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

THE TRADE MARK INFRINGEMENT CLAIM

Getty Images allege that Stability has committed acts of trade mark infringement pursuant to sections 10(1), 10(2) and 10(3) of the TMA by reason of the fact that it has affixed watermarks (referred to hereafter as “the Sign(s)” or “watermarks*” (Footnote: ⁶)), which are identical or similar to the Marks, to the synthetic image outputs of Stable Diffusion and used those Signs in relation to services, namely the provision of synthetic image outputs. Getty Images accept that not every synthetic image produced by Stable Diffusion produces a Sign which is similar or identical to the Marks, but they say that they have adduced evidence that some do.

The question of whether the use of a sign infringes a trade mark under these sections falls to be assessed as at the date that the use of the sign complained of was commenced (Levi Strauss & Co v Casucci SpA (C-145/05) [2006] E.C.R. I-3703 at [13]). It is impossible to identify when each Model will have generated an allegedly infringing watermark*, and so I accept Getty Images’ case that the relevant dates for present purposes must assume infringement from the date of release of each Model. In other words, August 2022 for v1.x; November 2022 for v2.0; Mid 2023 for SD XL and November 2023 for v1.6. Although Stability complains that this approach is undisciplined in the context of identifying the use on which Getty Images relies for the purposes of infringement (because it relies upon use of the Models per se, rather than use of an infringing sign), nevertheless, I consider it to be the only sensible approach in the very unusual circumstances of this case. None other was suggested by Stability.

It is common ground that Getty Images must establish, on the balance of probabilities, that each version of the Model has generated at least one output with a watermark* containing a Getty Mark and a watermark* containing an ISTOCK Mark in the United Kingdom. It is important to differentiate between outputs produced by different versions of the Model because (as is also common ground) the Models have not all been trained on the same dataset and different filters have been applied to the training data. Thus while one version of the Model may generate synthetic images which include the Sign, another version may not.

There is no pleaded defence of “de minimis” from Stability and Mr Cuddigan confirmed in submissions that Stability did not intend to run such a case. Thus there is no scope for Stability to contend (as it appeared at times to do in its opening submissions) that detailed issues of infringement will only arise “if the court decides that the evidence of watermark* incidence ‘in the wild’, in the UK, is a problem of substance”. However, in closing, Stability submitted that, in light of the evidence, a threshold issue arises, namely whether Getty Images have proved that any user of any version of the Model in the UK has ever been presented with a watermark* on a synthetic image generated by the Model. It is accordingly to this threshold question that I must turn first.

Stability admits in its Defence that “it may be possible” to generate synthetic image outputs which feature watermarks* and it accepted in closing that the Getty Watermark Experiments have shown that it is possible “to push” the Models in issue to generate watermarks*. However Stability says that these experiments were contrived and that when it comes to “non-legacy models” (which I understand to be a reference to SD XL and v1.6) the experiments failed. Stability relies upon a non-admission in its Defence as to the presence of watermarks* on output images generated by users of Stable Diffusion as the foundation for its argument (run at considerable length in closing) that Getty Images have been unable to prove at trial that any UK user of the various Models in issue has ever been presented with a watermark* on a synthetic image in the UK.

While Stability’s primary case is that this broad submission applies to all versions of the Models, Mr Cuddigan accepted in closing that “there is a different position in relation to different models” and that “there is a difference between v1.x and the others”. This difference arises because, as Stability acknowledged in its written opening submissions, “the incidence of watermarks* in outputs [from v.1x models] is non-trivial” and has “occurred to a non-trivial extent both in response to prompts designed by Getty to generate such watermarks*, and sporadically in relation to other prompts”. However, Mr Cuddigan explained in closing that this was intended as a reference to the incidence “worldwide” and that it involved no more than an acknowledgement that there is available evidence of commentary about the appearance of watermarks* in respect of v1.x, together with evidence that this has had an impact on users.

As for the remaining versions of the Model, Stability submits, in essence, that the available evidence is insufficient to satisfy Getty Images’ burden of proof, that much of the evidence of watermark* incidence has been produced using prompts that real world users would not use and that Getty Images has failed to advance any case based on the statistical probability of an infringing Sign being produced on a synthetic image generated from one or more versions of the Model.

Getty Images strongly resist this argument, pointing out that Stability’s Defence on the point is, at best, equivocal, and that Stability has failed to run any positive case asserting that watermarks* have not been generated by users of the various versions of Stable Diffusion in the UK. Getty Images submit that Stability has effectively already conceded the point on v1.x (notwithstanding Mr Cuddigan’s explanation in closing) and that, by reason of that concession, they had understood Stability to be advancing a case on wilful contrivance only in respect of v2.x, SD XL and v1.6 (and not v1.x).

Furthermore, Getty Images maintain that, in light of Stability’s non-admission in its Defence, Getty Images does not require much by way of evidence to discharge their burden of proof and they say that in fact they have ample evidence with which to do so. Owing to the absence of any plea by Stability that the generation of watermarks* by any of the Models in issue is de minimis, Getty Images submit that “it is certainly not necessary for Getty Images to prove that there is a particular probability of any of the Models generating watermarked images”.

To look at the threshold question in more detail, I must first examine the available evidence as to the generation of watermarks* generally by the Models. That evidence falls into four main categories: (i) the Expert evidence as to the scope for generation of watermarks*; (ii) Annex 8I to the PoC; (iii) the Getty Watermark Experiments and Annex 8H to the PoC; and (iv) Evidence of watermark* generation “in the wild”.