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

LEGAL RESPONSIBILITY FOR STABLE DIFFUSION v1.X

Before I turn to look at the detail of the claims made by Getty Images in these proceedings, I must first determine a dispute between the parties over whether Stable Diffusion v1.x was released and/or made available by Stability via the CompVis GitHub and CompVis Hugging Face platforms. This dispute was captured by Issues 13-15 of the List of Issues which, broadly, focused respectively on (13) the development of v1.x; (14) the publication and making available of v1.x on these platforms; and (15) the question of where all versions of Stable Diffusion have been hosted.

As to the latter, it is accepted by Getty Images that the evidence at trial has established that DreamStudio and the Developer Platform are and were hosted outside the United Kingdom and that there is no clear evidence as to where the Github and Hugging Face platforms are hosted. Ms Lane confirmed in closing that there is now no need for the court to make findings on this issue.

Stability accepts (as I have recorded earlier in this judgment) that it made Stable Diffusion v1.x available to the public (for a limited time) via an API Platform and also that it made v1.4 available to the public via DreamStudio Beta – acts which are capable of giving rise to liability for trade mark infringement and passing off. However, I understand it to be common ground that access to the public via DreamStudio and the API Platform would not involve downloading the Model – an important consideration in the context of the Secondary Infringement Claim where Getty Images argues that it is the downloading of the Model in the UK that amounts to importation. Users would only download the Model (or more accurately the relevant Model Card, model weights and source code) if they gained access to it via the Hugging Face or GitHub platforms.

Thus issue 14 specifically concerns the question of whether Stability published and/or made available Stable Diffusion v1.1 to v1.4 on these platforms in August 2022, when v1.x was first released. Stability accepts that it provided computing resources to CompVis via its AWS Cluster and that it announced the release of Stable Diffusion on its website, but it says that each checkpoint of v1.x was made available by CompVis itself on the Compvis Hugging Face and CompVis GitHub pages, which were under the control of CompVis. Thus Stability denies that it is responsible for trade mark infringement/passing off and acts of secondary infringement of copyright (assuming they took place) in respect of these versions of the Model when accessed from these platforms. If Stability is right on this score, then it is accepted by Getty Images that it cannot succeed in its claim of secondary infringement of copyright in relation to v1.1 to v1.4 and it cannot succeed in relation to its claims of trade mark infringement and passing off in respect of any infringing watermarks generated using these platforms.

Save that Ms Gagliano confirmed in her oral evidence that v1.x was located on CompVis pages on the Hugging Face and GitHub platforms, there was no other witness evidence to assist on this issue. Getty Images relies primarily upon an admission from Stability in its Defence that it “launched” Stable Diffusion in August 2022, together with an analysis of the available contemporary documents, both as to the development and subsequent release of v1.x.

I need deal with the development of v1.x only briefly. Getty Images’ pleaded case is that Stability would not have given access to the AWS Cluster to Mr Rombach and/or CompVis for the purpose of developing an open source diffusion model (or any other AI model) “if the development was not understood to be done (wholly, jointly or in some other way)” on Stability’s behalf or for Stability’s benefit. Stability maintains that the investment was purely speculative and that “there is no indication that anything came in return for it other than goodwill”. It points out that Mr Rombach ultimately became a Stability employee and worked on v2.x.

Notwithstanding the lack of witness evidence in support of Stability’s factual position, Getty Images’ pleading alone certainly does not provide any real basis for a finding that Stability was itself directly responsible for the development of v1.x. Furthermore, the documents on which Getty Images rely in their written opening submissions provide no support for such a case. These include (i) Chat messages from May and June 2022 involving Mr Mostaque (identified as “emad”) and Mr Rombach together with researchers at CompVis and Stability employees; and (ii) a Chat message from February 2023 between employees of Stability. These Chat messages establish no more than that by June 2022 Stability was involved in discussions with Mr Rombach, Mr Esser (identified as “pesser”) and the CompVis team in connection with the training and development of the Model; a proposition which appears to be accepted by Getty Images, which assert in their opening submissions that these documents show that Stability “was involved in and/or assisted” in the process of development prior to launch. The February 2023 Chat (which involves a discussion on what Stability’s employees viewed as misinformation contained in an article published in MIT Technology Review) takes matters no further; although it suggests that Stability did more than merely pay for use of the AWS Cluster in connection with the development of Stable Diffusion, it is inconclusive as to the extent of any direct involvement on the part of Stability.

