Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 1 of 33

4 UNITED STATES DISTRICT COURT

5 NORTHERN DISTRICT OF CALIFORNIA

7 SARAH ANDERSEN, et al., Case No. 23-cv-00201-WHO

Plaintiffs,
8
ORDER GRANTING IN PART AND
v. DENYING IN PART MOTIONS TO
9
DISMISS FIRST AMENDED
10 STABILITY AI LTD., et al., COMPLAINT
Defendants. Dkt. Nos. 160, 162, 163, 164, 169
11

12 Defendants Stability AI Ltd. and Stability AI, Inc. (collectively “Stability AI”),
Northern District of California
United States District Court

13 Midjourney, Inc., DeviantArt, Inc., and Runway AI, Inc. move to dismiss various claims from

14 plaintiffs’ First Amended Complaint (“FAC,” Dkt. No. 129). For the reasons discussed below,

15 those motions are GRANTED in part and DENIED in part.

16 BACKGROUND

17 Artists Sarah Andersen, Kelly McKernan, Karla Ortiz, Hawke Southworth, Grzegorz

18 Rutkowski, Gregory Manchess, Gerald Brom, Jingna Zhang, Julia Kaye, and Adam Ellis

19 (“Plaintiffs”1) filed this putative class action on behalf of artists challenging the defendants’

20 creation and/or use of Stable Diffusion, an artificial intelligence (“AI”) software product. They

21 allege that Stable Diffusion used plaintiffs’ artistic works as “training images” and as a result

22 Stable Diffusion can produce output images “in the style” of those images. See generally FAC. In

23 the FAC, plaintiffs allege claims against the three defendants identified in the original complaint

24 (Stability AI, Midjourney, and DeviantArt) and against a new defendant, Runway AI.2 Plaintiffs

25
1
26 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
27 (H. Southworth PKA Hawke Southworth, Grzegorz Rutkowski, Gregory Manchess, Gerald Brom,
Jingna Zhang, Julia Kaye, and Adam Ellis). FAC ¶¶ 14-23.
28 2
The basic factual allegations regarding how defendants’ AI products were trained and work were
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 2 of 33

1 contend that Runway AI worked with, helped train, and then distributed Stable Diffusion with

2 Stability AI. Plaintiffs assert that Runway made a text-to-image generator available via its online

3 AI image product called AI Magic Tools.

4 Plaintiffs’ claims center first around the creation of the LAION training sets, where five

5 billion images were allegedly scraped into datasets used by Stability and Runway to train the

6 versions of Stable Diffusion. FAC ¶ 4. Plaintiffs state that Midjourney likewise trained its

7 product using Stable Diffusion, and that all four defendants use Stable Diffusion in their AI

8 products; in doing so, those four defendants copy or utilize versions of plaintiffs’ artistic works.

9 Id. ¶¶ 5-6. Significantly, plaintiffs allege that the “LAION-5B dataset contains only URLs of

10 training images, not the actual training images. Therefore, anyone who wishes to use LAION-5B

11 for training their own machine learning model must first acquire copies of the actual training

12 images from their URLs using the img2dataset or other similar tool.” FAC ¶ 221. They also
Northern District of California
United States District Court

13 clarify their theory of direct infringement, adding allegations regarding CLIP-guided diffusion in

14 the training phase but also in use, after training. Id. ¶¶ 82-150.

15 Plaintiffs assert claims on behalf of six different classes:

16 “Injunctive Relief Class” under Rule 23(b)(2): All persons or entities nationalized or

17 domiciled in the United States that own a copyright interest in any work that was used to train any

18 version of an AI image product that was offered directly or incorporated into another product by

19 one or more Defendants during the Class Period. FAC ¶ 34.

20 “Damages Class” under Rule 23(b)(3): All persons or entities nationalized or domiciled in

21 the United States that own a copyright interest in any work that was used to train any version of an

22 AI image product that was offered directly or incorporated into another product by one or more

23 Defendants during the Class Period. Id.

24 “LAION-5B Damages Subclass” under Rule 23(b)(3): All persons or entities nationalized

25 or domiciled in the United States that own a registered copyright in any work in the LAION-5B

26
27
identified in my prior Order and will not be repeated here. See October 30, 2023, Order at Dkt.
28 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.
2
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 3 of 33

1 dataset that was used to train any version of an AI image product that was offered directly or

2 incorporated into another product by one or more Defendants during the Class Period.3 Id.

3 “LAION-400M Damages Subclass” under Rule 23(b)(3): All persons or entities

4 nationalized or domiciled in the United States that own a registered copyright in any work in the

5 LAION-400M dataset that was used to train any version of an AI image product that was offered

6 directly or incorporated into another product by one or more Defendants during the Class Period.

7 Id.

8 “DeviantArt Damages Subclass” under Rule 23(b)(3): All members of the Damages Class

9 who (1) maintained an account on DeviantArt; (2) posted copyrighted work on DeviantArt; and

10 (3) had that work used to train any version of an AI image product. Id.

11 “Midjourney Named Artist Class” under Rule 23(b)(3): All persons or entities who appear

12 on the Midjourney Names List and whose names were invoked within prompts of the Midjourney
Northern District of California
United States District Court

13 Image Product during the Class Period.4 Id.

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

15 Against Stability AI: (1) direct copyright infringement of the LAION-5B Registered

16 Works by training the Stability Models, including Stable Diffusion 2.0 and Stable Diffusion XL

17 1.0 on behalf of the LAION-5B Registered Plaintiffs and Damages Subclass; (2) inducement of

18 copyright infringement by distributing Stable Diffusion 2.0 and Stable Diffusion XL 1.0 for free

19 on behalf of the LAION-5B Registered Plaintiffs and Damages Subclass; (3) violations of the

20 Digital Millennium Copyright Act (“DMCA”) by removing and altering copyright management

21 information (“CMI”) of training images on behalf of all Plaintiffs, the Damages and the Injunctive

22 Classes; and (4) unjust enrichment under Cal. Bus. & Prof. Code § 17200 and California Common

23

24 3
The “LAION-5B Registered Plaintiffs” as used in the FAC include “the subset of plaintiffs who
25 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
26 Manchess, Julia Kaye, and Adam Ellis.” FAC ¶ 213.

