Andersen v. Stability AI Ltd.

Full title: SARAH ANDERSEN, et al., Plaintiffs, v. STABILITY AI LTD., et al.…

Court: United States District Court, Northern District of California

Date published: Aug 12, 2024

23-cv-00201-WHO (N.D. Cal. Aug. 12, 2024)

Opinion

23-cv-00201-WHO

08-12-2024

SARAH ANDERSEN, et al., Plaintiffs, v. STABILITY AI LTD., et al., Defendants.

William H. Orrick, United States District Judge

ORDER GRANTING IN PART AND DENYING IN PART MOTIONS TO DISMISS FIRST AMENDED COMPLAINT DKT. NOS. 160, 162, 163, 164, 169

William H. Orrick, United States District Judge

Defendants Stability AI Ltd. and Stability AI, Inc. (collectively “Stability AI”), Midjourney, Inc., DeviantArt, Inc., and Runway AI, Inc. move to dismiss various claims from plaintiffs' First Amended Complaint (“FAC,” Dkt. No. 129). For the reasons discussed below, those motions are GRANTED in part and DENIED in part.

BACKGROUND

Artists Sarah Andersen, Kelly McKernan, Karla Ortiz, Hawke Southworth, Grzegorz Rutkowski, Gregory Manchess, Gerald Brom, Jingna Zhang, Julia Kaye, and Adam Ellis (“Plaintiffs”¹) filed this putative class action on behalf of artists challenging the defendants' creation and/or use of Stable Diffusion, an artificial intelligence (“AI”) software product. They allege that Stable Diffusion used plaintiffs' artistic works as “training images” and as a result Stable Diffusion can produce output images “in the style” of those images. See generally FAC. In the FAC, plaintiffs allege claims against the three defendants identified in the original complaint (Stability AI, Midjourney, and DeviantArt) and against a new defendant, Runway AI.² Plaintiffs contend that Runway AI worked with, helped train, and then distributed Stable Diffusion with Stability AI. Plaintiffs assert that Runway made a text-to-image generator available via its online AI image product called AI Magic Tools.

1 In their FAC, plaintiffs reasserted claims for the three original plaintiffs (Sarah Anderson, Kelly McKernan, and Karla Ortiz) and added - without the court permission - seven additional plaintiffs (H. Southworth PKA Hawke Southworth, Grzegorz Rutkowski, Gregory Manchess, Gerald Brom, Jingna Zhang, Julia Kaye, and Adam Ellis). FAC ¶¶ 14-23.

2 The basic factual allegations regarding how defendants' AI products were trained and work were identified in my prior Order and will not be repeated here. See October 30, 2023, Order at Dkt. No. 117. To the extent new, material allegations have been added to the FAC, they will be addressed as part of the substantive analysis below.

Plaintiffs' claims center first around the creation of the LAION training sets, where five billion images were allegedly scraped into datasets used by Stability and Runway to train the versions of Stable Diffusion. FAC ¶ 4. Plaintiffs state that Midjourney likewise trained its product using Stable Diffusion, and that all four defendants use Stable Diffusion in their AI products; in doing so, those four defendants copy or utilize versions of plaintiffs' artistic works. Id. ¶¶ 5-6. Significantly, plaintiffs allege that the “LAION-5B dataset contains only URLs of training images, not the actual training images. Therefore, anyone who wishes to use LAION-5B for training their own machine learning model must first acquire copies of the actual training images from their URLs using the img2dataset or other similar tool.” FAC ¶ 221. They also clarify their theory of direct infringement, adding allegations regarding CLIP-guided diffusion in the training phase but also in use, after training. Id. ¶¶ 82-150.

Plaintiffs assert claims on behalf of six different classes:

“Injunctive Relief Class” under Rule 23(b)(2): All persons or entities nationalized or domiciled in the United States that own a copyright interest in any work that was used to train any version of an AI image product that was offered directly or incorporated into another product by one or more Defendants during the Class Period. FAC ¶ 34.

“Damages Class” under Rule 23(b)(3): All persons or entities nationalized or domiciled in the United States that own a copyright interest in any work that was used to train any version of an AI image product that was offered directly or incorporated into another product by one or more Defendants during the Class Period. Id.

“LAION-5B Damages Subclass” under Rule 23(b)(3): All persons or entities nationalized or domiciled in the United States that own a registered copyright in any work in the LAION-5B dataset that was used to train any version of an AI image product that was offered directly or incorporated into another product by one or more Defendants during the Class Period.³ Id.

3 The “LAION-5B Registered Plaintiffs” as used in the FAC include “the subset of plaintiffs who hold copyrights in these LAION-5B Registered Works that were registered before the filing of the initial complaint in this action, namely Sarah Andersen, Jingna Zhang, Gerald Brom, Gregory Manchess, Julia Kaye, and Adam Ellis.” FAC ¶ 213.

“LAION-400M Damages Subclass” under Rule 23(b)(3): All persons or entities nationalized or domiciled in the United States that own a registered copyright in any work in the LAION-400M dataset that was used to train any version of an AI image product that was offered directly or incorporated into another product by one or more Defendants during the Class Period. Id.

“DeviantArt Damages Subclass” under Rule 23(b)(3): All members of the Damages Class who (1) maintained an account on DeviantArt; (2) posted copyrighted work on DeviantArt; and (3) had that work used to train any version of an AI image product. Id.

“Midjourney Named Artist Class” under Rule 23(b)(3): All persons or entities who appear on the Midjourney Names List and whose names were invoked within prompts of the Midjourney Image Product during the Class Period.⁴ Id.

4 The “Midjourney Named Plaintiffs” are the plaintiffs whose names were disclosed by Midjourney as artists whose works were included and could be recreated through use of Midjourney's product; Grzegorz Rutkowski, Sarah Andersen, Karla Ortiz, Gerald Brom, and Julia Kaye. FAC ¶¶ 263, 264.

Plaintiffs assert the following claims against the different sets of defendants:

Against Stability AI: (1) direct copyright infringement of the LAION-5B Registered Works by training the Stability Models, including Stable Diffusion 2.0 and Stable Diffusion XL 1.0 on behalf of the LAION-5B Registered Plaintiffs and Damages Subclass; (2) inducement of copyright infringement by distributing Stable Diffusion 2.0 and Stable Diffusion XL 1.0 for free on behalf of the LAION-5B Registered Plaintiffs and Damages Subclass; (3) violations of the Digital Millennium Copyright Act (“DMCA”) by removing and altering copyright management information (“CMI”) of training images on behalf of all Plaintiffs, the Damages and the Injunctive Classes; and (4) unjust enrichment under Cal. Bus. & Prof. Code § 17200 and California Common Law on behalf of all Plaintiffs, the Damages and the Injunctive Classes.

Against Runway AI: (1) Direct copyright infringement of the LAION-5B Registered Works by training the Runway Models, including Stable Diffusion 1.5 on behalf of the LAION-5B Registered Plaintiffs, LAION-5B Subclass, and Karla Ortiz Individually; (2) Inducement of copyright infringement by distributing Stable Diffusion 1.5 for free on behalf of the LAION-5B Registered Plaintiffs and Subclass; (3) DMCA violations by removing and altering CMI of training images on behalf of all Plaintiffs, the Damages and Injunctive Classes; and (4) Unjust enrichment under Cal. Bus. & Prof. Code § 17200 and California Common Law on behalf of all Plaintiffs, the Damages and Injunctive Classes.

Against Midjourney: (1) Direct copyright infringement of the LAION-400M Registered Works by training the Midjourney 400M Models, including Midjourney Model version 1 on behalf of the LAION-400M Registered Plaintiffs and Damages Subclass; (2) Direct copyright infringement of the LAION-5B Registered Works by training the Midjourney 5B Models, including Midjourney Model version 5.2 on behalf of the LAION-5B Registered Plaintiffs and Damages Subclass; (3) DMCA violations by removing and altering CMI of training images on behalf of All Plaintiffs, the Damages and Injunctive Class; (4) Lanham Act - false endorsement by unauthorized commercial use of artists' names on behalf of the Midjourney Named Plaintiffs and Class; (5) Lanham Act - vicarious trade-dress violation by profiting from imitations of protectable trade dress on behalf of the Midjourney Named Plaintiffs and Class; and (6) Unjust enrichment under Cal. Bus. & Prof. Code § 17200 and California Common Law on behalf of all Plaintiffs, and the Damages and Injunctive Class.

Against DeviantArt: (1) Direct copyright infringement by copying the DreamUp-CompVis Model and incorporating it into DreamUp on behalf of the LAION-5B Registered Plaintiffs; (2) Breach of contract for violation of its Terms of Service on behalf of the DeviantArt Plaintiffs; (3) Unjust enrichment under Cal. Bus. & Prof. Code § 17200 and California Common Law on behalf of the DeviantArt Plaintiffs.

In October 2023, I largely granted the motions to dismiss brought by defendants Stability, Midjourney and DeviantArt. The only claim that survived was the direct infringement claim asserted against Stability, based on Stability's alleged “creation and use of ‘Training Images' scraped from the internet into the LAION datasets and then used to train Stable Diffusion.” October 2023 Order at 7. The remainder of the claims were dismissed with leave, so that plaintiffs could amend “to provide clarity regarding their theories of how each defendant separately violated their copyrights, removed or altered their copyright management information, or violated their rights of publicity and plausible facts in support.” Id. at 1.