In written closing submissions, Getty Images suggested that it might reasonably be inferred that the training datasets used to train Stable Diffusion v1.x are stored (in whatever form) within what have been referred to as “Stability’s S3 buckets” (“the AWS S3 Bucket”). While this might well be an available inference given that it is common ground that CompVis used Stability’s AWS computing resources to train v1.x and given that the available evidence suggests that Stability used the AWS S3 Bucket for the purposes of storage, I do not see that this elevates Stability’s involvement beyond mere assistance or involvement.

Absent a plea by Getty Images that Stability is jointly liable for any tortious acts of Mr Rombach and CompVis (and/or Runway) - a plea which was expressly disavowed in the recital to an Order of 28 March 2025 - mere “involvement” or “assistance” on the part of Stability in the development of the Model is plainly insufficient to give rise to any tortious liability (in the form of infringement of statutory intellectual property rights) that may arise by reason of its subsequent release to the public. As Stability submits, it is not enough for Stability to have facilitated the development of the Model by enabling access to (and providing the funding for) the AWS Cluster; it would also not be enough for Stability to have directed others to carry out any relevant act, or to have commissioned it, or even to have procured it – these are not acts which give rise to direct tortious liability in the case of statutory torts such as copyright or trade mark infringement (see Lifestyle Equities CV v Ahmed [2024] 2 WLR per Leggatt JSC at [92]-[95]).

In oral closing, I did not understand Getty Images to dispute Stability’s submissions on this point or seriously to maintain that issue 13 (development of the Model) was determinative of, or relevant to, any of the remaining causes of action.

Turning then to issue 14 (responsibility for release of v1.x), once again, Getty Images rely upon contemporaneous documents in support of their case that Stable Diffusion v1.x was released and/or made available for download via Hugging Face and GitHub as Stability’s own product, for which it was responsible and in respect of which it has direct liability. Specifically they rely upon the following documents in the period prior to 10 August 2022, when Stability announced the launch of Stable Diffusion on its website (emboldened words are emphasised by Getty Images in their submissions):

Getty Images contend that these documents show Stability taking “active steps” to raise funds commercially to exploit Stable Diffusion, treating Stable Diffusion as its own commercial product which it could launch and make available to the public as and when it saw fit, and making its own decisions about branding, packaging and marketing of the different versions of v1.x. This, say Getty Images, is indicative of an entity with control over and/or ultimate responsibility for the release of the product.

On a careful reading of these documents, I tend to disagree. Aside from the fact that Getty Images has no pleaded case of control or responsibility based on “active steps” in relation to branding, packaging, marketing or commercial exploitation, it is not at all clear to me that Getty Images’ interpretation of these documents is correct. I have already found on the evidence that Stability was involved in the development of Stable Diffusion together with Compvis and these documents are plainly supportive of that conclusion. But they say nothing determinative on the subject of whether Stability was in fact in control of, and responsible for, the release of Stable Diffusion v1.x on the CompVis Hugging Face and GitHub pages.

Certainly the documents establish a narrative of ongoing collaboration between Stability, Compvis and others – i.e. Mr Schuhmann’s express understanding that Stable Diffusion would be announced as a “cooperative Stability-EAI-LAION-Compvis effort” (which is not disputed by Mr Mostaque in his Chat reply, or by anyone else). The use of language referring to “our” Model and to actions that “we” will be taking is language that collaborators, working together, would use. Similarly the reference to the CompVis team now being “part of team Stability” appears to be an inclusive statement, designed to reflect the collaborative nature of the relationship between the two organisations; a relationship which is also reflected in a post made by Mr Mostaque nearly a year later, on 4 June 2023, in which he records that Stability was “a collaborator in the development of the first release of Stable Diffusion”. None of this evidence establishes on balance that Stability in fact took control over, and responsibility for, the launch of v1.x such that it will have direct liability for any tortious claims that may be established in respect of the release of v1.x to the GitHub and Hugging Face platforms, as Getty Images contend.