27
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
28 Midjourney’s product; Grzegorz Rutkowski, Sarah Andersen, Karla Ortiz, Gerald Brom, and Julia
Kaye. FAC ¶¶ 263, 264.
3
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 4 of 33

1 Law on behalf of all Plaintiffs, the Damages and the Injunctive Classes.

2 Against Runway AI: (1) Direct copyright infringement of the LAION-5B Registered

3 Works by training the Runway Models, including Stable Diffusion 1.5 on behalf of the LAION-

4 5B Registered Plaintiffs, LAION-5B Subclass, and Karla Ortiz Individually; (2) Inducement of

5 copyright infringement by distributing Stable Diffusion 1.5 for free on behalf of the LAION-5B

6 Registered Plaintiffs and Subclass; (3) DMCA violations by removing and altering CMI of

7 training images on behalf of all Plaintiffs, the Damages and Injunctive Classes; and (4) Unjust

8 enrichment under Cal. Bus. & Prof. Code § 17200 and California Common Law on behalf of all

9 Plaintiffs, the Damages and Injunctive Classes.

10 Against Midjourney: (1) Direct copyright infringement of the LAION-400M Registered

11 Works by training the Midjourney 400M Models, including Midjourney Model version 1 on

12 behalf of the LAION-400M Registered Plaintiffs and Damages Subclass; (2) Direct copyright
Northern District of California
United States District Court

13 infringement of the LAION-5B Registered Works by training the Midjourney 5B Models,

14 including Midjourney Model version 5.2 on behalf of the LAION-5B Registered Plaintiffs and

15 Damages Subclass; (3) DMCA violations by removing and altering CMI of training images

16 on behalf of All Plaintiffs, the Damages and Injunctive Class; (4) Lanham Act — false

17 endorsement by unauthorized commercial use of artists’ names on behalf of the Midjourney

18 Named Plaintiffs and Class; (5) Lanham Act — vicarious trade-dress violation by profiting from

19 imitations of protectable trade dress on behalf of the Midjourney Named Plaintiffs and Class; and

20 (6) Unjust enrichment under Cal. Bus. & Prof. Code § 17200 and California Common Law on

21 behalf of all Plaintiffs, and the Damages and Injunctive Class.

22 Against DeviantArt: (1) Direct copyright infringement by copying the DreamUp–CompVis

23 Model and incorporating it into DreamUp on behalf of the LAION-5B Registered Plaintiffs; (2)

24 Breach of contract for violation of its Terms of Service on behalf of the DeviantArt Plaintiffs; (3)

25 Unjust enrichment under Cal. Bus. & Prof. Code § 17200 and California Common Law on behalf

26 of the DeviantArt Plaintiffs.

27 In October 2023, I largely granted the motions to dismiss brought by defendants Stability,

28 Midjourney and DeviantArt. The only claim that survived was the direct infringement claim
4
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 5 of 33

1 asserted against Stability, based on Stability’s alleged “creation and use of ‘Training Images’

2 scraped from the internet into the LAION datasets and then used to train Stable Diffusion.”

3 October 2023 Order at 7. The remainder of the claims were dismissed with leave, so that plaintiffs

4 could amend “to provide clarity regarding their theories of how each defendant separately violated

5 their copyrights, removed or altered their copyright management information, or violated their

6 rights of publicity and plausible facts in support.” Id. at 1.

7 As one example, I required plaintiffs on amendment to address the following deficiencies

8 with a second theory of direct infringement, separate from the creating and use of images for

9 training theory:

10 Plaintiffs will be required to amend to clarify their theory with respect

to compressed copies of Training Images and to state facts in support
11 of how Stable Diffusion – a program that is open source, at least in
part – operates with respect to the Training Images. If plaintiffs
12 contend Stable Diffusion contains “compressed copies” of the
Northern District of California

Training Images, they need to define “compressed copies” and

United States District Court

13 explain plausible facts in support. And if plaintiffs’ compressed

copies theory is based on a contention that Stable Diffusion contains
14 mathematical or statistical methods that can be carried out through
algorithms or instructions in order to reconstruct the Training Images
15 in whole or in part to create the new Output Images, they need to
clarify that and provide plausible facts in support.
16
October 2023 Order at 8-9; see also id. 9 n.7 (“Plaintiffs’ second theory of direct infringement –
17
that Stable Diffusion is a ‘derivative work’ because it contains compressed copies of billions of
18
copyrighted images and by incorporating Stable Diffusion into” defendants’ own AI products,
19
defendants are “liable for producing works that have been ‘transformed’ based on plaintiffs’ works
20
[] fails for the same reasons.”).
21
Plaintiffs added defendant Runway AI and seven new plaintiffs when it filed the FAC.
22
Each defendant moves to dismiss.
23
LEGAL STANDARD
24
Under FRCP 12(b)(6), a district court must dismiss a complaint if it fails to state a claim
25
upon which relief can be granted. To survive a Rule 12(b)(6) motion to dismiss, the plaintiff must
26
allege “enough facts to state a claim to relief that is plausible on its face.” Bell Atl. Corp. v.
27
Twombly, 550 U.S. 544, 570 (2007). A claim is facially plausible when the plaintiff pleads facts
28
5
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 6 of 33

1 that “allow the court to draw the reasonable inference that the defendant is liable for the

2 misconduct alleged.” Ashcroft v. Iqbal, 556 U.S. 662, 678 (2009) (citation omitted). There must

3 be “more than a sheer possibility that a defendant has acted unlawfully.” Id. While courts do not

4 require “heightened fact pleading of specifics,” a plaintiff must allege facts sufficient to “raise a

5 right to relief above the speculative level.” Twombly, 550 U.S. at 555, 570.

6 In deciding whether the plaintiff has stated a claim upon which relief can be granted, the

7 Court accepts the plaintiff’s allegations as true and draws all reasonable inferences in favor of the

8 plaintiff. See Usher v. City of Los Angeles, 828 F.2d 556, 561 (9th Cir. 1987). However,

9 the court is not required to accept as true “allegations that are merely conclusory, unwarranted

10 deductions of fact, or unreasonable inferences.” In re Gilead Scis. Sec. Litig., 536 F.3d 1049,

11 1055 (9th Cir. 2008). If the court dismisses the complaint, it “should grant leave to amend even if

12 no request to amend the pleading was made, unless it determines that the pleading could not
Northern District of California
United States District Court

13 possibly be cured by the allegation of other facts.” Lopez v. Smith, 203 F.3d 1122, 1127 (9th Cir.

14 2000). In making this determination, the court should consider factors such as “the presence or

15 absence of undue delay, bad faith, dilatory motive, repeated failure to cure deficiencies by

16 previous amendments, undue prejudice to the opposing party and futility of the proposed

17 amendment.” Moore v. Kayport Package Express, 885 F.2d 531, 538 (9th Cir. 1989).

18 DISCUSSION

19 I. ADDITION OF NEW PLAINTIFFS & CLAIMS

20 Stability objects to the addition of seven new plaintiffs to the FAC, as well as the addition

21 of two new claims against it (induced infringement, Count 2) and unjust enrichment (Count 4).

22 Stability MTD, Dkt. No. 162 at 13-14. DeviantArt similarly objects to the addition of new

23 plaintiffs and the addition of the unjust enrichment claim. DeviantArt MTD, Dkt. No. 163, at 23-

24 24. Defendants argue that the addition of the new plaintiffs and claims exceeds the scope of the

25 leave to amend I allowed plaintiffs in the October 2023 Order and, therefore, the new plaintiffs

26 and claims should be stricken.