As one example, I required plaintiffs on amendment to address the following deficiencies with a second theory of direct infringement, separate from the creating and use of images for training theory:

Plaintiffs will be required to amend to clarify their theory with respect to compressed copies of Training Images and to state facts in support of how Stable Diffusion - a program that is open source, at least in part - operates with respect to the Training Images. If plaintiffs contend Stable Diffusion contains “compressed copies” of the Training Images, they need to define “compressed copies” and explain plausible facts in support. And if plaintiffs' compressed copies theory is based on a contention that Stable Diffusion contains mathematical or statistical methods that can be carried out through algorithms or instructions in order to reconstruct the Training Images in whole or in part to create the new Output Images, they need to clarify that and provide plausible facts in support.

October 2023 Order at 8-9; see also id. 9 n.7 (“Plaintiffs' second theory of direct infringement - that Stable Diffusion is a ‘derivative work' because it contains compressed copies of billions of copyrighted images and by incorporating Stable Diffusion into” defendants' own AI products, defendants are “liable for producing works that have been ‘transformed' based on plaintiffs' works [] fails for the same reasons.”).

Plaintiffs added defendant Runway AI and seven new plaintiffs when it filed the FAC. Each defendant moves to dismiss.

LEGAL STANDARD

Under FRCP 12(b)(6), a district court must dismiss a complaint if it fails to state a claim upon which relief can be granted. To survive a Rule 12(b)(6) motion to dismiss, the plaintiff must allege “enough facts to state a claim to relief that is plausible on its face.” Bell Atl. Corp. v. Twombly, 550 U.S. 544, 570 (2007). A claim is facially plausible when the plaintiff pleads facts that “allow the court to draw the reasonable inference that the defendant is liable for the misconduct alleged.” Ashcroft v. Iqbal, 556 U.S. 662, 678 (2009) (citation omitted). There must be “more than a sheer possibility that a defendant has acted unlawfully.” Id. While courts do not require “heightened fact pleading of specifics,” a plaintiff must allege facts sufficient to “raise a right to relief above the speculative level.” Twombly, 550 U.S. at 555, 570.

In deciding whether the plaintiff has stated a claim upon which relief can be granted, the Court accepts the plaintiff's allegations as true and draws all reasonable inferences in favor of the plaintiff. See Usher v. City of Los Angeles, 828 F.2d 556, 561 (9th Cir. 1987). However, the court is not required to accept as true “allegations that are merely conclusory, unwarranted deductions of fact, or unreasonable inferences.” In re Gilead Scis. Sec. Litig., 536 F.3d 1049, 1055 (9th Cir. 2008). If the court dismisses the complaint, it “should grant leave to amend even if no request to amend the pleading was made, unless it determines that the pleading could not possibly be cured by the allegation of other facts.” Lopez v. Smith, 203 F.3d 1122, 1127 (9th Cir. 2000). In making this determination, the court should consider factors such as “the presence or absence of undue delay, bad faith, dilatory motive, repeated failure to cure deficiencies by previous amendments, undue prejudice to the opposing party and futility of the proposed amendment.” Moore v. Kayport Package Express, 885 F.2d 531, 538 (9th Cir. 1989).

DISCUSSION

I. ADDITION OF NEW PLAINTIFFS & CLAIMS

Stability objects to the addition of seven new plaintiffs to the FAC, as well as the addition of two new claims against it (induced infringement, Count 2) and unjust enrichment (Count 4). Stability MTD, Dkt. No. 162 at 13-14. DeviantArt similarly objects to the addition of new plaintiffs and the addition of the unjust enrichment claim. DeviantArt MTD, Dkt. No. 163, at 2324. Defendants argue that the addition of the new plaintiffs and claims exceeds the scope of the leave to amend I allowed plaintiffs in the October 2023 Order and, therefore, the new plaintiffs and claims should be stricken.

Plaintiffs respond that they were given broad “leave to amend and attempt to cure the deficiencies identified” in their claims and the ability of the then three named plaintiffs to pursue the claims asserted. Because leave was not cabined to “solely” or exclusively correct the deficiencies, plaintiffs argue that the additions were permissible and in any event should be allowed under Rule 15's liberal amendment standard. See Oppo. to Stability, Dkt. No. 174, at 22; Oppo. to DeviantArt, Dkt. No. 177, at 21-22.

Plaintiffs are correct that leave to amend under Rule 15 is “freely given,” especially at the start of a case. However, once a complaint is dismissed and a court grants plaintiffs specific leave to amend to address identified legal or factual deficiencies, adding plaintiffs or claims beyond those previously alleged requires either requesting and securing leave, which plaintiffs did not do, or the consent of defendants, which plaintiffs did not seek. See Fed. R. Civ. P, 15(a)(2) (“In all other cases, a party may amend its pleading only with the opposing party's written consent or the court's leave.”).⁵

5 The only case relied on by plaintiffs is totally inapposite. It addresses a situation where plaintiffs were given leave to amend to file a second amended complaint and revised motion for class certification by stipulation, and after the court consolidated cases, defendants moved to dismiss the new plaintiffs based on statute of limitations grounds. See Bos. Ret. Sys. v. Uber Techs., Inc., No. 19-CV-06361-RS, 2021 WL 4503137, at *2 (N.D. Cal. Oct. 1, 2021).

That said, I would have granted leave to amend if plaintiffs had sought leave, given the lack of prejudice to defendants. At this juncture, where each defendant has addressed the newly added claims and plaintiffs, it would elevate form over substance to grant the motions to dismiss on this ground. And I will assume that plaintiffs have implicitly sought leave to amend to include the new plaintiffs and claims in a Second Amended Complaint. I will grant leave and address the arguments defendants make against the added claims and plaintiffs.⁶

6 As explained below, the end result is that plaintiffs may file a Second Amended Complaint including the new plaintiffs and may attempt to plead unjust enrichment claims against any defendant based on theories (if any) that are not preempted by the Copyright Act.

II. STABILITY AI MOTION TO DISMISS

A. Induced Copyright Infringement

Stability challenges plaintiffs' first theory of induced copyright infringement interpreting the claim as one alleging that the Stable Diffusion models themselves are infringing works. Under this theory, Stability is inducing infringement by distributing the models when any third-party downloads, uses, or deploys the models provided by Stability. See FAC, Count Two (alleging contributory infringement by “distributing Stable Diffusion 2.0 and Stable Diffusion XL 1.0 for free” on “behalf of the LAION-5B Registered Plaintiffs and Damages Subclass”) ¶¶ 233-236.⁷

7 Stability does not challenge plaintiffs' claim of direct copyright infringement, as that survived the prior motion to dismiss. October 2023 Order at 7.

Stability argues, first, that this theory is simply a repackaged direct infringement theory, that by “distributing” Stable Diffusion, Stability violates plaintiffs' exclusive rights of distribution of their works.⁸ But whether this is a direct infringement claim (where liability is imposed against Stability for distributing copyrighted works) or more properly characterized as an inducement claim (where liability is imposed because Stability induces or otherwise causes others to copy protectible material) depends on how Stable Diffusion works and is implemented by users other than Stability itself. Any potential overlap - or potential requirement for plaintiffs to elect one claim or another - is better addressed on summary judgment, after discovery.

8 Stability also contends this theory is another take on the “derivative” works claim I dismissed in the October 2023 Order. However, that claim was dismissed in large part because of the ambiguity and lack of plausible facts that Stable Diffusion was itself a “derivative work” because it contained compressed copies of billions of copyrighted images. See October 2023 Order at 10 n.7. Plaintiffs were given leave to amend their compressed copies theory in support of direct copyright infringement with respect to the fully trained Stable Diffusion product, and as discussed with respect to the motions to dismiss below, have plausibly done so.

Stability also argues that the inducement claim must be dismissed to the extent plaintiffs are alleging that Stability encourages the use of Stable Diffusion to create infringing outputs. That theory is barred, according to Stability, because plaintiffs fail to allege any facts that Stability promoted the use of Stable Diffusion to “infringe copyright, as shown by clear expression or other affirmative steps taken to foster infringement.” Metro-Goldwyn-Mayer Studios Inc. v. Grokster, Ltd., 545 U.S. 913, 919 (2005). Stability argues that given the lack of allegations supporting a specific intent by Stability to promote infringement, the claim fails.

However, plaintiffs point to one statement by Stability's CEO that Stability took 100,000 gigabytes of images and compressed it to a two-gigabyte file that can “recreate” any of those images.⁹ Stability responds that the “isolated” use of the word “create” by its CEO cannot demonstrate that it intended to foster infringement. Stability argues that “clear allegations of active steps to encourage direct infringement” are especially important here, where plaintiffs do not dispute that Stable Diffusion is capable of substantial noninfringing uses, like creating art from inputs that do not reference particular artists or invoke particular artists' styles or have any substantial similarity to plaintiffs' works.

9 See FAC ¶ 4 (“Emad Mostaque described it thus: ‘Stable Diffusion is the model itself. It's a collaboration that we did with a whole bunch of people ... We took 100,000 gigabytes of images and compressed it to a two-gigabyte file that can recreate any of those [images] and iterations of those.'”).