Mr Passos’ intervention in the 8 August 2022 Chat expressly acknowledges the collaboration between Stability (as a commercial entity) and Compvis (as the academic research vehicle) and he certainly suggests the desirability of a separation between the two when it comes to branding. However, while it is unclear what knowledge he had as to the arrangements for the release of v1.x, his understanding appears to be that the Model/its academic codebase was a CompVis product.

On 10 August 2022, Stability posted an announcement in the following terms on its website (emphasis added):

“Stability AI and our collaborators are proud to announce the first stage of the release of Stable Diffusion to researchers. Our friends at Hugging Face will host the model weights once you get access. The code is available here [https://github.com/CompVis/stable-diffusion], and the model card is here [https://huggingface.co/CompVis/stablediffusion]. We are working together towards a public release soon. This has been led by Patrick Esser from Runway and Robin Rombach from the Machine Vision & Learning research group at LMU Munich (formerly CompVis lab at Heidelberg University) building on their prior work on Latent Diffusion Models at CVPR’22, combined with support from communities at Eleuther AI, LAION, and our own generative AI team.

The model builds upon the work of the team at CompVis and Runway in their widely used latent diffusion model combined with insights from the conditional diffusion models by our lead generative AI developer Katherine Crowson, Dall-E 2 by Open AI, Imagen by Google Brain, and many others. We are delighted that AI media generation is a cooperative field and hope it can continue this way to bring the gift of creativity to all.”

The announcement went on to set out quotes from various interested individuals, including the following:

“We are excited to see what will be built with the current models as well as to see what further works will be coming out of open, collaborative research efforts!” – Patrick [Esser] (Runway) and Robin [Rombach] (LMU)

“We're excited that state of the art text-to-image models are being built openly and we are happy to collaborate with CompVis and Stability.ai towards safely and ethically release the models to the public and help democratize ML capabilities with the whole community” – Apolinário, ML Art Engineer, Hugging Face”

“We are delighted to release the first in a series of benchmark open source Stable Diffusion models that will enable billions to be more creative, happy and communicative. This model builds on the work of many excellent researchers and we look forward to the positive effect of this and similar models on society and science in the coming years as they are used by billions worldwide”. - Emad, CEO, Stability AI

In a second announcement entitled “Stable Diffusion Public Release” which was published on the Stability website on 22 August 2022, Stability announced the public release of Stable Diffusion, providing a link to its own website at https://stability.ai/stablediffusion and again providing a link to the CompVis Hugging Face and CompVis GitHub pages. Later in the announcement, Stability directed potential users to DreamStudio “[f]or more control and rapid generation”.

In my judgment, far from evidencing that Stability itself published the Model on the CompVis Hugging Face and CompVis GitHub pages, the 10 August 2022 announcement emphasises the collaboration involved in its development together with the leading role of CompVis and Runway. Importantly, the links to the code and to the Model Card which were provided in both the 10 and 22 August 2022 announcements are to CompVis pages on GitHub and Hugging Face; the Stability GitHub page that Mr Mostaque had apparently envisaged during the 26 July 2022 Chat was not in place at this time (although by 22 August 2022 Stability was providing access to v1.4 of the Model via DreamStudio (as it accepts)).

The CompVis GitHub page (URL https://github.com/ComPVis/stable-diffusion) is headed “CompVis/stable-diffusion” and a link is immediately provided to the Ommer Lab, the lab of Professor Ommer at https://ommer-lab.com/research/latent-diffusion-models/. Under the heading “Stable Diffusion” is the following post from CompVis:

“Stable Diffusion was made possible thanks to a collaboration with Stability AI and Runway and builds upon our previous work”.