27 Plaintiffs respond that they were given broad “leave to amend and attempt to cure the

28 deficiencies identified” in their claims and the ability of the then three named plaintiffs to pursue
6
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 7 of 33

1 the claims asserted. Because leave was not cabined to “solely” or exclusively correct the

2 deficiencies, plaintiffs argue that the additions were permissible and in any event should be

3 allowed under Rule 15’s liberal amendment standard. See Oppo. to Stability, Dkt. No. 174, at 22;

4 Oppo. to DeviantArt, Dkt. No. 177, at 21-22.

5 Plaintiffs are correct that leave to amend under Rule 15 is “freely given,” especially at the

6 start of a case. However, once a complaint is dismissed and a court grants plaintiffs specific leave

7 to amend to address identified legal or factual deficiencies, adding plaintiffs or claims beyond

8 those previously alleged requires either requesting and securing leave, which plaintiffs did not do,

9 or the consent of defendants, which plaintiffs did not seek. See Fed. R. Civ. P, 15(a)(2) (“In all

10 other cases, a party may amend its pleading only with the opposing party’s written consent or the

11 court’s leave.”).5

12 That said, I would have granted leave to amend if plaintiffs had sought leave, given the
Northern District of California
United States District Court

13 lack of prejudice to defendants. At this juncture, where each defendant has addressed the newly

14 added claims and plaintiffs, it would elevate form over substance to grant the motions to dismiss

15 on this ground. And I will assume that plaintiffs have implicitly sought leave to amend to include

16 the new plaintiffs and claims in a Second Amended Complaint. I will grant leave and address the

17 arguments defendants make against the added claims and plaintiffs.6

18 II. STABILITY AI MOTION TO DISMISS

19 A. Induced Copyright Infringement
20 Stability challenges plaintiffs’ first theory of induced copyright infringement interpreting

21 the claim as one alleging that the Stable Diffusion models themselves are infringing works. Under

22 this theory, Stability is inducing infringement by distributing the models when any third-party

23
5
24 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
25 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.
26 19-CV-06361-RS, 2021 WL 4503137, at *2 (N.D. Cal. Oct. 1, 2021).
6
27 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
28 defendant based on theories (if any) that are not preempted by the Copyright Act.

7
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 8 of 33

1 downloads, uses, or deploys the models provided by Stability. See FAC, Count Two (alleging

2 contributory infringement by “distributing Stable Diffusion 2.0 and Stable Diffusion XL 1.0 for

3 free” on “behalf of the LAION-5B Registered Plaintiffs and Damages Subclass”) ¶¶ 233-236.7

4 Stability argues, first, that this theory is simply a repackaged direct infringement theory,

5 that by “distributing” Stable Diffusion, Stability violates plaintiffs’ exclusive rights of distribution

6 of their works.8 But whether this is a direct infringement claim (where liability is imposed against

7 Stability for distributing copyrighted works) or more properly characterized as an inducement

8 claim (where liability is imposed because Stability induces or otherwise causes others to copy

9 protectible material) depends on how Stable Diffusion works and is implemented by users other

10 than Stability itself. Any potential overlap – or potential requirement for plaintiffs to elect one

11 claim or another – is better addressed on summary judgment, after discovery.

12 Stability also argues that the inducement claim must be dismissed to the extent plaintiffs
Northern District of California
United States District Court

13 are alleging that Stability encourages the use of Stable Diffusion to create infringing outputs. That

14 theory is barred, according to Stability, because plaintiffs fail to allege any facts that Stability

15 promoted the use of Stable Diffusion to “infringe copyright, as shown by clear expression or other

16 affirmative steps taken to foster infringement.” Metro-Goldwyn-Mayer Studios Inc. v. Grokster,

17 Ltd., 545 U.S. 913, 919 (2005). Stability argues that given the lack of allegations supporting a

18 specific intent by Stability to promote infringement, the claim fails.

19 However, plaintiffs point to one statement by Stability’s CEO that Stability took 100,000

20 gigabytes of images and compressed it to a two-gigabyte file that can “recreate” any of those

21 images.9 Stability responds that the “isolated” use of the word “create” by its CEO cannot

22

23
7
Stability does not challenge plaintiffs’ claim of direct copyright infringement, as that survived
the prior motion to dismiss. October 2023 Order at 7.
24 8
Stability also contends this theory is another take on the “derivative” works claim I dismissed in
25 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
26 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
27 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.
28 9
See FAC ¶ 4 (“Emad Mostaque described it thus: ‘Stable Diffusion is the model itself. It’s a
8
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 9 of 33

1 demonstrate that it intended to foster infringement. Stability argues that “clear allegations of

2 active steps to encourage direct infringement” are especially important here, where plaintiffs do

3 not dispute that Stable Diffusion is capable of substantial noninfringing uses, like creating art from

4 inputs that do not reference particular artists or invoke particular artists’ styles or have any

5 substantial similarity to plaintiffs’ works.

6 The theory of this case is not similar to – for example – a case asserting contributory

7 infringement based on the sale of VCRs where, after discovery, plaintiff had no evidence of

8 defendant’s intent to induce infringement. The Supreme Court explained that, in those

9 circumstances, intent could not be “based on presuming or imputing intent to cause infringement

10 solely from the design or distribution of a product capable of substantial lawful use, which the

11 distributor knows is in fact used for infringement.” See Grokster, Ltd., 545 U.S. at 933

12 (discussing holding of Sony Corp. of America v. Universal City Studios, Inc., 464 U.S. 417
Northern District of California
United States District Court

13 (1984)). Instead, this is a case where plaintiffs allege that Stable Diffusion is built to a significant

14 extent on copyrighted works and that the way the product operates necessarily invokes copies or

15 protected elements of those works. The plausible inferences at this juncture are that Stable

16 Diffusion by operation by end users creates copyright infringement and was created to facilitate

17 that infringement by design. In addition to the comment of Stability’s CEO, plaintiffs reference

18 articles by academics and others that training images can sometimes be reproduced as outputs

19 from the AI products. FAC ¶¶ 90, 130-139.

20 Whether true and whether the result of a glitch (as Stability contends) or by design

21 (plaintiffs’ contention) will be tested at a later date. The allegations of induced infringement are

22 sufficient.

23 B. DMCA
24 Stability also moves again to dismiss plaintiffs’ DMCA claims asserted under 17 U.S.C.

25 section 1202(a) for providing or distributing false copyright management information (“CMI”)

26
27
collaboration that we did with a whole bunch of people … We took 100,000 gigabytes of images
28 and compressed it to a two-gigabyte file that can recreate any of those [images] and iterations of
those.’”).
9
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 10 of 33

1 and under section 1202(b)(1) for intentional removal of CMI.10 I dismissed plaintiffs’ DMCA

2 claim with leave to amend in the October 2023 Order, explaining:

3 [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
4 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
5 development of generative AI models to capture not only images but
any accompanying text because that accompanying text is necessary
6 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
7 text CMI present in the images the named plaintiffs included with
their online images that they contend was stripped or altered in
8 violation of the DMCA during the training of Stable Diffusion or the
use of the end-products. Plaintiffs must, on amendment, identify the
9 particular types of their CMI from their works that they believe were
removed or altered.
10
In addition, plaintiffs must clarify and then allege plausible facts
11 regarding which defendants they contend did the stripping or altering
in violation of the DMCA and when that occurred.
12
Northern District of California

October 2023 Order at 18.