The theory of this case is not similar to - for example - a case asserting contributory infringement based on the sale of VCRs where, after discovery, plaintiff had no evidence of defendant's intent to induce infringement. The Supreme Court explained that, in those circumstances, intent could not be “based on presuming or imputing intent to cause infringement solely from the design or distribution of a product capable of substantial lawful use, which the distributor knows is in fact used for infringement.” See Grokster, Ltd., 545 U.S. at 933 (discussing holding of Sony Corp. of America v. Universal City Studios, Inc., 464 U.S. 417 (1984)). Instead, this is a case where plaintiffs allege that Stable Diffusion is built to a significant extent on copyrighted works and that the way the product operates necessarily invokes copies or protected elements of those works. The plausible inferences at this juncture are that Stable Diffusion by operation by end users creates copyright infringement and was created to facilitate that infringement by design. In addition to the comment of Stability's CEO, plaintiffs reference articles by academics and others that training images can sometimes be reproduced as outputs from the AI products. FAC ¶¶ 90, 130-139.

Whether true and whether the result of a glitch (as Stability contends) or by design (plaintiffs' contention) will be tested at a later date. The allegations of induced infringement are sufficient.

B. DMCA

Stability also moves again to dismiss plaintiffs' DMCA claims asserted under 17 U.S.C. section 1202(a) for providing or distributing false copyright management information (“CMI”) and under section 1202(b)(1) for intentional removal of CMI.¹⁰ I dismissed plaintiffs' DMCA claim with leave to amend in the October 2023 Order, explaining:

10 Section 1202(a) prohibits the knowing, with the intent to induce or enable infringement, provision or distribution of false CMI. Section 1202(b)(1) governs unpermitted removal or alteration of CMI and distribution of works with removed or altered CMI. See 17 U.S.C. § 1202 et seq.

[E]ach plaintiff must identify the exact type of CMI included in their online works that were online and that they have a good faith belief were scraped into the LAION datasets or other datasets used to train Stable Diffusion. At the hearing, plaintiffs argued that it is key for the development of generative AI models to capture not only images but any accompanying text because that accompanying text is necessary to the models' ability to “train” on key words associated with those images. Tr. at 9:13-24. But there is nothing in the Complaint about text CMI present in the images the named plaintiffs included with their online images that they contend was stripped or altered in violation of the DMCA during the training of Stable Diffusion or the use of the end-products. Plaintiffs must, on amendment, identify the particular types of their CMI from their works that they believe were removed or altered.

In addition, plaintiffs must clarify and then allege plausible facts regarding which defendants they contend did the stripping or altering in violation of the DMCA and when that occurred.

October 2023 Order at 18.

1. Claim Under 1202(a)

Stability moves to dismiss the subsection (a) claim regarding false CMI. Plaintiffs allege that “Stability distributes the Stability Models under the MIT License (see, e.g. -https://github.com/Stability-AI/stablediffusion/blob/main/LICENSE). Within this license, Stability asserts copyright in the Stability Models. By asserting copyright in the Stability Models, which infringe the copyrights of the LAION-5B Plaintiffs, Stability is providing and distributing false CMI in violation of 17 U.S.C. § 1202(a).” FAC ¶ 248.

Stability argues that this claim fails because Stability's generic license does not suggest any association at all with plaintiffs' works, and therefore was not made “in connection with” plaintiffs' works which is necessary to support a claim under 1202(a). See Logan v. Meta Platforms, Inc., 636 F.Supp.3d 1052, 1062 (N.D. Cal. 2022) (plaintiff failed to allege facts showing defendant “conveyed CMI in connection” with plaintiff's photos where “Meta's allegedly false CMI is ‘[a] [generic] copyright tag on the bottom of each Facebook user page,' separated from the rest of the content on the webpage, and is not located on or next to” plaintiff's photos); but see Post Univ. v. Course Hero, Inc., No. 3:21-CV-1242 (JBA), 2023 WL 5507845, at *4 (D. Conn. Aug. 25, 2023) (“Thus, at the motion to dismiss stage, unless it is ‘implausible' that a viewer could understand the information to be referring to the defendant as the work's copyright holder, dismissal is inappropriate.”). In addition, Stability contends that plaintiffs fail to allege facts plausibly satisfying the “double scienter” required under (a); that Stability knowingly provided false CMI with the intent to induce or enable infringement.

I agree on both counts. The generic license that accompanies use of Stable Diffusion on its face claims rights to Stable Diffusion as a work, not to the LAION dataset and not any works that were used to create the LAION dataset. It is implausible that a viewer reading the license disclosure for Stable Diffusion would understand that Stability is claiming rights to or conveying any false information regarding the rights of the plaintiffs whose copyrighted works are among the billions of images in the LAION datasets.

The 1202(a) claim is dismissed again, this time with prejudice as plaintiffs do not identify a basis for this claim.

2. Claim Under 1202(b)(1)

Similarly, Stability contends that plaintiffs have failed to plead plausible facts that Stability AI intentionally removed or altered CMI from plaintiffs' works during the training process for Stable Diffusion, and failed to allege the double-scienter requirement that Stability did so in order to facilitate infringement. In the FAC, plaintiffs allege that “Stability directly copied the LAION-5B Works and used these Statutory Copies as training data for the Stability Models. The works copied by Stability included CMI, including in the form of distinctive marks such as watermarks or signatures, and as the captions in the image-text pairs. The training process is designed to remove or alter CMI from the training images. Therefore, Stability intentionally removed or altered CMI from the Plaintiffs' works in violation of 17 U.S.C. § 1202(b)(1).” FAC ¶ 245. Plaintiffs support that allegation by pointing to some of plaintiffs' images that were used as Training Images in LAION-5B and that contained CMI, comparing them to images that were created, for example, when plaintiffs' names were used as Midjourney image prompts. FAC ¶¶ 189-200. Plaintiffs have also, as directed in the October 2023 Order, identified the CMI present on their works that they contend has been stripped by Stability. Id. ¶¶ 239-241. Finally, plaintiffs allege that Stability engaged in knowing removal of CMI, as supported by their allegations regarding how the diffusion process works, how training images are used, and based on plausible allegations regarding Stability AI's prominent role in the funding of LAION. Id. ¶¶ 245-247.

Stability, however, raises a new argument on this round of motions, based on a recent opinion from the Hon. Jon S. Tigar. Doe 1 v. GitHub, Inc., No. 22-CV-06823-JST, 2024 WL 235217, at *8 (N.D. Cal. Jan. 22, 2024). Stability contends that because the output images are admittedly not identical to the Training Images, there can be no liability for any removal of CMI that occurred during the training process. That is because failing to affix CMI to a “different work” is not “removal” under Section 1202. Judge Tigar wrote:

Defendants now ask the Court to address an unresolved argument from the prior briefing-namely, that “[Section] 1202(b) claims lie only when CMI is removed or altered from an identical copy of a copyrighted work.” ECF No. 107-3 at 20 (emphasis added); see ECF No. 109-3 at 23-24. Defendants argue that because Plaintiffs' new allegations state that output from Copilot is often a modification of their licensed works, as opposed to an “identical copy,” they have effectively pleaded themselves out of their Section 1202(b)(1) and 1202(b)(3) claims. ECF No. 109-3 at 23. Agreeing with Defendants on both fronts, the Court finds that it is not precluded from analyzing this claim anew and that Section 1202(b) claims require that copies be “identical.”

Id. * 8; see also Doe 1 v. GitHub, Inc., No. 22-CV-06823-JST, 2024 WL 1643691, at *2 (N.D. Cal. Apr. 15, 2024) (rejecting motion for reconsideration because there was no allegation that any output was identical to any plaintiff's work in its entirety, so there could be no “removal” of CMI for purposes of DMCA).¹¹

11 In so holding, Judge Tigar followed Advanta-STAR Auto. Rsch. Corp. of Am. v. Search Optics, LLC, No. 22-CV-1186 TWR, 2023 WL 3366534, at *12 (S.D. Cal. May 9, 2023); Kirk Kara Corp. v. W. Stone & Metal Corp., No . CV 20-1931-DMG, 2020 WL 5991503, at *6 (C.D. Cal. Aug. 14, 2020); Frost-Tsuji Architects v. Highway Inn, Inc., No. CIV. 13-00496 SOM, 2015 WL 263556, at *3 (D. Haw. Jan. 21, 2015), aff'd, 700 Fed.Appx. 674 (9th Cir. 2017) (no section 1202(b) violation where the allegedly infringing drawing was “not identical”).

Judge Tigar's analysis disagrees with a case from the Southern District of Texas, where the court rejected the “identicality” requirement. See ADR Int'l Ltd. v. Inst. for Supply Mgmt. Inc., 667 F.Supp.3d 411, 427 (S.D. Tex. 2023) (“[b]ased on the plain wording of the statute, the Court is not persuaded that the DMCA includes an ‘identical copy' requirement,” noting the copying should be “substantially similar” but does not have to be “perfect”).

Recognizing that this issue is unsettled, I agree with the reasoning of Judge Tigar that followed other district court decisions within the Ninth Circuit. Because there are no allegations that any output from Stable Diffusion was identical to a plaintiff's work, the DMCA section 1202(b) claim fails as well.¹²

12 The only allegations regarding identical outputs in the FAC are those regarding researcher's ability to reproduce identical images to training images - that as noted above supports the plausibility of the copyright infringement claims - but there are no allegations that Stable Diffusion can produce a work identical to one of the plaintiffs here, sufficient to show actionable removal of CMI under section 1202(b).

