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

Fecha: 04-Nov-2025

THE WITNESSES AND EVIDENCE

(D)

THE WITNESSES AND EVIDENCE

77.

The court heard from a total of 19 witnesses over the course of only a few days. They were cross-examined with concision and focus. Some of their evidence is of course no longer relevant to the issues that remain to be decided. With only one exception identified below, I formed the impression that all of the witnesses were honest and that they were all trying their best to assist the court.

Getty Images’ Witnesses of Fact

78.

Getty Images called 10 witnesses:

Heather Cameron, Director of Legal Risk Mitigation at Getty Images (Seattle) Inc, a management company which supports various entities in the Getty Images Corporate Group, including the First Claimant. Ms Cameron has worked at Getty Images for 26 years. She has provided four witness statements in these proceedings which (amongst other things) cover the history of Getty Images, the Getty Images Websites, the Getty Images Database and the investment made in it, the Getty Images and iStock watermarks; Getty Images’ AI Generator, damage to Getty Images’ business and Getty Images licensing practice. In her fourth statement, Ms Cameron gives evidence of an analysis of Copyright Works and Visual Assets identified in the LAION datasets, designed to identify the number of Copyright Works and Visual Assets on which Stable Diffusion was trained. Although Stability contends that no weight can be placed on the evidence in her fourth statement, it makes no personal criticism of Ms Cameron in relation to that evidence.

ii)

Andrea Gagliano, Senior Director of Artificial Intelligence and Machine Learning at Getty Images (Seattle) Inc. Ms Gagliano has worked at Getty Images since 2018. In her statement, Ms Gagliano addresses (amongst other things) her knowledge of, and investigation into, Stability and Stable Diffusion, including the datasets on which Stable Diffusion was trained and the means of accessing Stable Diffusion and generating an image from it. Ms Gagliano also gives evidence about Getty Images’ own AI model, GAI.

iii)

David Stanley, Vice President of Marketing at the Third Claimant, who joined Getty Images in July 2005. In his statement, Mr Stanley gives evidence about the ISTOCK and Getty Images logos and branding, watermarks, his concerns about Getty Images’ reputation, and the potential confusion and damage caused by Stable Diffusion.

iv)

Tanja Malnar, Localization Project Manager for the Fifth Claimant. Ms Malnar gives evidence in her statement of the process by which contributors uploaded and licensed content to the Fifth Claimant through its website prior to 2017. Ms Malnar gave her evidence remotely from Canada.

Emma Varty, a solicitor at Fieldfisher. In her first statement Ms Varty gives evidence as to the conduct of Getty Images’ various experiments and the synthetic outputs generated. In her second statement Ms Varty explains the manner in which she identified the prompts used to generate Getty Images/ISTOCK watermarks in connection with the Getty Watermark Experiments. In her third statement, Ms Varty addresses the enrolment process that has been in place since 2017 for Getty Images’ exclusively licensed contributors. Although Stability criticises the way in which Getty Images went about conducting the various experiments described by Ms Varty, Mr Cuddigan very properly did not criticise Ms Varty herself for the approach that had been taken.

vi)

Mr Barwick, a photographer and President of the Sixth Claimant and author of Sample Works E-K.

vii)

two freelance photographers: Alberto Rodriguez who captured SOCI Work A8; and Alexander Livesey, who captured SOCI Works A5 and A6. Both gave their evidence remotely from the USA.

viii)

Andreas Rentz, a staff photographer for Getty Images DevCo Deutschland GmbH who gave evidence with the assistance of an interpreter. Mr Rentz captured SOCI Works A10 and A11.

ix)

Gregory Shamus, a staff photographer employed by the First Claimant who captured SOCI Work A7. Mr Shamus gave his evidence remotely from the USA.

79.

In addition, Getty Images rely upon statements from 3 witnesses who were not required to give evidence and whose evidence was therefore not challenged: Matthew Prichard, Senior Director of International Tax employed by Getty Images (Seattle), Inc; Tessa O’Neill, a Director of the Second Claimant since 2015; and Jonathan Lockwood, Vice President, Corporate Counsel at the Third Claimant. Given the narrowing of Getty Images’ case at trial, very little of the evidence in these statements remains relevant.