United States District Court

13
1. Claim Under 1202(a)
14
Stability moves to dismiss the subsection (a) claim regarding false CMI. Plaintiffs allege
15
that “Stability distributes the Stability Models under the MIT License (see, e.g. —
16
https://github.com/Stability-AI/stablediffusion/blob/main/LICENSE). Within this license, Stability
17
asserts copyright in the Stability Models. By asserting copyright in the Stability Models, which
18
infringe the copyrights of the LAION-5B Plaintiffs, Stability is providing and distributing false
19
CMI in violation of 17 U.S.C. § 1202(a).” FAC ¶ 248.
20
Stability argues that this claim fails because Stability’s generic license does not suggest
21
any association at all with plaintiffs’ works, and therefore was not made “in connection with”
22
plaintiffs’ works which is necessary to support a claim under 1202(a). See Logan v. Meta
23
Platforms, Inc., 636 F. Supp. 3d 1052, 1062 (N.D. Cal. 2022) (plaintiff failed to allege facts
24
showing defendant “conveyed CMI in connection” with plaintiff’s photos where “Meta’s allegedly
25

26
10
27 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
28 alteration of CMI and distribution of works with removed or altered CMI. See 17 U.S.C. § 1202
et seq.
10
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 11 of 33

1 false CMI is ‘[a] [generic] copyright tag on the bottom of each Facebook user page,’ separated

2 from the rest of the content on the webpage, and is not located on or next to” plaintiff’s photos);

3 but see Post Univ. v. Course Hero, Inc., No. 3:21-CV-1242 (JBA), 2023 WL 5507845, at *4 (D.

4 Conn. Aug. 25, 2023) (“Thus, at the motion to dismiss stage, unless it is ‘implausible’ that a

5 viewer could understand the information to be referring to the defendant as the work’s copyright

6 holder, dismissal is inappropriate.”). In addition, Stability contends that plaintiffs fail to allege

7 facts plausibly satisfying the “double scienter” required under (a); that Stability knowingly

8 provided false CMI with the intent to induce or enable infringement.

9 I agree on both counts. The generic license that accompanies use of Stable Diffusion on its

10 face claims rights to Stable Diffusion as a work, not to the LAION dataset and not any works that

11 were used to create the LAION dataset. It is implausible that a viewer reading the license

12 disclosure for Stable Diffusion would understand that Stability is claiming rights to or conveying
Northern District of California
United States District Court

13 any false information regarding the rights of the plaintiffs whose copyrighted works are among the

14 billions of images in the LAION datasets.

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

16 a basis for this claim.

17 2. Claim Under 1202(b)(1)

18 Similarly, Stability contends that plaintiffs have failed to plead plausible facts that Stability

19 AI intentionally removed or altered CMI from plaintiffs’ works during the training process for

20 Stable Diffusion, and failed to allege the double-scienter requirement that Stability did so in order

21 to facilitate infringement. In the FAC, plaintiffs allege that “Stability directly copied the LAION-

22 5B Works and used these Statutory Copies as training data for the Stability Models. The works

23 copied by Stability included CMI, including in the form of distinctive marks such as watermarks

24 or signatures, and as the captions in the image-text pairs. The training process is designed to

25 remove or alter CMI from the training images. Therefore, Stability intentionally removed or

26 altered CMI from the Plaintiffs’ works in violation of 17 U.S.C. § 1202(b)(1).” FAC ¶ 245.

27 Plaintiffs support that allegation by pointing to some of plaintiffs’ images that were used as

28 Training Images in LAION-5B and that contained CMI, comparing them to images that were
11
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 12 of 33

1 created, for example, when plaintiffs’ names were used as Midjourney image prompts. FAC ¶¶

2 189-200. Plaintiffs have also, as directed in the October 2023 Order, identified the CMI present

3 on their works that they contend has been stripped by Stability. Id. ¶¶ 239-241. Finally, plaintiffs

4 allege that Stability engaged in knowing removal of CMI, as supported by their allegations

5 regarding how the diffusion process works, how training images are used, and based on plausible

6 allegations regarding Stability AI’s prominent role in the funding of LAION. Id. ¶¶ 245-247.

7 Stability, however, raises a new argument on this round of motions, based on a recent

8 opinion from the Hon. Jon S. Tigar. Doe 1 v. GitHub, Inc., No. 22-CV-06823-JST, 2024 WL

9 235217, at *8 (N.D. Cal. Jan. 22, 2024). Stability contends that because the output images are

10 admittedly not identical to the Training Images, there can be no liability for any removal of CMI

11 that occurred during the training process. That is because failing to affix CMI to a “different

12 work” is not “removal” under Section 1202. Judge Tigar wrote:

Northern District of California
United States District Court

13 Defendants now ask the Court to address an unresolved argument

from the prior briefing—namely, that “[Section] 1202(b) claims lie
14 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
15 No. 109-3 at 23–24. Defendants argue that because Plaintiffs' new
allegations state that output from Copilot is often a modification of
16 their licensed works, as opposed to an “identical copy,” they have
effectively pleaded themselves out of their Section 1202(b)(1) and
17 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
18 this claim anew and that Section 1202(b) claims require that copies
be “identical.”
19
Id. * 8; see also Doe 1 v. GitHub, Inc., No. 22-CV-06823-JST, 2024 WL 1643691, at *2 (N.D.
20
Cal. Apr. 15, 2024) (rejecting motion for reconsideration because there was no allegation that any
21
output was identical to any plaintiff’s work in its entirety, so there could be no “removal” of CMI
22
for purposes of DMCA).11
23
Judge Tigar’s analysis disagrees with a case from the Southern District of Texas, where the
24

25 11
In so holding, Judge Tigar followed Advanta-STAR Auto. Rsch. Corp. of Am. v. Search Optics,
26 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.
27 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 F. App'x 674 (9th Cir. 2017) (no section 1202(b)
28 violation where the allegedly infringing drawing was “not identical”).

12
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 13 of 33

1 court rejected the “identicality” requirement. See ADR Int'l Ltd. v. Inst. for Supply Mgmt. Inc.,

2 667 F. Supp. 3d 411, 427 (S.D. Tex. 2023) (“[b]ased on the plain wording of the statute, the Court

3 is not persuaded that the DMCA includes an ‘identical copy’ requirement,” noting the copying

4 should be “substantially similar” but does not have to be “perfect”).