The DMCA claims against Stability are DISMISSED WITH PREJUDICE.¹³

13 Runway and Midjourney also move to dismiss the DMCA claim asserted against them. The arguments raised by those defendants, and plaintiffs' responses, are based on the defendants' use of Stable Diffusion and materially identical to the arguments addressed above. As a result, the DMCA claims asserted against Runway and Midjourney are likewise DISMISSED WITH PREJUDICE and will not be addressed further.

C. Unjust Enrichment

In addition to arguing that leave to amend was not granted to assert this claim, addressed above, Stability argues that the unjust enrichment claim cannot proceed because it is preempted by the Copyright Act. It also contends that plaintiffs fail to allege that they have an inadequate remedy at law as required by Sonner v. Premier Nutrition Corp., 971 F.3d 834 (9th Cir. 2020).

To support this claim, plaintiffs allege that Stability AI “has unjustly misappropriated the LAION-5B Works in order to develop, train and promote the Stability Models, enabling it to receive profit and other benefits.” FAC ¶252. “By using Plaintiffs' works to train, develop and promote the Stability Models, Plaintiffs and the Class were deprived of the benefit of the value of their works, including monetary damages.” Id. ¶ 254. Stability points out that this claim is expressly based on the use of plaintiffs' copyrighted works without consent, and as a result is covered by and preempted by the Copyright Act.

“Section 301 of the Act seeks ‘to preempt and abolish any rights under the common law or statutes of a State that are equivalent to copyright and that extend to works,' so long as the rights fall ‘within the scope of the Federal copyright law.' [] ‘We have adopted a two-part test,' in accordance with section 301, ‘to determine whether a state law claim is preempted by the Act. Laws,' [] First, we decide ‘whether the ‘subject matter' of the state law claim falls within the subject matter of copyright as described in 17 U.S.C. §§ 102 and 103.' [] Second, assuming it does, we determine ‘whether the rights asserted under state law are equivalent to the rights contained in 17 U.S.C. § 106, which articulates the exclusive rights of copyright holders.'” Maloney v. T3Media, Inc., 853 F.3d 1004, 1010 (9th Cir. 2017) (quoting Laws v. Sony Music Entm't, Inc., 448 F.3d 1134, 1137-38 (9th Cir. 2006)).

Plaintiffs correctly note that to “survive preemption, the state cause of action must protect rights which are qualitatively different from the copyright rights. The state claim must have an extra element which changes the nature of the action.” Laws, 448 F.3d at 1143. Plaintiffs do not identify an “extra element” required under California's Unfair Competition Law (“UCL”) or common law (which the unjust enrichment claims are based on) that changes the nature of those state law claims to protect something other than rights protected under Copyright Act. Instead, they assert that the heart of their unjust enrichment claim is Stability being unjustly enriched by its “image product's ability to mimic Plaintiffs' artistic style and benefit from their notoriety and reputation as sought-after artists.” Oppo. to Stability AI at 20. More specifically, they allege that the Stable Diffusion models use “CLIP-guided diffusion” that relies on prompts including artists' names to generate an image. Therefore, the “crux” of plaintiffs' unjust enrichment claim revolved not around plaintiffs' works but around plaintiffs themselves and their “artistic personas.” Id.

The problem with plaintiffs' theory is that it is not in the FAC. The unjust enrichment claim against Stability (and the other defendants) is tied instead to use of plaintiffs' works. See FAC ¶ 42 (“Whether the use of Plaintiffs and Class members' works to train, develop, and promote Defendants AI Image Products constitute an unjust benefit conferred upon Defendants to Plaintiffs' detriment”); see also id. ¶¶ 252-255 (alleging unjust misappropriation of works).

As alleged, the unjust enrichment claim against Stability added to the FAC without leave of court is preempted by the Copyright Act. It is DISMISSED. If plaintiffs have a good faith theory of unjust enrichment that falls outside the scope of the protections provided by the Copyright Act, they are given leave to make one last attempt to state an unjust enrichment claim.¹⁴

14 Because of this conclusion, I need not reach Stability's argument under Sonner v. Premier Nutrition Corp., 971 F.3d 834 (9th Cir. 2020). I have however, repeatedly held that a plaintiff satisfies Sonner at the pleading stage by simply pleading inadequate remedies at law. See Costa v. Apple, Inc., No. 23-CV-01353-WHO, 2023 WL 7389276, at *4 (N.D. Cal. Aug. 9, 2023). If plaintiffs reassert unjust enrichment claims in their Second Amended Complaint, they should address that issue. The unjust enrichment claims asserted against Runway and Midjourney raise identical issues. Those claims are, therefore, DISMISSED with one last leave to amend and will not be addressed further. The unjust enrichment claim asserted against DeviantArt is addressed below.

Stability's motion to dismiss is DENIED concerning the challenged copyright claims, but GRANTED for the DMCA claims without leave to amend and GRANTED for the unjust enrichment claim with leave to amend.

III. RUNWAY AI MOTION TO DISMISS

Newly added defendant Runway AI is primarily alleged to have trained or assisted in the training at least Stable Diffusion 1.5, using Training Images from the LAION dataset. See, e.g., FAC ¶¶ 4, 163, 176, 342-345. In addition to challenging plaintiffs' DMCA and unjust enrichment/UCL claims - which fail for the same reasons identified above with respect to Stability - Runway argues that the infringement claims fail given the allegations alleged with respect to its particular conduct.

A. Request for Judicial Notice

As an initial matter, Runway asks me to take judicial notice of a motion to dismiss filed in another case in this District, as well as the complaint from the case. Runway RJN (Dkt. No. 164), Exs. A&B. Plaintiffs oppose, arguing that judicial notice of court records from other cases is inappropriate for the purpose Runway seeks; to encourage me to follow the “approach” of those other courts. “On a Rule 12(b)(6) motion to dismiss, when a court takes judicial notice of another court's opinion, it may do so ‘not for the truth of the facts recited therein, but for the existence of the opinion, which is not subject to reasonable dispute over its authenticity.'” Southern Cross Overseas Agencies, Inc. v. Wah Kwong Shipping Group Ltd., 181 F.3d 410, 426-27 (3rd Cir.1999).” Lee v. City of Los Angeles, 250 F.3d 668, 690 (9th Cir. 2001). Notifying me of the existence of opinions from the judges in other cases may be accomplished by a far simpler method, citing them as persuasive authority. Judicial notice is not an appropriate method to suggest that I should follow the analysis of other courts addressing different cases with different facts.

Runway also asks me to take judicial notice - under the doctrine of incorporation - of the full contents of three academic articles cited in plaintiffs' FAC. Id., Exs. C, D & E. Plaintiffs object to this request, noting that one article is only mentioned once in the FAC and the others a few more times as support for the plausibility of plaintiffs' assertion that Stable Diffusion contains “compressed copies” of the Training Images based on how these models generally work. Plaintiffs argue that it is inappropriate to take “notice” of the debated truth, meaning or implications of the articles to foreclose their claims. I agree. I will not take judicial notice of the full contents of the academic articles to resolve disputes of fact, or the legal implications from undisputed facts, at this juncture.

Runway's request for judicial notice is DENIED.

B. Direct Copyright Infringement

Runway does not move to dismiss first direct copyright infringement claim asserted against it. See Count 11, FAC ¶¶ 347-349. That claim is based on Runway's alleged use of Training Images to train Stable Diffusion 1.5, a claim that survived Stability AI's prior motion to dismiss. October 2023 Order at 7 (discussing the “Training Theory”). Runway does challenge plaintiffs' two other direct infringement theories. First, the “Model Theory” is based on the theory that the Stable Diffusion1.5 product itself - after it was trained - is “an infringing Statutory Copy” of plaintiffs' works or a “Statutory Derivative Work” because it represents a transformation of plaintiffs' works. See FAC ¶¶ 209, 350. Second, Runway challenges plaintiffs' “Distribution Theory” of infringement, based on allegations that Runway infringes plaintiffs' exclusive distribution rights because distributing Stable Diffusion 1.5 is equivalent to distributing plaintiffs' works. FAC ¶ 352.

As with Stability, because Runway does not challenge the use of the images for training purposes, I need not address the other theories of direct infringement. However, I note that both the model theory and the distribution theory of direct infringement depend on whether plaintiffs' protected works are contained, in some manner, in Stable Diffusion as distributed and operated. That these works may be contained in Stable Diffusion as algorithmic or mathematical representations - and are therefore fixed in a different medium than they may have originally been produced in - is not an impediment to the claim at this juncture. 1 Nimmer on Copyright § 2.09[D][1] (2024) (“A work is no less a motion picture (or other audiovisual work) whether the images are embodied in a videotape, videodisc, or any other tangible form.”).

Plaintiffs addressed the deficiencies in their prior complaint, alleging additional facts in the FAC concerning how the training images remain in and are used by Stable Diffusion. See, e.g., FAC ¶¶ 71, 83, 88-90, 150.¹⁵ Plaintiffs rely, with respect to Runway as with Stability, on comments from Stability's CEO regarding the contents of the model and their ability to reproduce works, as well as academic papers indicating that the Stable Diffusion models are capable of producing very similar if not identical works to at least some training images. FAC ¶¶ 122-150. And plaintiffs also rely on use of their names as prompts in the Runway products to create outputs mimicking aspects of plaintiffs' protected works as evidence that their protected works are being copied or distributed in Runway's product. Id. ¶¶ 163-169.