Getty Images’ CEA Notices

80.

Getty Images have served four Civil Evidence Act Notices (the “Getty CEA Notices”) identifying documents on which they intend to rely at trial:

An updated CEA Notice dated 11 April 2025 attaching documents relied upon by Getty Images on the subjects of: (a) Getty Images’ business; (b) copyright subsistence and ownership in connection with the SOCI Works and the Sample Works; (c) the Second Claimant and the issue of database right subsistence and ownership; (d) Getty Images’ Trade Marks; and (e) Stability and Stable Diffusion.

ii)

A CEA Notice dated 10 March 2025 attaching documents relating to the exclusively licensed works.

iii)

A CEA Notice dated 24 March 2025 attaching documents relevant to the SOCI Works.

iv)

A CEA Notice dated 27 May 2025 attaching screenshots showing the number of downloads of various versions of Stable Diffusion from Hugging Face.

Stability’s Witnesses of Fact

81.

Stability called 9 witnesses:

Cedric Wagrez, Head of Data Strategy and Operations at Stability and an employee of Stability AI Japan KK. Mr Wagrez joined Stability in October 2023. In his first statement, Mr Wagrez explains his limited knowledge of the datasets used to train Stable Diffusion models released prior to his involvement with Stability. In his second statement he explains the investigations performed by Stability to establish whether certain images were used in training various versions of Stable Diffusion and he addresses the question of the number of Copyright Works used in training those versions of Stable Diffusion. Although Mr Wagrez’s evidence was plainly honest (and Getty Images does not seek to suggest otherwise), I accept that the preparation of his first witness statement involved a very high level of generality which he then sought to remedy in his second statement.

ii)

Daria Bakshandaeva and Mikhail Konstantinov, research scientists who joined Stability as employees in the spring of 2022 and worked as part of what became known as the DeepFloyd Team. Both gave their evidence with the assistance of an interpreter. Although the importance of their evidence is now much reduced, I should observe that I agree with Getty Images that Mr Konstantinov was not a candid witness, that he was overly careful with his answers and that he resisted making appropriate concessions. This did not apply to Ms Bakshandaeva.

iii)

Dennis Niedworok, Head of Applied Research at Stability. Mr Niedworok joined Stability in December 2022 as a ‘model fine-tuner’ and became a member of the Applied team. In his statement he describes the process of “fine-tuning” a model and explains the work he was involved with at Stability.

iv)

Kate Hodesdon, a machine learning engineer who has been employed by Stability in the UK since February 2023. Ms Hodesdon works for the Inference Team, which is led by Anilkumar Bandari, Stability’s Director of Engineering. In her statement Ms Hodesdon explains her work on inference, i.e. the process of running models once they have been trained.

Nojus Cebatoriunas, an Information Technology specialist at Stability who joined as an employee in 2021 and is based in London. While no criticism is made of his honesty by Getty Images, they contend (and Mr Cebatoriunas accepted in cross-examination) that the process he undertook to search for laptops available to Stability in the UK was quite obviously flawed. However, this is of little significance in the context of the remaining claims.

vi)

Scott Trowbridge, VP of Business Development who has been employed by Stability in the UK since March 2023. His role involves hiring individuals and supporting any sales-related or partnership opportunities. In his statement, Mr Trowbridge gives evidence about Stability’s commercial customers and its licensing practices.

vii)

Stephan Auerhahn, was formerly a Principal Infrastructure Engineer in the API / Data / Infrastructure team at Stability based at all times in Los Angeles. Mr Auerhahn joined Stability in September 2022. In his statement he deals with data location and storage services, how inference works for the Developer Platform and DreamStudio, and the filtering used on those platforms. Mr Auerhahn gave his evidence remotely from the USA.

viii)

Anilkumar Bandari, Director of Engineering at Stability based in Seattle. Mr Bandari joined Stability in August 2023. In his statement, Mr Bandari gives evidence about the Engineering team. Mr Bandari gave his evidence remotely from the USA.

82.