Later on the page (which includes detailed advice for downloading and sampling the Model together with details of the checkpoints that were being provided), CompVis explains that:

“Thanks to a generous compute donation from Stability AI and support from LAION we were able to train a Latent Diffusion Model on 512x512 images from a subset of the LAION-5B database”

“The weights are available via the CompVis organization at Hugging Face under a license which contains specific use-based restrictions to prevent misuse and harm as informed by the model card, but otherwise remains permissive. While commercial use is permitted under the terms of the license, we do not recommend using the provided weights for services or products without additional safety mechanisms and considerations, since there are known limitations and biases of the weights, and research on safe and ethical deployment of general text-to-image models is an ongoing effort. The weights are research artifacts and should be treated as such.” (emphasis added).

The license referred to was a CreativeML Open RAIL-M license dated 22 August 2022 in which “Licensor” was defined as “the copyright owner or entity authorized by the copyright owner that is granting the license, including the persons or entities that may have rights to the Model and/or distributing the Model”. The copyright owners were identified as “Robin Rombach and Patrick Esser and contributors” (in contrast to the position in respect of v2.0 to which I have referred above). The CompVis Hugging Face page is a sub-page of the CompVis directory at https://huggingface.co/CompVis and is headed “CompVis/stable-diffusion-v1-1”.

Consistent with the evidence provided by the CompVis GitHub and Hugging Face pages is an article in The Verge by James Vincent dated 15 September 2022 attached by Getty Images to their CEA notice which refers to Mr Mostaque observing that the Model “is being released by a research institute as a generalized model”. The author of the article comments that by referring to a “research institute” Mr Mostaque “is referring to the fact that the technical license for Stable Diffusion has been released by the Ludwig Maximillian University of Munich’s CompVis lab, though Stability funded and shaped its development”.

In light of the available evidence, I reject Getty Images’ submission that the use of hyperlinks on Stability’s website directing end users to CompVis pages on GitHub and Hugging Face is evidence of Stability taking “active steps” to make available Stable Diffusion to the public in the UK via these pages. In my judgment the provision of these links was no more than the provision of information as to where a prospective user could find the weights, Model Card and code. It did not involve Stability actually making these resources available itself on these platforms. The fact that the code and Model Card were released on a CompVis page is entirely consistent with a collaboration between Stability and CompVis which involved Stability (as the commercial arm) funding the development of, being involved in, and publicising the launch of, Stable Diffusion v1.x, and CompVis (as the research and development arm) wishing to release the fruits of its labours in the form of the Model Card and source code on its own Hugging Face and GitHub pages from the outset. At this time, as the witness statement of Zachary Evans confirms, Mr Rombach was not an employee of Stability.

Further evidence which appears to me to be supportive of this conclusion is contained in an announcement published on the Stability website on 24 November 2022 (after Mr Rombach had become an employee of Stability) in which Stability announced the release of v2.0 and explained that:

“The dynamic team of Robin Rombach (Stability AI) and Patrick Esser (Runway ML) from the CompVis Group at LMU Munich, headed by Prof. Dr. Bjӧrn Ommer, led the original Stable Diffusion V1 release. They built on their prior work in the lab with Latent Diffusion Models and got critical support from LAION and Eleuther AI…Robin is now leading the effort with Katherine Crowson at Stability AI to create the next generation of media models with our broader team”.

On this occasion, (in contrast to the release of v1.x) Stability refers anyone interested in the release of v2.0 of the Model to “our GitHub: https://github.com/Stability-AI/Stable Diffusion” (emphasis added) and states that “[w]e are releasing these models into the Stability AI API Platform (platform.stability.ai) and DreamStudio in the next few days”. The Stability GitHub page refers to the fact that:

“The weights are available via the StabilityAI organization at Hugging Face under the CreativeML Open RAIL++-M License” (emphasis added).

On balance, having reviewed the available documentary evidence, I consider there to be no basis for a finding that Stability actively made v1.x available and/or published the source code and model weights for v1.x to the public via the CompVis Hugging Face and CompVis GitHub platforms, and/or that it was the “moving force” behind this publication, and I reject Getty Images’ case to that effect. I do not consider the fact that Stability accepts that it collaborated in the development of v1.x, that it was involved in its launch, or that it published the Model elsewhere, to be inconsistent with this conclusion and, as I have already said, there is no case to the effect that Stability is a joint tortfeasor together with CompVis (or anyone else) for any allegedly infringing acts.