15 Runway's and the other defendants' repeated reliance on “run of the mill” copyright cases where a showing of substantial similarity between works is required when determining whether an inference of copying can be supportedsee, e.g., Hanagami v. Epic Games, Inc., 85 F.4th 931, 935 (9th Cir. 2023) - or liability imposedsee, e.g., Narell v. Freeman, 872 F.2d 907, 910 (9th Cir. 1989) - are unhelpful in this case where the copyrighted works themselves are alleged to have not only been used to train the AI models but also invoked in their operation. Whether that use is “substantial enough” either in operation or output of images to qualify for the fair use defense will be tested on summary judgment. See, e.g., Authors Guild v. Google, Inc., 804 F.3d 202, 226 (2d Cir. 2015) (affirming summary judgment that Google's “program does not allow access in any substantial way to a book's expressive content”).

Runway disputes the accuracy of those assertions - including the full meaning and import of the academic articles relied on by plaintiffs - and argues that the prompt-examples do not support assertions of direct copyright infringement absent express identification of outputs that are “substantially similar” to plaintiffs' copyrighted works. But the allegations at this juncture are sufficient to allow the direct infringement claims to proceed. Whether evidence can support each of the theories and whether plaintiffs will need to choose between theories (e.g., between direct infringement based on selling a product containing effective copies of copyrighted works or violating plaintiffs' rights to restrict distribution of their works) will be addressed at summary judgment.¹⁶

16 Runway, like the other defendants, picks up on my questions from the prior oral argument to argue that plaintiffs should be able to point to source code from the “open source” AI products to identify where in the Stable Diffusion and defendants' AI products copies of plaintiffs' works are stored. See, e.g., Runway Reply at 4-5. In the FAC, plaintiffs challenge the use of the label “open source,” pointing out that various components of the models, including weights files, are not open for inspection by all. FAC ¶¶ 147-148. Those allegations satisfy the Court's prior queries. Runway also relies on a decision by a different judge in this court rejecting derivative infringement theories in Kadrey v. Meta Platforms, Inc., No. 23-CV-03417-VC, 2023 WL 8039640, at *1 (N.D. Cal. Nov. 20, 2023). In Kadrey, the Hon. Vince Chhabria considered copyright infringement with respect to copyrighted written works that were used to train large language models (“LLaMA”), software programs designed to produce naturalistic text outputs in response to user prompts. He dismissed plaintiffs' derivative copyright theories because plaintiffs failed to allege that the “LLaMA models themselves” could be understood “as a recasting or adaptation of any of the plaintiffs' books” and because there were no allegations that the outputs of those models could be “understood as recasting, transforming, or adapting the plaintiffs' books” or otherwise producing outputs that were “substantially similar” to aspects of plaintiffs' books. Kadrey, 2023 WL 8039640 at *1. The products at issue here - image generators allegedly trained on, relying on, and perhaps able to invoke copyrighted images - and the necessary allegations regarding the products' training and operations, are materially different from those in Kadrey. Whether substantial similarity remains a hurdle to specific theories - including any derivative infringement theory - depends in part on what the evidence shows concerning how these products operate and, presumably, whether and what the products can produce substantially similar outputs as a result of “overtraining” on specific images or by design.

Runway's motion to dismiss the direct infringement claims is DENIED.

C. Induced Infringement

Runway argues that plaintiffs have also failed to allege two elements of the induced infringement claim: acts of infringement by third parties using Runway's products and that Runway promoted use of Stable Diffusion to infringe. Runway claims that the only support for the induced infringement claim are comments plaintiffs identify by Stability executives, not statements by anyone associated with Runway.

Plaintiffs allege that Runway helped train and develop Stable Diffusion, and therefore, knew that the product allegedly uses or invokes the training images in its operation. Those allegations, combined with allegations that Runway actively induces others to download Stable Diffusion by distributing Stable Diffusion through popular coding websites and also by making selling its products (including AI Magic Tools) that include Stable Diffusion, are sufficient. FAC ¶¶ 83, 163, 185, 352, 358.¹⁷

17 Runway's argument that plaintiffs' induced infringement claim must fail because plaintiffs do not plead that Runway's models lack “substantial non-infringing uses” is rejected for the reasons discussed above with respect to Stability. See supra at 8-9.

Runway's motion to dismiss is DENIED on the infringement claims, GRANTED with prejudice on the DMCA claims, and GRANTED with leave to amend on the unjust enrichment/UCL claim.

IV. MIDJOURNEY MOTION TO DISMISS

In addition to moving to dismiss the DMCA and unjust enrichment claims addressed above, Midjourney moves to dismiss the copyright claims and Lanham Act false endorsement and trade dress claims¹⁸ asserted against it.

18 15 U.S.C. § 1125(a) et seq.

A. Copyright

1. Registration

Midjourney argues, with respect to three of the named plaintiffs - Anderson, Kaye, and Brom - that the evidence of their registration of newly identified copyrighted works is insufficient. With respect to Anderson and Kaye, Midjourney asserts that a subset of each artists' works identified as being both copyrighted and included in the LAION datasets used to train the AI products are compilations. It contends that copyright protection only extends to the new material in compilations and Anderson and Kaye fail to identify which works within the subset of compilations are the new material. It also argues that for two of three of the works identified by Brom, the copyright registrations extend only to text and not artwork. Midjourney Mot. at 6-8.

It is undisputed that plaintiffs who do not have valid copyright protections will not be able to pursue copyright claims based on un-registered works or works whose registrations covered only text. It is also undisputed that each of the named plaintiffs who claim their copyright-protected works were included in the LAION datasets have at least one work whose registration is facially valid. At this juncture, therefore, the Copyright Act claims survive against Midjourney and the other defendants. However, the identification in the FAC and exhibits of unprotected works is not irrelevant; plaintiffs rely on some of those works to plausibly demonstrate that their works were used as training images and that their works or elements of their works can be recreated through the AI products. The identification of those works may not prove liability under the Copyright Act, but they do provide support for the plausibility of plaintiffs' Copyright Act theories.

2. Use as Training Images

In my prior Order, I required plaintiffs to clarify the basis of their copyright claims against Midjourney, as plaintiffs' allegations regarding whether and how Midjourney used plaintiffs' images in the training of its product were insufficient. October 2023 Order at 13-14. In their FAC, plaintiffs now allege that Midjourney separately trained its product on the LAION400M and LAION5B datasets. See FAC ¶¶ 266, 274. Plaintiffs also allege that Midjourney incorporates Stable Diffusion into its own AI product. Id. ¶¶ 169-170.

Midjourney contests the adequacy of these new allegations with respect to the training of images, arguing plaintiffs must identify specific, individual registered works that each artist contends Midjourney actually used for training. Given the unique facts of this case - including the size of the LAION datasets and the nature of defendants' products, including the added allegations disputing the transparency of the “open source” software at the heart of Stable Diffusion - that level of detail is not required for plaintiffs to state their claims. Instead, plaintiffs have added to their FAC more detailed allegations regarding the training and use of the LAOIN datasets by defendants generally and Midjourney specifically. Plaintiffs have plausible allegations showing why they believe their works were included in the LAION datasets. And plaintiffs plausibly allege that the Midjourney product produces images - when their own names are used as prompts - that are similar to plaintiffs' artistic works. See FAC Exs. F&G.

Midjourney nonetheless argues that these examples are insufficient because some of the identified works are not registered and the resulting outputs could just as likely be the result of training on unregistered works or utilizing only unprotected elements from plaintiffs' works. But plaintiffs' reliance on those exhibits are not to establish copyright infringement as a matter of law.

Instead, they are relied on to support the plausibility of plaintiffs' copyright theories (that all or most of the works in the LAION datasets were used by Midjourney and the other defendants to train their AI products, and that plaintiffs' works or their protected elements that are contained in the AI products as the works or protected elements can be recreated by using the AI products). The FAC allegations and the exhibits help plaintiffs cross the plausibility threshold. Whether plaintiffs will be able to prove their claims is a different matter and those claims will be tested on an evidentiary basis at summary judgment.¹⁹

19 Echoing the arguments made by Runway, Midjourney also challenges the model and distribution theories. As above, given the FAC's plausible allegations regarding training of Midjourney's product and that “copies” or protected elements of some of some plaintiffs' registered works remain in Midjourney's AI product, these theories survive to be tested on summary judgment.

Midjourney's motion to dismiss the Copyright Act claims is DENIED.

B. Lanham Act

In place of the right of publicity claims asserted against Midjourney that were dismissed in my October 2023 Order, five plaintiffs (Anderson, Brom, Kaye, Ortiz and Rutkowski) now assert Lanham Act claims based on theories of false endorsement and trade dress. These plaintiffs allege that their names appeared on the list of 4700 artists posted by Midjourney's CEO on Discord, the platform where Midjourney's AI product operates. Midjourney's CEO promoted the list as describing the various styles of artistic works its AI product could produce. FAC ¶¶ 261-262, 305.²⁰ Plaintiffs also allege that Midjourney has itself published user-created images that incorporate the plaintiff artists' names in Midjourney's “showcase” site. Id. ¶ 325(b) & Ex. K.

20 FAC ¶ 305 (“Midjourney's use of the Midjourney Named Plaintiffs' names was purely to advertise its image generator. This use does not contribute significantly to a matter of public interest. The purpose of publishing over 4700 names in the Midjourney Name List was to promote and highlight the capabilities of Midjourney's image generator to emulate and create work that is indistinguishable from that of the artists whose names were published.”); ¶ 309 (“A reasonably prudent consumer in the marketplace for art products likely would be confused as to whether the Midjourney Named Plaintiffs included in the Midjourney Name List sponsored or approved of Midjourney's image generator.”).