5 Recognizing that this issue is unsettled, I agree with the reasoning of Judge Tigar that

6 followed other district court decisions within the Ninth Circuit. Because there are no allegations

7 that any output from Stable Diffusion was identical to a plaintiff’s work, the DMCA section

8 1202(b) claim fails as well.12

9 The DMCA claims against Stability are DISMISSED WITH PREJUDICE.13

10 C. Unjust Enrichment
11 In addition to arguing that leave to amend was not granted to assert this claim, addressed

12 above, Stability argues that the unjust enrichment claim cannot proceed because it is preempted by
Northern District of California
United States District Court

13 the Copyright Act. It also contends that plaintiffs fail to allege that they have an inadequate

14 remedy at law as required by Sonner v. Premier Nutrition Corp., 971 F.3d 834 (9th Cir. 2020).

15 To support this claim, plaintiffs allege that Stability AI “has unjustly misappropriated the

16 LAION-5B Works in order to develop, train and promote the Stability Models, enabling it to

17 receive profit and other benefits.” FAC ¶252. “By using Plaintiffs’ works to train, develop and

18 promote the Stability Models, Plaintiffs and the Class were deprived of the benefit of the value of

19 their works, including monetary damages.” Id. ¶ 254. Stability points out that this claim is

20 expressly based on the use of plaintiffs’ copyrighted works without consent, and as a result is

21 covered by and preempted by the Copyright Act.

22

23 12
The only allegations regarding identical outputs in the FAC are those regarding researcher’s
24 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
25 Diffusion can produce a work identical to one of the plaintiffs here, sufficient to show actionable
removal of CMI under section 1202(b).
26 13
Runway and Midjourney also move to dismiss the DMCA claim asserted against them. The
27 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
28 DMCA claims asserted against Runway and Midjourney are likewise DISMISSED WITH
PREJUDICE and will not be addressed further.
13
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 14 of 33

1 “Section 301 of the Act seeks ‘to preempt and abolish any rights under the common law or

2 statutes of a State that are equivalent to copyright and that extend to works,’ so long as the rights

3 fall ‘within the scope of the Federal copyright law.’ [] ‘We have adopted a two-part test,’ in

4 accordance with section 301, ‘to determine whether a state law claim is preempted by the Act.

5 Laws,’ [] First, we decide ‘whether the ‘subject matter’ of the state law claim falls within the

6 subject matter of copyright as described in 17 U.S.C. §§ 102 and 103.’ [] Second, assuming it

7 does, we determine ‘whether the rights asserted under state law are equivalent to the rights

8 contained in 17 U.S.C. § 106, which articulates the exclusive rights of copyright holders.’”

9 Maloney v. T3Media, Inc., 853 F.3d 1004, 1010 (9th Cir. 2017) (quoting Laws v. Sony Music

10 Entm't, Inc., 448 F.3d 1134, 1137-38 (9th Cir. 2006)).

11 Plaintiffs correctly note that to “survive preemption, the state cause of action must protect

12 rights which are qualitatively different from the copyright rights. The state claim must have an
Northern District of California
United States District Court

13 extra element which changes the nature of the action.” Laws, 448 F.3d at 1143. Plaintiffs do not

14 identify an “extra element” required under California’s Unfair Competition Law (“UCL”) or

15 common law (which the unjust enrichment claims are based on) that changes the nature of those

16 state law claims to protect something other than rights protected under Copyright Act. Instead,

17 they assert that the heart of their unjust enrichment claim is Stability being unjustly enriched by its

18 “image product’s ability to mimic Plaintiffs’ artistic style and benefit from their notoriety and

19 reputation as sought-after artists.” Oppo. to Stability AI at 20. More specifically, they allege that

20 the Stable Diffusion models use “CLIP-guided diffusion” that relies on prompts including artists’

21 names to generate an image. Therefore, the “crux” of plaintiffs’ unjust enrichment claim revolved

22 not around plaintiffs’ works but around plaintiffs themselves and their “artistic personas.” Id.

23 The problem with plaintiffs’ theory is that it is not in the FAC. The unjust enrichment

24 claim against Stability (and the other defendants) is tied instead to use of plaintiffs’ works. See

25 FAC ¶ 42 (“Whether the use of Plaintiffs and Class members’ works to train, develop, and

26 promote Defendants AI Image Products constitute an unjust benefit conferred upon Defendants to

27 Plaintiffs’ detriment”); see also id. ¶¶ 252-255 (alleging unjust misappropriation of works).

28 As alleged, the unjust enrichment claim against Stability added to the FAC without leave
14
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 15 of 33

1 of court is preempted by the Copyright Act. It is DISMISSED. If plaintiffs have a good faith

2 theory of unjust enrichment that falls outside the scope of the protections provided by the

3 Copyright Act, they are given leave to make one last attempt to state an unjust enrichment claim.14

4 Stability’s motion to dismiss is DENIED concerning the challenged copyright claims, but

5 GRANTED for the DMCA claims without leave to amend and GRANTED for the unjust

6 enrichment claim with leave to amend.

7 III. RUNWAY AI MOTION TO DISMISS

8 Newly added defendant Runway AI is primarily alleged to have trained or assisted in the

9 training at least Stable Diffusion 1.5, using Training Images from the LAION dataset. See, e.g.,

10 FAC ¶¶ 4, 163, 176, 342-345. In addition to challenging plaintiffs’ DMCA and unjust

11 enrichment/UCL claims – which fail for the same reasons identified above with respect to

12 Stability – Runway argues that the infringement claims fail given the allegations alleged with
Northern District of California
United States District Court

13 respect to its particular conduct.

14 A. Request for Judicial Notice

15 As an initial matter, Runway asks me to take judicial notice of a motion to dismiss filed in

16 another case in this District, as well as the complaint from the case. Runway RJN (Dkt. No. 164),

17 Exs. A&B. Plaintiffs oppose, arguing that judicial notice of court records from other cases is

18 inappropriate for the purpose Runway seeks; to encourage me to follow the “approach” of those

19 other courts. “On a Rule 12(b)(6) motion to dismiss, when a court takes judicial notice of another

20 court’s opinion, it may do so ‘not for the truth of the facts recited therein, but for the existence of

21 the opinion, which is not subject to reasonable dispute over its authenticity.’” Southern Cross

22 Overseas Agencies, Inc. v. Wah Kwong Shipping Group Ltd., 181 F.3d 410, 426–27 (3rd

23

24 14
Because of this conclusion, I need not reach Stability’s argument under Sonner v. Premier
25 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.
26 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
27 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
28 not be addressed further. The unjust enrichment claim asserted against DeviantArt is addressed
below.
15
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 16 of 33

1 Cir.1999).” Lee v. City of Los Angeles, 250 F.3d 668, 690 (9th Cir. 2001). Notifying me of the

2 existence of opinions from the judges in other cases may be accomplished by a far simpler

3 method, citing them as persuasive authority. Judicial notice is not an appropriate method to

4 suggest that I should follow the analysis of other courts addressing different cases with different

5 facts.