In addition, Stability relies upon the statements of 7 witnesses who were not required to give evidence and whose evidence was therefore not challenged. Five of these witnesses are existing employees of Stability: Chantelle Palmer, Head of People at Stability since June 2023; Peter O’Donoghue, Chief Financial Officer of Stability since the Summer of 2022; Reshinth Adithyam, a research scientist who joined Stability as an employee in December 2022; Zachary Evans, Head of Audio Research and an employee of Stability since June 2022; Christian Burrows, Head Security Engineer and an employee of Stability since March 2024. None of this evidence is of any significance in the context of Getty Images’ narrowed case.

83.

The remaining two witnesses whose evidence was not challenged by Getty Images are associates at Bird & Bird:

Aneesah Kabba-Kamara, whose first statement explains: (a) how inference can be undertaken using Stable Diffusion v.1.4, v.2.1 and XL 1.0 running on a local machine; (b) how inference can be undertaken using Stable Diffusion v.1.6 via the Developer Platform; and (c) how inference can be undertaken using Stable Diffusion v.1.6 and XL 1.0 via DreamStudio. In her second short statement, Ms Kabba-Kamara addresses an omission from the information provided in her first statement.

ii)

William Wortley, who gives evidence (amongst other things) about the presence of certain images in the LAION-2B dataset which contain similar material to SOCI images A7, A10 and A11. This evidence is no longer of any relevance to the remaining issues.

84.

Stability chose not to rely on a witness statement from Oliver Toromanoff, a former Lead Software Engineer who has now left his employment with Stability.

85.

Until the day before he was due to give oral evidence, it appeared that Stability would also be relying upon the evidence of Richard Vencu, a former Stability employee who was involved with the training infrastructure at Stability between February 2022 and September 2024. One of Mr Vencu’s two witness statements was relied upon by Stability’s technical expert in his report and various references were made to Mr Vencu’s evidence in Stability’s Opening Skeleton.

86.

However, Stability took the decision not to call Mr Vencu and so is no longer able to rely on the two statements signed by him in the proceedings (although Stability still relies upon documents that were exhibited to his statements, without objection from Getty Images). No reasons were provided to the court at the time for this decision. However during closing submissions Stability provided the court with a letter from Bird & Bird dated 27 June 2025 in which factors relevant to the decision not to rely upon Mr Vencu’s evidence were explained.

87.

In summary, these were that Mr Vencu’s evidence related solely to Stability’s computing resources, that Getty Images had dropped its claim in respect of “Final Training” of Stable Diffusion on Visual Assets in the UK prior to the trial and that his activities prior to working for Stability (when he had worked for LAION) were irrelevant to the matters before the Court. Furthermore, the letter relied upon the potential for Mr Vencu to be cross examined about an assertion (which Stability had itself made in cross-examination of Ms Gagliano) that the LAION database had originally included CSAM material which had been removed in RE-LAION (a more recent dataset compiled by LAION). Bird & Bird explained that the assessment required in order to determine whether it was fair and proper to ask Mr Vencu to attend for cross examination included consideration of whether Mr Vencu would need an opportunity to obtain his own legal advice prior to giving evidence.

88.

Getty Images invites me to draw various adverse inferences by reason of Mr Vencu’s absence from the trial, to which I shall return in due course.

89.

In closing, Getty Images suggested that the court might draw adverse inferences by reason of the failure on the part of Stability to rely upon evidence from either Mr Mostaque or Tom Mason, Stability’s Chief Technical Officer. However, in my judgment, this would be inappropriate. Aside from the fact that the relevance of their potential evidence was said by Getty Images in its opening submissions to go specifically to the Training and Development Claim (now abandoned), no attempt was made during the trial to establish whether these potential witnesses were alive, well and available to give evidence – an obviously relevant consideration (see Efobi v Royal Mail Group Ltd [2021] UKSC 33, per Lord Leggatt JSC at [41]). In closing, Getty Images did not seek to identify in any detail the relevant evidence that these missing witnesses would have been able to give, just as they did not seek to identify other existing evidence that might have a bearing on the points on which these missing witnesses might have given evidence. Accordingly I decline to draw any inferences by reason of the absence of these witnesses.

Stability’s CEA Notices

90.