1. False Endorsement

“To prevail on its Lanham Act trademark claim, a plaintiff must prove: (1) that it has a protectible ownership interest in the mark; and (2) that the defendant's use of the mark is likely to cause consumer confusion.” Rearden LLC v. Rearden Commerce, Inc., 683 F.3d 1190, 1202-203 (9th Cir. 2012) (quoting Network Automation, Inc. v. Advanced Sys. Concepts, Inc., 638 F.3d 1137, 1144 (9th Cir. 2011)). A false designation of origin claim likewise requires a showing of a likelihood of consumer confusion. New W. Corp. v. NYM Co. of Cal., Inc., 595 F.2d 1194, 1201 (9th Cir. 1979) (“Whether we call the violation infringement, unfair competition or false designation of origin, the test is identical[:] is there a ‘likelihood of confusion?'”).

Midjourney contends that plaintiffs fail to allege the first, necessary element of their false endorsement claim: falsity. They note that plaintiffs do not allege that the Midjourney AI product cannot recognize works by those included on the Midjourney Names List, and argue that just because the Names List exists and was promoted by the Midjourney CEO, that by itself cannot support an inference of endorsement. Midjourney points to a part of the Discord thread (identified by plaintiffs as the source of the Midjourney CEO's identification of the Names List) not relied on by plaintiffs, where the CEO indicated that the names on the list came from “wikipedia and magic the gathering.” Midjourney MTD, Dkt. No. 169, at 18. Midjourney argues that this part of the Discord thread is judicially noticeable and dispels any inference that the artists on the lists could plausibly be considered to have endorsed the Midjourney product.

As discussed above, judicial notice of other comments in the thread is not appropriate to dispute the facts plaintiffs otherwise plausibly assert, especially as the plaintiffs dispute the accuracy and inferences to be drawn from the totality of the messages in that thread. And even if the Names List itself was insufficient to support an inference of false endorsement, plaintiffs also allege that their names were used in connection with works included in Midjourney's “showcase.” FAC ¶ 325b. Whether or not a reasonably prudent consumer would be confused or misled by the Names List and showcase to conclude that the included artists were endorsing the Midjourney product can be tested at summary judgment. See Adobe Sys. Inc. v. A & S Elecs., Inc., 153 F.Supp.3d 1136, 1142 (N.D. Cal. 2015).²¹

21 While the court in Smith v. Chanel, Inc., 402 F.2d 562 (9th Cir. 1968) explained that the “Lanham Act does not prohibit a commercial rival's truthfully denominating his goods a copy of a design in the public domain, though he uses the name of the designer to do so,” that was a case where the seller expressly advertised its product as “equivalent” to the trademarked product, which, the court recognized was promoting competition. Id. at 565-66. That case was also admittedly not one with allegations of “misrepresentation or confusion as to source or sponsorship,” as here. Id. Whether the use of plaintiffs' names and works in Midjourney's advertising/promotion or the operation of its product was misleading or is truthful comparative advertising can be tested at summary judgment.

Midjourney also argues that to get past the First Amendment protection provided to expressive works, plaintiffs must, but have not, alleged that Midjourney's use of plaintiffs' names to invoke their styles has “‘no artistic relevance to the underlying work whatsoever, or, if it has some artistic relevance, unless the [use of trademark or other identifying material] explicitly misleads as to the source or the content of the work.'” Brown v. Elec. Arts, Inc., 724 F.3d 1235, 1239 (9th Cir. 2013) (quoting Rogers v. Grimaldi, 875 F.2d 994 (2d Cir.1989)). However, the plaintiffs have plausibly alleged that the use of their names by Midjourney in the List and showcase misleads consumers regarding source and endorsement. Open questions also remain whether Midjourney promoting its product for commercial gain for use by others to create artistic images is itself expressive use that creates “artistic relevance” to plaintiffs' underlying works. Discovery may show that it is or that is it not. Unlike in Brown v. Elec. Arts, Inc., where plaintiff alleged only that his likeness was used in a video game, we do not yet have the sort of record, or sort of alleged use, that could support dismissal of the claim at the motion to dismiss stage.

2. Vicarious Trade Dress

The same five plaintiffs (Anderson, Brom, Kaye, Ortiz and Rutkowski) also allege a trade dress claim based on the use of their names in connection with the Midjourney AI product's use of a “CLIP-guided model” that has been trained on the work of the Midjourney Named Plaintiffs and allows users to create works capturing the “trade dress of each of the Midjourney Named Plaintiffs [that] is inherently distinctive in look and feel as used in connection with their artwork and art products.” FAC ¶¶ 321. The CLIP model, plaintiffs assert, works as a trade dress database that can recall and recreate the elements of each artist's trade dress. FAC ¶¶ 83, 320. Plaintiffs point to examples showing how Midjourney recreates works with their trade dress in Ex. F to the FAC.

Midjourney argues that plaintiffs have failed to state this claim because they have not adequately identified the “concrete elements” of each plaintiff's trade dress. See, e.g., YZ Prods. Inc. v. Redbubble, Inc., 545 F.Supp.3d 756, 767 (N.D. Cal. 2021) (“images alone are insufficient to provide adequate notice dismissing trade dress claim,” and requiring plaintiff to identify “the concrete elements” of their protected trade dress). It acknowledges that plaintiffs have identified aspects of their trade dress, see FAC ¶ 319,²² but asserts that those descriptions are impermissibly broad. See, e.g., Walker & Zanger, Inc. v. Paragon Indus., Inc., 549 F.Supp.2d 1168, 1176 (N.D. Cal. 2007 (granting summary judgment on the “some of the elements of plaintiff's proposed trade dress [that] are overbroad”). Plaintiffs refute that characterization, arguing instead that they have identified the “set of recurring visual elements and artistic techniques, the particular combination of which are distinctive to each of the Midjourney Named Plaintiffs.” Id. ¶ 319.

22 FAC ¶ 319 (“ a. Sarah Andersen is known for work that is simple, cartoony, and often strictly in black and white. In particular, she is known for “Sarah's Scribbles,” a comic featuring a young woman with dark hair, big eyes, and a striped shirt. b. b. Karla Ortiz is known for a mixture of classical realism and impressionism, often delving into fantastical, macabre and surrealist themes, and inspired by the technical prowess of American Renaissance movements with a strong influence of contemporary media. c. Gerald Brom is known for gritty, dark, fantasy images, painted in traditional media, combining classical realism, gothic and counterculture aesthetics. d. Grzegorz Rutkowski is known for lavish fantasy scenes rendered in a classical painting style. e. Julia Kaye is known for three-panel black-and-white comics, loosely inked with a thin fixed-width pen, wherein each individual comic is a microvignette in the artist's life.”).

While the images from Exhibit F on their own would be insufficient identification, and while some of the alleged “concrete elements” identified in the FAC are, standing alone, vague and possibly overbroad, those elements cannot be considered alone but as a whole in the context of plaintiffs' other, plausible allegations. Arcsoft, Inc. v. Cyberlink Corp., 153 F.Supp.3d 1057, 1068 (N.D. Cal. 2015). Here, the combination of identified elements and images, when considered with plaintiffs' allegations regarding how the CLIP model works as a trade dress database, and Midjourney's use of plaintiffs' names in its Midjourney Name List and showcase, provide sufficient description and plausibility for plaintiffs' trade dress claim.

Midjourney also argues that artistic elements or styles identified for each artist that are allegedly re-creatable by using its product are functional, and therefore not protected.²³ It ignores, however, the Ninth Circuit's test for determining non-functionality and instead relies on a series of inapposite cases addressing jewelry, wooden cutouts, and keychains. See, e.g., Int'l Ord. of Job's Daughters v. Lindeburg & Co., 633 F.2d 912, 918 (9th Cir. 1980) (use of name and emblem of fraternal organization were functional aesthetic components of jewelry); Crafty Prods., Inc. v. Michaels Companies, Inc., 424 F.Supp.3d 983, 993 (S.D. Cal. 2019), aff'd sub nom. Crafty Prods., Inc. v. Fuqing Sanxing Crafts Co., 839 Fed.Appx. 95 (9th Cir. 2020) (plaintiffs failed to identify “concrete elements” of their trade dress and instead attempted to capture the “entire design” of hundreds of different wooden cardboard pieces); Au-Tomotive Gold, Inc. v. Volkswagen of Am., Inc., 457 F.3d 1062, 1064 (9th Cir. 2006) (reversing summary judgment protecting defendants' use of trademarked symbols in keychains and license plate holders, as “[t]he doctrine of aesthetic functionality does not provide a defense against actions to enforce the trademarks against such poaching” or source identification).

23 See Arcsoft, 153 F.Supp.3d at 1068 (“To determine whether a claimed trade dress is functional, the Ninth Circuit considers several factors: ‘(1) whether the design yields a utilitarian advantage, (2) whether alternative designs are available, (3) whether advertising touts the utilitarian advantages of the design, and (4) whether the particular design results from a comparatively simple or inexpensive method of manufacture.'” (quoting Apple Inc. v. Samsung Elecs. Co., 786 F.3d 983, 991 (Fed.Cir.2015)). “In applying these factors and evaluating functionality, ‘it is crucial that [the court] focus not on the individual elements, but rather on the overall visual impression that the combination and arrangement of those elements create.'” Arcsoft, 153 F.Supp.3d at 1068 (quoting Clicks Billiards, Inc. v. Sixshooters, Inc., 251 F.3d 1252, 1259 (9th Cir.2001)).

