text encoders (OpenCLIP-ViT/G and CLIP-ViT/L). The SD XL Model Card explains that it is a distilled version of SD XL 1.0 and that it is based on the novel training method called ADD. These Model Cards no longer record how these Models were trained and, although they continue to identify “Limitations”, they do not say anything about the presence of adult, violent or sexual content in the training dataset.

Each of the Model Cards for v1.x, v2.x and SD XL identifies “Out-of-Scope Use” in the following terms: “[t]he model was not trained to be factual or true representations of people or events and therefore using the model to generate such content is out-of-scope for the abilities of this model”.

On 18 July 2023, Stability launched the Developer Platform API (“the Developer Platform”) at https://platform.stability.ai designed as a reboot of the prior API Platform. This platform makes APIs available to subscribers, thereby enabling subscribers’ applications to request services from various versions of the Model, remotely hosted outside the UK by Stability. It does not enable subscribers to download model weights or associated source code but instead enables them to run inference on AWS. Originally Stable Diffusion v1.x and v2.x variants were made available through APIs but support for these variants has since been discontinued.

In addition to the Stable Diffusion XL Models, Stability also developed a further Stable Diffusion Model known as Stable Diffusion 1.6 (“v1.6”). This Model is based on the architecture of Stable Diffusion XL and is optimized to generate images which are 512x512 in resolution. V1.6 was made available for use in around November 2023. The model weights for v1.6 have not been published and it would appear that v1.6 is not available to download. The Developer Platform now facilitates access to SD XL and to v1.6.

It is common ground that all versions of Stable Diffusion were trained using various subsets of the LAION-5B dataset (“LAION-5B” and “LAION-Subsets”) assembled by LAION e.V. (“LAION”), a not-for-profit organisation registered in Hamburg, Germany. LAION researchers first announced the publication of a dataset known as LAION-400M (“LAION-400M”) on 20 August 2021. LAION-400M, as its name suggests, comprised approximately 400 million CLIP-filtered (Contrastive Language-Image Pre-training) URL-text pairs (Footnote: ⁴) (URLs with associated alt-text captions) and was created by LAION by filtering the Common Crawl public web archive for images and storing these alongside their HTML alt-text (hidden text associated with the image on a web page). The Latent Diffusion model developed by CompVis was trained on LAION-400M.

LAION researchers announced the publication of LAION-5B and the LAION-5B Subsets on 31 March 2022. Stability admits that at or around this time, it donated support to LAION in the form of hosting services comprising access to the AWS Cluster. The available evidence suggests that LAION-5B was created in a similar manner to LAION-4400M, but with the addition of further filtering steps. LAION-5B comprises metadata including 5.85 billion URL-text pairs. The LAION-Subsets (said by Getty Images to contain millions of Copyright Works) were produced by filtering LAION-5B against specific sets of requirements and together comprise approximately 3 billion URL-text pairs from the LAION 5B dataset. Information contained in the Model Cards for the various iterations of Stable Diffusion indicates that these Models were trained on LAION-Subsets which included LAION-2B-en, LAION-improved-aesthetics, LAION-aesthetics v2 5+, LAION-high-resolution and LAION-A. The Model Card for v2.0 gives a ‘shout out’ to “The DeepFloyd team at Stability AI, for creating the subset of LAION-5B dataset used to train the model”.

It is common ground that for training purposes it would have been necessary to download the images from the URLs in the LAION-Subsets – a process known as materialisation. Stability says that the training process involved downloading and storing copies of each image obtained from the URLs in the relevant dataset on Amazon Simple Storage Service (“Amazon S3”) on the AWS Cluster, retrieving those images and then making temporary copies of them in the VRAM of the GPUs performing the training on the AWS Cluster. Owing to the abandonment of the Training and Development Claim, Getty Images no longer seek to advance a case that this training process took place in the UK. Stability’s case has always been that training took place on the AWS Cluster outside the UK, but it has sought no declaration to this effect and accordingly I need make no finding on the point.

Stability accepts that “at least some” LAION-Subsets contain URLs referencing images on the Getty Images Websites and that “at least some” images from the Getty Images Websites were used during the training of Stable Diffusion. However, there is a dispute over whether this court should make any findings as to the number of images from the Getty Images Websites which were so used, it being Stability’s case that the particular images used will depend on the starting dataset and the filters applied to it for each training run. Filters, such as the NSFW filter to which I have already referred, may be applied to the training dataset to remove (or attempt to remove) undesirable images. I shall return to this in due course.

Stability also accepts that Stable Diffusion may be used to generate synthetic images which include the Marks in the form of Getty Images’ watermarks. However, it contends, broadly, that (i) where such images are generated by a user, this is the result of third party use of Stable Diffusion and not a statement or commercial communication attributable to Stability, or for which Stability is responsible in law; (ii) any such generation of watermarks does not amount to use of any sign in those watermarks in the course of trade; and that (iii) “watermarked” synthetic image outputs will only be generated with wilful contrivance of the user.