6 Runway also asks me to take judicial notice – under the doctrine of incorporation – of the

7 full contents of three academic articles cited in plaintiffs’ FAC. Id., Exs. C, D & E. Plaintiffs

8 object to this request, noting that one article is only mentioned once in the FAC and the others a

9 few more times as support for the plausibility of plaintiffs’ assertion that Stable Diffusion contains

10 “compressed copies” of the Training Images based on how these models generally work.

11 Plaintiffs argue that it is inappropriate to take “notice” of the debated truth, meaning or

12 implications of the articles to foreclose their claims. I agree. I will not take judicial notice of the
Northern District of California
United States District Court

13 full contents of the academic articles to resolve disputes of fact, or the legal implications from

14 undisputed facts, at this juncture.

15 Runway’s request for judicial notice is DENIED.

16 B. Direct Copyright Infringement

17 Runway does not move to dismiss first direct copyright infringement claim asserted

18 against it. See Count 11, FAC ¶¶ 347-349. That claim is based on Runway’s alleged use of

19 Training Images to train Stable Diffusion 1.5, a claim that survived Stability AI’s prior motion to

20 dismiss. October 2023 Order at 7 (discussing the “Training Theory”). Runway does challenge

21 plaintiffs’ two other direct infringement theories. First, the “Model Theory” is based on the theory

22 that the Stable Diffusion1.5 product itself – after it was trained – is “an infringing Statutory Copy”

23 of plaintiffs’ works or a “Statutory Derivative Work” because it represents a transformation of

24 plaintiffs’ works. See FAC ¶¶ 209, 350. Second, Runway challenges plaintiffs’ “Distribution

25 Theory” of infringement, based on allegations that Runway infringes plaintiffs’ exclusive

26 distribution rights because distributing Stable Diffusion 1.5 is equivalent to distributing plaintiffs’

27 works. FAC ¶ 352.

28 As with Stability, because Runway does not challenge the use of the images for training
16
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 17 of 33

1 purposes, I need not address the other theories of direct infringement. However, I note that both

2 the model theory and the distribution theory of direct infringement depend on whether plaintiffs’

3 protected works are contained, in some manner, in Stable Diffusion as distributed and operated.

4 That these works may be contained in Stable Diffusion as algorithmic or mathematical

5 representations – and are therefore fixed in a different medium than they may have originally been

6 produced in – is not an impediment to the claim at this juncture. 1 Nimmer on Copyright §

7 2.09[D][1] (2024) (“A work is no less a motion picture (or other audiovisual work) whether the

8 images are embodied in a videotape, videodisc, or any other tangible form.”).

9 Plaintiffs addressed the deficiencies in their prior complaint, alleging additional facts in the

10 FAC concerning how the training images remain in and are used by Stable Diffusion. See, e.g.,

11 FAC ¶¶ 71, 83, 88-90, 150.15 Plaintiffs rely, with respect to Runway as with Stability, on

12 comments from Stability’s CEO regarding the contents of the model and their ability to reproduce
Northern District of California
United States District Court

13 works, as well as academic papers indicating that the Stable Diffusion models are capable of

14 producing very similar if not identical works to at least some training images. FAC ¶¶ 122-150.

15 And plaintiffs also rely on use of their names as prompts in the Runway products to create outputs

16 mimicking aspects of plaintiffs’ protected works as evidence that their protected works are being

17 copied or distributed in Runway’s product. Id. ¶¶ 163-169.

18 Runway disputes the accuracy of those assertions – including the full meaning and import

19 of the academic articles relied on by plaintiffs – and argues that the prompt-examples do not

20 support assertions of direct copyright infringement absent express identification of outputs that are

21 “substantially similar” to plaintiffs’ copyrighted works. But the allegations at this juncture are

22 sufficient to allow the direct infringement claims to proceed. Whether evidence can support each

23

24
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
25 inference of copying can be supported – see, e.g., Hanagami v. Epic Games, Inc., 85 F.4th 931,
935 (9th Cir. 2023) – or liability imposed – see, e.g., Narell v. Freeman, 872 F.2d 907, 910 (9th
26 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
27 “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
28 Cir. 2015) (affirming summary judgment that Google’s “program does not allow access in any
substantial way to a book’s expressive content”).
17
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 18 of 33

1 of the theories and whether plaintiffs will need to choose between theories (e.g., between direct

2 infringement based on selling a product containing effective copies of copyrighted works or

3 violating plaintiffs’ rights to restrict distribution of their works) will be addressed at summary

4 judgment.16

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

6 C. Induced Infringement
7 Runway argues that plaintiffs have also failed to allege two elements of the induced

8 infringement claim: acts of infringement by third parties using Runway’s products and that

9 Runway promoted use of Stable Diffusion to infringe. Runway claims that the only support for

10 the induced infringement claim are comments plaintiffs identify by Stability executives, not

11 statements by anyone associated with Runway.

12 Plaintiffs allege that Runway helped train and develop Stable Diffusion, and therefore,
Northern District of California
United States District Court

13 knew that the product allegedly uses or invokes the training images in its operation. Those

14 allegations, combined with allegations that Runway actively induces others to download Stable

15 Diffusion by distributing Stable Diffusion through popular coding websites and also by making

16

17 16
Runway, like the other defendants, picks up on my questions from the prior oral argument to
18 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
19 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
20 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
21 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
22 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
23 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
24 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”
25 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
26 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.
27 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
28 operate and, presumably, whether and what the products can produce substantially similar outputs
as a result of “overtraining” on specific images or by design.
18
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 19 of 33

1 selling its products (including AI Magic Tools) that include Stable Diffusion, are sufficient. FAC

2 ¶¶ 83, 163, 185, 352, 358.17

3 Runway’s motion to dismiss is DENIED on the infringement claims, GRANTED with

4 prejudice on the DMCA claims, and GRANTED with leave to amend on the unjust

5 enrichment/UCL claim.

6 IV. MIDJOURNEY MOTION TO DISMISS

7 In addition to moving to dismiss the DMCA and unjust enrichment claims addressed

8 above, Midjourney moves to dismiss the copyright claims and Lanham Act false endorsement and

9 trade dress claims18 asserted against it.