Midjourney complains that plaintiffs intend to assert a monopoly over some of the elements plaintiffs identify as critical parts of their protected trade dress. But that ignores the plausible allegations that the CLIP model functions as a trade dress database and the use of its product to produce works based on the names of these plaintiffs. Plaintiffs have no protection over “simple, cartoony drawings” or “gritty fantasy paintings.” Midjourney Reply, Dkt. No. 184 at 13. But their trade dress claims must be considered viewing all identified elements as well as the nature of the use of the CLIP model and their names. Those issues will be tested on an evidentiary basis. This claim will not be dismissed based on “aesthetic functionality.”

Midjourney contends that the trade dress claim must nonetheless be dismissed because plaintiffs allege their trade dress is “either” distinctive or, in the alternative, has acquired secondary meaning.²⁴ However, plaintiffs pleaded secondary meaning. See FAC ¶ 328 (“Each of the Midjourney Named Plaintiffs' trade dress possesses secondary meaning because the trade dress of their art products invoke a mental association by a substantial segment of potential consumers between the trade dress and the creator of the art product.”). Proof of intentional copying by Midjourney has been alleged, especially considering the allegations regarding the CLIP model functioning as a trade dress database, the express use of plaintiffs' names in the Midjourney Names List to promote the product and the use of some plaintiffs' names and likenesses of their works in the Midjourney showcase, and the “mental recognition” of Midjourney's calling out of plaintiffs by name.

24 As the case relied on by Midjourney, Art Attacks Ink, LLC v. MGA Ent. Inc., 581 F.3d 1138, (9th Cir. 2009), explains: “‘Secondary meaning can be established in many ways, including (but established place in the market; and proof of intentional copying by the defendant.' Filipino Yellow Pages, Inc. v. Asian Journal Publ'ns, Inc., 198 F.3d 1143, 1151 (9th Cir.1999). To show secondary meaning, a plaintiff must demonstrate ‘a mental recognition in buyers' and potential buyers' minds that products connected with the [mark] are associated with the same source.' Japan Telecom v. Japan Telecom Am., 287 F.3d 866, 873 (9th Cir.2002) (internal quotation omitted).” 581 F.3d at 1145.

Finally, Midjourney contests the sufficiency of plaintiffs' vicarious trade dress claim, arguing that plaintiffs have failed to allege facts of “joint ownership or control” over the infringing product by Midjourney and any end user. Plaintiffs allege:

Midjourney exercises control over the infringing images by including the CLIP model in its image pipeline, and by marketing artist-name prompts as a key feature of its image generator via the Midjourney Name List. Without the CLIP model, Midjourney's users would not be able to infringe on the Midjourney Named Plaintiffs' trade-dress rights or those of the other artists on the Midjourney Name List.

FAC ¶ 326. These allegations support the claim for vicarious trade dress infringement. See, e.g., Y.Y.G.M. SA v. Redbubble, Inc., No. 219CV04618RGKJPR, 2020 WL 3984528, at *9 (C.D. Cal. July 10, 2020 (“To impose vicarious liability the shared control must extend in some way to the infringed intellectual property itself.”).

Midjourney's motion to dismiss the Lanham Act claims is DENIED.²⁵

25 I acknowledge Midjourney's reliance on treatises and cases that caution against “extending trademark law to intrude into the domain of copyright law.” See McCarthy on Trademarks and Unfair Competition § 10:40.50 (5th ed.); see also id. (“Is there such a thing as trademark protection for the ‘style' of an artist? Courts have almost uniformly said no.”); Leigh v. Warner Bros., 10 F.Supp.2d 1371, 1380-81 (S.D. Ga. 1998), aff'd in part, rev'd in part on other grounds, 212 F.3d 1210 (11th Cir. 2000) (no trademark claim based on use of plaintiff's not limited to) direct consumer testimony; survey evidence; exclusivity, manner, and length of use of a mark; amount and manner of advertising; amount of sales and number of customers; copyrighted photograph because the work “merely identifie[d] the artist rather than any products or services” he sold, it could not “be protected as a trademark,” even if it was an example of his “unique artistic style.”). These cases do not readily fit the allegations plaintiffs raise regarding how Midjourney's product and other CLIP model products function as trade dress databases, combined with Midjourney's use of plaintiffs' names and showcase examples calling out named plaintiffs. This argument is better determined on a full record.

C. Request for Judicial Notice

Midjourney requests judicial notice of: (1) messages from a Discord thread, where Midjourney CEO David Holz released “our style list” and posted a link to a spreadsheet on Google Docs called “Midjourney Style List” that contained the named of 4700 artists, under the doctrine of incorporation by reference, Dkt. No. 170-1, Ex. 1; (2) messages in a different Discord thread that plaintiffs rely on in part, where Holz discusses Midjourney's image-prompt tool, Dkt. No. 170-1, Ex. 2; (3) another Discord thread discussing the image-prompt tool; and (4) a transcript from the hearing before Judge Chhabria in the Kadrey case.

As noted before, judicial notice of documents from the Kadrey case is not necessary, as either side can point to the orders from that case to argue their persuasiveness or differences with respect to the sufficiency of the allegations in this case. See supra at 15-16. Judicial notice is also not appropriate for any of the three Discord threads to dispute the facts plausibly alleged. For example, Midjourney does not dispute that the list of artists was disclosed and promoted by its CEO, which is the central reason for which plaintiffs rely on that Discord thread and its attachment. Midjourney seeks to rely on the thread to dispute the meaning of the comments by Holz and point to messages that are not relied on by plaintiffs (and whose meanings are disputed by plaintiffs) to foreclose plaintiffs' claims. That is not appropriate.

Midjourney's request for judicial notice is DENIED.

V. DEVIANTART MOTION TO DISMISS

A. Copyright Claim

DeviantArt moves again to dismiss the copyright claims asserted against it by attempting to set itself apart from the other defendants. It points out that it has not been alleged to have trained any AI model, but simply to have implemented and used the AI tools provided by Stability and others. It argues that holding it liable for that conduct would make the millions of third parties who have downloaded and implemented the AI products challenged in this case - many of which are open source software - liable for infringement, which is unsupportable.

In the October 2023 Order, I dismissed the copyright infringement claim asserted against DeviantArt, explaining:

In addition to providing clarity regarding their definition of and theory with respect to the inclusion of compressed copies of Training Images in Stable Diffusion, plaintiffs shall also provide more facts that plausibly show how DeviantArt is liable for direct copyright infringement when, according to plaintiffs' current allegations, DeviantArt simply provides its customers access to Stable Diffusion as a library. Plaintiffs do cite testimony from DeviantArt's CEO that DeviantArt uses Stable Diffusion because Stability allowed DeviantArt to “modify” Stable Diffusion. Compl. ¶ 129. The problem is that there are no allegations what those modifications might be or why, given the structure of Stable Diffusion, any compressed copies of copyrighted works that may be present in Stable Diffusion would be copied within the meaning of the Copyright Act by DeviantArt or its users when they use DreamUp. Nor do plaintiffs provide plausible facts regarding DeviantArt “distributing” Stable Diffusion to its users when users access DreamUp through the app or through DeviantArt's website.

October 2023 Order at 10.

DeviantArt argues, first, that plaintiffs' copyright act claims are barred by the October 2023 Order because plaintiffs still do not allege that DeviantArt itself copied or used their works to train Stable Diffusion or trained any other program. The FAC alleges only that DeviantArt incorporates and relies on Stable Diffusion for its DreamUp product. FAC ¶¶ 6, 387-392, 395-397. Plaintiffs have added allegations to their FAC, however, regarding how copies or protected elements of their works remain, in some format, in Stable Diffusion and how those works can be invoked by use of all of the Stable Diffusion versions. The actual operation of Stable Diffusion 1.4 and whether the amount of any plaintiff's copyrighted works in that program suffices for copyright infringement or a fair use defense concerning DeviantArt remains to be tested at summary judgment.

Moreover, while plaintiffs admittedly did not include assertions or examples in the FAC of outputs from DreamUp that appear to copy elements of their works (as they did for the other defendants), their added allegations regarding the use of the LAION datasets to train Stable

Diffusion and how Stable Diffusion versions operates, including specific examples and academic references to the operation of Stable Diffusion 1.4 used by DreamUp, suffice. See FAC ¶¶ 132137, 388-397. For example, in paragraph 393 plaintiffs assert:

On information and belief, by the end of training, Stable Diffusion 1.4 was capable of reproducing protected expression from each of the LAION-5B Registered Works that was in each case substantially similar to that registered work, because- a. In the Carlini Paper, Nicholas Carlini tested Stable Diffusion 1.4 and found that it could emit stored copies of its training images; b. The training procedure for Stable Diffusion 1.4 was very similar to that of Stable Diffusion 1.5, which was shown in Exhibit E: Runway text prompts and Exhibit H: Runway image prompts to be capable of emitting stored copies of protected expression.

See also ¶ 394 (“Therefore, like Stable Diffusion 1.5, Stable Diffusion 1.4 also qualifies as an infringing Statutory Copy of the LAION-5B Registered Works. Because Stable Diffusion 1.4 represents a transformation of the LAION-5B Registered Works into an alternative form, Stable Diffusion 1.4 also qualifies as an infringing Statutory Derivative Work.”).