In closing, Getty Images sought to rely upon evidence from Mr Wagrez to the effect that when he was putting together his first witness statement he had understood that he was being asked about what data had been used to train v.1, v.2 and XL of the Model. It was put to him in cross examination that he therefore understood that Stability was “responsible for those models”, to which he responded “Yes”. However, I do not consider this evidence to be of any real assistance. Aside from the fact that Mr Wagrez did not join Stability until October 2023 and so can have no direct knowledge of the details of the release of Stable Diffusion v1.x, the question was posed in the abstract, without any attempt to explain what was meant by “responsible for those models”. Given that Stability accepts that it did make v1.x available via access mechanisms other than the CompVis Hugging Face and Compvis GitHub pages, Mr Wagrez’s evidence takes matters no further. Similarly, although it was suggested by Getty Images that Mr Auerhahn “did not distinguish” in his evidence between v1.x and later versions of Stable Diffusion and therefore the submission was made that “it is clear” that he thought that responsibility for Stable Diffusion resided with Stability, regardless of the version or sub-version, Mr Auerhahn’s questioning was not directed at the question of whether Stability was “responsible” for releasing v1.x via any particular access mechanism.

I note in passing that the parties agreed upon an issue for disclosure specifically addressed to the question of Stability’s responsibility for the publication of Stable Diffusion on CompVis GitHub and CompVis Hugging Face and identified various potentially relevant categories of documents, but, aside from the documents to which I have referred, no other evidence is relied upon by Getty Images. I reject Ms Lane’s oral submission that the use of the CompVis page was mere “happenstance”.

I also reject Getty Images’ case (advanced in opening submissions) that Stability is responsible for publishing and/or making Stable Diffusion v1.x available via Mr Esser’s Github page prior to the official launch of v1.x on the CompVis GitHub page. There is no support for such an allegation in the Particulars of Claim and it appears to be based upon the content of the 28 June 2022 email to Amazon referred to above. Getty Images provided no explanation as to why publication on Mr Esser’s page could amount to publication to the public by Stability and I did not understand the point to be pursued in closing.

Finally I note that the question of responsibility for v1.x was an issue on which Getty Images invited me to draw an adverse inference by reason of the absence of Mr Vencu as a witness. Ms Lane drew my attention to Efobi v Royal Mail Group Ltd [2021] 1 WLR 3863 per Lord Leggatt JSC at [41], and submitted that she would have wanted to question Mr Vencu on the issue of the release of Stable Diffusion v1.x, not least because his name appears on various relevant documents. Ms Lane is right on that score – Mr Vencu was a party to various of the Chats to which I have referred above and occasionally contributed to them. Furthermore, as Getty Images point out, Mr Vencu was employed by Stability from February 2022 and his role involved providing infrastructure, compute and storage services to research teams. No doubt he would have had relevant evidence to give on the development and training of Stable Diffusion and it is certainly possible that he would have had material evidence to give on the issue of “making available to the public”.

Always subject to my being satisfied with the explanation for Mr Vencu’s absence, then I accept that if Getty Images had established a prima facie case to the effect that Stability released v1.x to the public via the CompVis Hugging Face and GitHub pages, then it may have been appropriate to infer that nothing Mr Vencu would have said in evidence (had he been called) would have undermined that case. However, in my judgment I need not consider the matter further. As set out above, the available evidence goes the other way and, in the circumstances, I do not consider it appropriate to draw any inference by reason of Mr Vencu’s absence and certainly not an inference that would be inconsistent with the evidence contained in the contemporaneous documents.

In conclusion, for the reasons set out above, I find that Stability bears no direct liability for any tortious acts alleged in these proceedings arising by reason of the release of Stable Diffusion v1.x via the CompVis GitHub and CompVis Hugging Face pages.

I note, however, that there is of course no dispute that Stability did make v1.x available via an API Platform site and (in the case of v1.4 only) via its own DreamStudio platform. All references to v1.x hereafter are intended to be references only to v1.x as made available via these two access mechanisms.

It is common ground that Stability is responsible for the release of the other versions of Stable Diffusion in dispute in these proceedings.