10 A. Copyright
11 1. Registration
12 Midjourney argues, with respect to three of the named plaintiffs – Anderson, Kaye, and
Northern District of California
United States District Court

13 Brom – that the evidence of their registration of newly identified copyrighted works is

14 insufficient. With respect to Anderson and Kaye, Midjourney asserts that a subset of each artists’

15 works identified as being both copyrighted and included in the LAION datasets used to train the

16 AI products are compilations. It contends that copyright protection only extends to the new

17 material in compilations and Anderson and Kaye fail to identify which works within the subset of

18 compilations are the new material. It also argues that for two of three of the works identified by

19 Brom, the copyright registrations extend only to text and not artwork. Midjourney Mot. at 6-8.

20 It is undisputed that plaintiffs who do not have valid copyright protections will not be able

21 to pursue copyright claims based on un-registered works or works whose registrations covered

22 only text. It is also undisputed that each of the named plaintiffs who claim their copyright-

23 protected works were included in the LAION datasets have at least one work whose registration is

24 facially valid. At this juncture, therefore, the Copyright Act claims survive against Midjourney

25

26 17
Runway’s argument that plaintiffs’ induced infringement claim must fail because plaintiffs do
27 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.
28 18
15 U.S.C. § 1125(a) et seq.
19
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 20 of 33

1 and the other defendants. However, the identification in the FAC and exhibits of unprotected

2 works is not irrelevant; plaintiffs rely on some of those works to plausibly demonstrate that their

3 works were used as training images and that their works or elements of their works can be

4 recreated through the AI products. The identification of those works may not prove liability under

5 the Copyright Act, but they do provide support for the plausibility of plaintiffs’ Copyright Act

6 theories.

7 2. Use as Training Images

8 In my prior Order, I required plaintiffs to clarify the basis of their copyright claims against

9 Midjourney, as plaintiffs’ allegations regarding whether and how Midjourney used plaintiffs’

10 images in the training of its product were insufficient. October 2023 Order at 13-14. In their

11 FAC, plaintiffs now allege that Midjourney separately trained its product on the LAION400M and

12 LAION5B datasets. See FAC ¶¶ 266, 274. Plaintiffs also allege that Midjourney incorporates
Northern District of California
United States District Court

13 Stable Diffusion into its own AI product. Id. ¶¶ 169-170.

14 Midjourney contests the adequacy of these new allegations with respect to the training of

15 images, arguing plaintiffs must identify specific, individual registered works that each artist

16 contends Midjourney actually used for training. Given the unique facts of this case – including the

17 size of the LAION datasets and the nature of defendants’ products, including the added allegations

18 disputing the transparency of the “open source” software at the heart of Stable Diffusion – that

19 level of detail is not required for plaintiffs to state their claims. Instead, plaintiffs have added to

20 their FAC more detailed allegations regarding the training and use of the LAOIN datasets by

21 defendants generally and Midjourney specifically. Plaintiffs have plausible allegations showing

22 why they believe their works were included in the LAION datasets. And plaintiffs plausibly

23 allege that the Midjourney product produces images – when their own names are used as prompts

24 – that are similar to plaintiffs’ artistic works. See FAC Exs. F&G.

25 Midjourney nonetheless argues that these examples are insufficient because some of the

26 identified works are not registered and the resulting outputs could just as likely be the result of

27 training on unregistered works or utilizing only unprotected elements from plaintiffs’ works. But

28 plaintiffs’ reliance on those exhibits are not to establish copyright infringement as a matter of law.
20
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 21 of 33

1 Instead, they are relied on to support the plausibility of plaintiffs’ copyright theories (that all or

2 most of the works in the LAION datasets were used by Midjourney and the other defendants to

3 train their AI products, and that plaintiffs’ works or their protected elements that are contained in

4 the AI products as the works or protected elements can be recreated by using the AI products).

5 The FAC allegations and the exhibits help plaintiffs cross the plausibility threshold. Whether

6 plaintiffs will be able to prove their claims is a different matter and those claims will be tested on

7 an evidentiary basis at summary judgment.19

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

9 B. Lanham Act
10 In place of the right of publicity claims asserted against Midjourney that were dismissed in

11 my October 2023 Order, five plaintiffs (Anderson, Brom, Kaye, Ortiz and Rutkowski) now assert

12 Lanham Act claims based on theories of false endorsement and trade dress. These plaintiffs allege
Northern District of California
United States District Court

13 that their names appeared on the list of 4700 artists posted by Midjourney’s CEO on Discord, the

14 platform where Midjourney’s AI product operates. Midjourney’s CEO promoted the list as

15 describing the various styles of artistic works its AI product could produce. FAC ¶¶ 261-262,

16 305.20 Plaintiffs also allege that Midjourney has itself published user-created images that

17 incorporate the plaintiff artists’ names in Midjourney’s “showcase” site. Id. ¶ 325(b) & Ex. K.

18 1. False Endorsement
19 “To prevail on its Lanham Act trademark claim, a plaintiff must prove: (1) that it has a

20 protectible ownership interest in the mark; and (2) that the defendant’s use of the mark is likely to

21
19
22 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
23 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
24 summary judgment.

25
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
26 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
27 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
28 Midjourney Named Plaintiffs included in the Midjourney Name List sponsored or approved of
Midjourney’s image generator.”).
21
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 22 of 33

1 cause consumer confusion.” Rearden LLC v. Rearden Commerce, Inc., 683 F.3d 1190, 1202–203

2 (9th Cir. 2012) (quoting Network Automation, Inc. v. Advanced Sys. Concepts, Inc., 638 F.3d

3 1137, 1144 (9th Cir. 2011)). A false designation of origin claim likewise requires a showing of a

4 likelihood of consumer confusion. New W. Corp. v. NYM Co. of Cal., Inc., 595 F.2d 1194, 1201

5 (9th Cir. 1979) (“Whether we call the violation infringement, unfair competition or false

6 designation of origin, the test is identical[:] is there a ‘likelihood of confusion?’”).

7 Midjourney contends that plaintiffs fail to allege the first, necessary element of their false

8 endorsement claim: falsity. They note that plaintiffs do not allege that the Midjourney AI product

9 cannot recognize works by those included on the Midjourney Names List, and argue that just

10 because the Names List exists and was promoted by the Midjourney CEO, that by itself cannot

11 support an inference of endorsement. Midjourney points to a part of the Discord thread (identified

12 by plaintiffs as the source of the Midjourney CEO’s identification of the Names List) not relied on
Northern District of California
United States District Court

13 by plaintiffs, where the CEO indicated that the names on the list came from “wikipedia and magic

14 the gathering.” Midjourney MTD, Dkt. No. 169, at 18. Midjourney argues that this part of the

15 Discord thread is judicially noticeable and dispels any inference that the artists on the lists could

16 plausibly be considered to have endorsed the Midjourney product.

17 As discussed above, judicial notice of other comments in the thread is not appropriate to

18 dispute the facts plaintiffs otherwise plausibly assert, especially as the plaintiffs dispute the

19 accuracy and inferences to be drawn from the totality of the messages in that thread. And even if

20 the Names List itself was insufficient to support an inference of false endorsement, plaintiffs also

21 allege that their names were used in connection with works included in Midjourney’s “showcase.”

22 FAC ¶ 325b. Whether or not a reasonably prudent consumer would be confused or misled by the

23 Names List and showcase to conclude that the included artists were endorsing the Midjourney

24 product can be tested at summary judgment. See Adobe Sys. Inc. v. A & S Elecs., Inc., 153 F.