Stability relies upon two CEA Notices (the “Stability CEA Notices”):

a CEA Notice dated 23 January 2025, attaching a capture of the DiffusionDB Webpage. DiffusionDB is a dataset that (as at that date) contained 14 million images generated by Stable Diffusion using text to image prompts specified by real users. It was created by scraping user-generated images on the public Stable Diffusion Discord platform (a platform that allows users to generate images through an API); and

ii)

a CEA Notice dated 2 April 2025 attaching captures of the iStock Website, some documents from Companies House and a Photo Archive News Article.

The Technical Experts

91.

Each side adduced evidence from eminent experts in the field of AI models (together “the Experts”):

Getty Images rely upon the evidence of Professor Hany Farid, a Professor at the University of California, Berkeley with a joint appointment in Electrical Engineering & Computer Sciences and the School of Information. He is a member of the Berkeley Artificial Intelligence Research Lab and the Center for Innovation in Vision and Optics. He has a PhD from the University of Pennsylvania. He is co-founder of two companies dedicated (amongst other things) to developing software to detect manipulated images. His academic research focuses on digital forensics, forensic science, misinformation, image analysis and human perception. Over his 25 year academic career he has published over 200 academic papers in these areas of scholarship. He is the recipient of an Alfred P Sloan Fellowship, a John Simon Guggenheim Fellowship and is a Fellow of the National Academy of Inventors.

ii)

Stability relies upon the evidence of Professor Thomas Brox, a Professor for Pattern Recognition and Image Processing and Head of the Computer Vision Group at the University of Freiburg in Germany for the last 15 years. He has worked in the field of image processing and computer vision since 2002. He has an engineering doctorate from the Saarland University in Germany. He published some of the first works on deep learning that went beyond classification and he is known in the field for creating the convolutional encoder-decoder architecture known as the U-Net, which has become standard architecture for a very broad set of image processing tasks and is used in Stable Diffusion and other latent diffusion models. Professor Brox has worked as a part-time employee for AWS and is co-author of around 300 academic papers.

92.

The Experts agreed a Joint Statement setting out the many areas on which they were able to agree. I asked at the PTR that they identify the remaining areas on which they disagreed and they then produced a further Joint Statement entitled “Statement of Areas of Disagreement”. In fact this latter document merely evidences the extent of the yet further agreement between the Experts. Nonetheless, both Experts were cross-examined and it became apparent during that cross-examination that there are areas of nuance in their respective reports on which they are not in complete agreement.

93.

In reality, however, there are no conflicts of substance between the Experts in relation to any of the issues left in dispute. They both gave clear and helpful evidence to the court and neither side sought to (or needed to) criticise the opposing Expert.

94.

In closing, Getty Images took issue with the instructions that had been provided to Professor Brox, submitting that they were in some respects deficient, a submission with which I agree. Specifically, Professor Brox had not been shown evidence which was capable of being relevant to the opinions he expressed in his report. In the event, however, this omission does not appear to me to undermine the evidence given by Professor Brox, whose responses under cross-examination were quite obviously direct, cogent and technically sound. As Stability submitted in closing (and consistent with his duties as an independent expert), Professor Brox readily agreed with propositions that were put to him as appropriate without regard to whether they might support or detract from Stability’s case.

95.

Shortly prior to the close of the trial and at my request, the Experts very helpfully produced an Agreed Technical Primer which provides a summary of the technology that is most relevant to the issues arising in this case. In so far as I have sought to summarise the technology behind Stable Diffusion in various parts of this judgment, I have drawn on the Agreed Technical Primer and am extremely grateful to the Experts for the additional work they have undertaken in putting it together.

96.

Unless the contrary is stated, where I record the views of one or other expert in this judgment I understand those views to be unchallenged and I accept them.

Evidence of New York Law

97.

Prior to the trial, the parties each served Statements of Case on New York law, relevant to the Licensing Issue, but neither side sought to rely on any expert evidence of New York law at trial. Instead, the parties have been able to agree upon a neutral statement of New York law, to which I shall return in due course.

The Trial Bundle

98.

Although initially the subject of a dispute, it is now agreed between the parties that all disclosure documents included in the trial bundle are admissible as evidence of the truth of their contents pursuant to CPR PD 32 paragraph 27.2.