DeviantArt asks me to review to the full content of one of the academic articles plaintiffs rely on, the Carlini Study, in particular the article's conclusion that for the 350,000 training images studied, only 109 output images were “near-copies” of the training images. See DeviantArt Reply at 4-5. DeviantArt argues that it is simply not plausible that LAION 1.4 can reproduce any of plaintiffs' copyrighted works given that the set of training images Carlini selected were the “most-duplicated” examples, numbering in the millions of impressions in the datasets. Id. Plaintiffs' reference to Carlini, however, is only one part of the allegations that help make plaintiffs' allegations plausible regarding how these products operate and how “copies” of the plaintiffs' registered works are captured in some form in the products. As noted above, it is not appropriate to rely on the defendants' assertions regarding the method or sample size used by Carlini and their conclusions from it to foreclose plaintiffs' claims; plaintiffs aggressively dispute the implications from the sample size and results of that study. Finally, while the differences between Stable Diffusion 1.4 and the other Stable Diffusion versions might be legally significant based on what the evidence shows after discovery, at this juncture the allegations about the common training of those versions and how they all operate are sufficient to keep plaintiffs' copyright infringement allegations against DeviantArt alive.²⁶

26 Given this conclusion, I need not separately reach the question of whether plaintiffs can assert a theory of derivative copyright infringement based on the use of the diffusion model itself. As defendants note, the soundness of this theory was questioned in my October 2023 Order and by other judges in this District, with respect to output images. See October 2024 Order at 10-13; Tremblay v. OpenAI, Inc., No. 23-CV-03223-AMO, 2024 WL 557720, at *3 (N.D. Cal. Feb. 12, 2024). Judge Chhabria also rejected the derivative theory as applied to the AI model in Kadrey v. Meta Platforms, Inc., No. 23-CV-03417-VC, 2023 WL 8039640, at *1 (N.D. Cal. Nov. 20, 2023). But as noted above, the allegations regarding the training and operation of the language models at issue in Kadrey are significantly different than the image creation models at issue here.

Finally, DeviantArt contends that any use of plaintiffs' works in Stable Diffusion 1.4 should be considered fair use as a matter of law, given the huge size of the training datasets and plaintiffs apparent inability to use Stable Diffusion 1.4 to reproduce any works that look similar to their copyrighted works. Whether DreamUp operates in a way that could draw upon or otherwise reproduce plaintiffs' works to an extent that violates the Copyright Act and whether a fair use defense applies are issues that must be tested on an evidentiary basis. See, e.g., Authors Guild v. Google, Inc., 804 F.3d 202, 226 (2d Cir. 2015) (applying fair use defense upon summary judgment record).

DeviantArt's motion to dismiss the Copyright Act claims is DENIED.

B. Breach of Contract

DeviantArt moves again to dismiss the breach of contract claim that plaintiffs assert based on the same provision of DeviantArt's Terms of Service (“ToS”) that I considered in the October 2023 Order. In that Order I explained:

In opposition, plaintiffs focus on § 16 of the TOS, a provision not identified or quoted in their Complaint:

16. Copyright in Your Content

DeviantArt does not claim ownership rights in Your Content. For the sole purpose of enabling us to make your Content available through the Service, you grant to DeviantArt a nonexclusive, royalty-free license to reproduce, distribute, reformat, store, prepare derivative works based on, and publicly display and perform Your Content. Please note that when you upload Content, third parties will be able to copy, distribute and display your Content using readily available tools on their computers for this purpose although other than by linking to your Content on DeviantArt any use by a third party of your Content could violate paragraph 4 of these Terms and Conditions [preserving copyright rights in the original owner

of the copyright and disclaiming any ownership interest of DeviantArt in the posted work] unless the third party receives permission from you by license.

Plaintiffs argue that DeviantArt breached this provision when “it incorporated Stable Diffusion into its own AI Image product knowing that Stability had scraped DeviantArt's artists' work.” Oppo to DeviantArt MTD [Dkt. No. 65] at 22. However, section 16 provides a limited license to DeviantArt and warns that third parties may be able to copy and violate content-owners' rights. It does not clearly cover the conduct that plaintiffs accuse DeviantArt of in this suit; offering for use a product that a third party may have created in part by using material posted on DeviantArt's own site. There are no facts alleged supporting an allegation that DeviantArt itself exceeded the scope of the limited license.

DeviantArt also challenges the ability of plaintiff McKernan and the unspecified “others” to sue DeviantArt for breach claims based on contractual provisions prohibiting other users (presumably here, Stability) from using DeviantArt content for commercial uses. The Complaint does not allege and is devoid of facts supporting the inference that Stability is bound by the TOS or that plaintiff McKernan or others are third party beneficiaries of specific provisions in the TOS who may sue to enforce terms of agreements entered between DeviantArt and Stability.

The breach of contract claim is DISMISSED with leave to amend. If plaintiffs attempt to amend this claim, they must identify the exact provisions in the TOS they contend DeviantArt breached and facts in support of breach of each identified provision. To the extent plaintiffs rely on provisions that appear to protect or benefit DeviantArt but not the users, or contracts.

In the FAC, plaintiffs reallege a violation of Section 16, based on the same theory as above, and DeviantArt argues it fails for the same reason. See FAC ¶¶ 420-422. I agree. For the reasons discussed in the October 2023 Order, plaintiffs cannot state a breach of contract claim as a matter of law based on allegations that DeviantArt knew Stable Diffusion was trained on LAION datasets that had been scraped in part from DeviantArt's website. As before, there are no allegations that DeviantArt did anything to permit the scraping of images from its site. Plaintiffs admit that DeviantArt played no role in the scraping or training. Nothing in the TOS precludes DeviantArt from using Stable Diffusion, even if it had knowledge that some of the images used in its training were scraped from its own site and its members' works.

Plaintiffs argue that DeviantArt breached this provision because it is now using - through its use of Stable Diffusion - its members' works for purposes beyond “the sole purpose of enabling us to make your Content available through the Service.” But nothing in Section 16 limits DeviantArt's ability to use plaintiffs' works that are available from another source, i.e., the LAION datasets as incorporated into Stable Diffusion. The breach claim fails, again, as a matter of law.

Plaintiffs added a breach of the implied covenant claim in the FAC, premised also on a breach of Section 16 of the ToS. See ¶ 422b; Dkt. No. 177 at 16.²⁷ “[T]o state a claim for breach of the implied covenant of good faith and fair dealing, a plaintiff must identify the specific contractual provision that was frustrated.” Rockridge Tr. v. Wells Fargo, N.A., 985 F.Supp.2d 1110, 1156 (N.D. Cal. 2013). DeviantArt argues that the implied covenant claim fails because plaintiffs do not identify any contractual provision that obligated DeviantArt to protect its users from competition from DeviantArt or anyone else, or to offer protection from data scraping, or that would otherwise preclude DeviantArt from using a tool developed by third parties even if that tool was based in part on works scraped from its site without its involvement. I agree. There is no basis to find any provision of the ToS was frustrated by DeviantArt's alleged conduct.

27 FAC ¶422b provides: “DeviantArt breached the implied covenant of good faith and fair dealing. The release of DreamUp unleashed a flood of AI-generated images on DeviantArt that immediately began drowning out the work of human artists like the DeviantArt Plaintiffs. By releasing DreamUp, DeviantArt put itself into competition with the DeviantArt Plaintiffs and its other artist members, undermining their very purpose in being on DeviantArt in the first place. DeviantArt's bad faith was further exemplified by its hasty addition of a permissive new ‘Data Scraping & Machine Learning Activities' provision to its Terms of Service after DeviantArt's members complained about the unfairness of DreamUp.”

The breach of contract claim is DISMISSED. As plaintiffs did not contest this claim at oral argument, despite my tentative ruling order identifying my intent to dismiss this claim (Dkt. No. 193), and did not suggest any facts they could allege to salvage their breach claim in an amended complaint, this claim is DISMISSED with prejudice.

C. Unjust Enrichment

Consistent with the analysis above, the newly added unjust enrichment claim against DeviantArt based on use of plaintiffs' “works to develop and promote DreamUp and the DreamUp-CompVis Model” that deprived plaintiffs “the benefit of the value of their works” FAC ¶ 433, is dismissed as preempted under the Copyright Act. In their opposition to DeviantArt's motion, plaintiffs imply that the unjust enrichment claim against DeviantArt could be based upon a different ground; misrepresentations made by DeviantArt to plaintiffs. Plaintiffs' Oppo. to DeviantArt at 19. Plaintiffs are given one last attempt to amend their unjust enrichment claims against each defendant. If the theory underlying plaintiffs' unjust enrichment claim against DeviantArt rests on different facts and theories from the unjust enrichment claim asserted against the other defendants, plaintiffs should make that clear.

CONCLUSION

Defendants' motions to dismiss the DMCA claims are GRANTED and the DMCA claims are DISMISSED WITH PREJUDICE. Defendants' motions to dismiss the unjust enrichment claims are GRANTED and those claims are DISMISSED with leave to amend. Defendants' motions to dismiss the Copyright Act claims are DENIED. Midjourney's motion to dismiss the Lanham Act claims is DENIED. DeviantArt's motion to dismiss the breach of contract and breach of the implied covenant of good faith and fair dealing claims is GRANTED and those claims are DISMISSED WITH PREJUDICE.