25 Supp. 3d 1136, 1142 (N.D. Cal. 2015).21

26
21
27 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
28 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
22
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 23 of 33

1 Midjourney also argues that to get past the First Amendment protection provided to

2 expressive works, plaintiffs must, but have not, alleged that Midjourney’s use of plaintiffs’ names

3 to invoke their styles has “‘no artistic relevance to the underlying work whatsoever, or, if it has

4 some artistic relevance, unless the [use of trademark or other identifying material] explicitly

5 misleads as to the source or the content of the work.’” Brown v. Elec. Arts, Inc., 724 F.3d 1235,

6 1239 (9th Cir. 2013) (quoting Rogers v. Grimaldi, 875 F.2d 994 (2d Cir.1989)). However, the

7 plaintiffs have plausibly alleged that the use of their names by Midjourney in the List and

8 showcase misleads consumers regarding source and endorsement. Open questions also remain

9 whether Midjourney promoting its product for commercial gain for use by others to create artistic

10 images is itself expressive use that creates “artistic relevance” to plaintiffs’ underlying works.

11 Discovery may show that it is or that is it not. Unlike in Brown v. Elec. Arts, Inc., where plaintiff

12 alleged only that his likeness was used in a video game, we do not yet have the sort of record, or
Northern District of California
United States District Court

13 sort of alleged use, that could support dismissal of the claim at the motion to dismiss stage.

14 2. Vicarious Trade Dress

15 The same five plaintiffs (Anderson, Brom, Kaye, Ortiz and Rutkowski) also allege a trade

16 dress claim based on the use of their names in connection with the Midjourney AI product’s use of

17 a “CLIP-guided model” that has been trained on the work of the Midjourney Named Plaintiffs and

18 allows users to create works capturing the “trade dress of each of the Midjourney Named Plaintiffs

19 [that] is inherently distinctive in look and feel as used in connection with their artwork and art

20 products.” FAC ¶¶ 321. The CLIP model, plaintiffs assert, works as a trade dress database that

21 can recall and recreate the elements of each artist’s trade dress. FAC ¶¶ 83, 320. Plaintiffs point

22 to examples showing how Midjourney recreates works with their trade dress in Ex. F to the FAC.

23 Midjourney argues that plaintiffs have failed to state this claim because they have not

24 adequately identified the “concrete elements” of each plaintiff’s trade dress. See, e.g., YZ Prods.,

25

26
the court recognized was promoting competition. Id. at 565-66. That case was also admittedly not
27 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
28 operation of its product was misleading or is truthful comparative advertising can be tested at
summary judgment.
23
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 24 of 33

1 Inc. v. Redbubble, Inc., 545 F. Supp. 3d 756, 767 (N.D. Cal. 2021) (“images alone are insufficient

2 to provide adequate notice dismissing trade dress claim,” and requiring plaintiff to identify “the

3 concrete elements” of their protected trade dress). It acknowledges that plaintiffs have identified

4 aspects of their trade dress, see FAC ¶ 319,22 but asserts that those descriptions are impermissibly

5 broad. See, e.g., Walker & Zanger, Inc. v. Paragon Indus., Inc., 549 F. Supp. 2d 1168, 1176 (N.D.

6 Cal. 2007 (granting summary judgment on the “some of the elements of plaintiff’s proposed trade

7 dress [that] are overbroad”). Plaintiffs refute that characterization, arguing instead that they have

8 identified the “set of recurring visual elements and artistic techniques, the particular combination

9 of which are distinctive to each of the Midjourney Named Plaintiffs.” Id. ¶ 319.

10 While the images from Exhibit F on their own would be insufficient identification, and

11 while some of the alleged “concrete elements” identified in the FAC are, standing alone, vague

12 and possibly overbroad, those elements cannot be considered alone but as a whole in the context
Northern District of California
United States District Court

13 of plaintiffs’ other, plausible allegations. Arcsoft, Inc. v. Cyberlink Corp., 153 F. Supp. 3d 1057,

14 1068 (N.D. Cal. 2015). Here, the combination of identified elements and images, when

15 considered with plaintiffs’ allegations regarding how the CLIP model works as a trade dress

16 database, and Midjourney’s use of plaintiffs’ names in its Midjourney Name List and showcase,

17 provide sufficient description and plausibility for plaintiffs’ trade dress claim.

18 Midjourney also argues that artistic elements or styles identified for each artist that are

19 allegedly re-creatable by using its product are functional, and therefore not protected.23 It ignores,

20

21
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
22 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,
23 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,
24 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.
25 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.”).
26 23
See Arcsoft, 153 F. Supp. 3d at 1068 (“To determine whether a claimed trade dress is functional,
27 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
28 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
24
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 25 of 33

1 however, the Ninth Circuit’s test for determining non-functionality and instead relies on a series of

2 inapposite cases addressing jewelry, wooden cutouts, and keychains. See, e.g., Int'l Ord. of Job's

3 Daughters v. Lindeburg & Co., 633 F.2d 912, 918 (9th Cir. 1980) (use of name and emblem of

4 fraternal organization were functional aesthetic components of jewelry); Crafty Prods., Inc. v.

5 Michaels Companies, Inc., 424 F. Supp. 3d 983, 993 (S.D. Cal. 2019), aff'd sub nom. Crafty

6 Prods., Inc. v. Fuqing Sanxing Crafts Co., 839 F. App'x 95 (9th Cir. 2020) (plaintiffs failed to

7 identify “concrete elements” of their trade dress and instead attempted to capture the “entire

8 design” of hundreds of different wooden cardboard pieces); Au-Tomotive Gold, Inc. v. Volkswagen

9 of Am., Inc., 457 F.3d 1062, 1064 (9th Cir. 2006) (reversing summary judgment protecting

10 defendants’ use of trademarked symbols in keychains and license plate holders, as “[t]he doctrine

11 of aesthetic functionality does not provide a defense against actions to enforce the trademarks

12 against such poaching” or source identification).

Northern District of California
United States District Court

13 Midjourney complains that plaintiffs intend to assert a monopoly over some of the

14 elements plaintiffs identify as critical parts of their protected trade dress. But that ignores the

15 plausible allegations that the CLIP model functions as a trade dress database and the use of its

16 product to produce works based on the names of these plaintiffs. Plaintiffs have no protection

17 over “simple, cartoony drawings” or “gritty fantasy paintings.” Midjourney Reply, Dkt. No. 184

18 at 13. But their trade dress claims must be considered viewing all identified elements as well as

19 the nature of the use of the CLIP model and their names. Those issues will be tested on an

20 evidentiary basis. This claim will not be dismissed based on “aesthetic functionality.”

21 Midjourney contends that the trade dress claim must nonetheless be dismissed because

22 plaintiffs allege their trade dress is “either” distinctive or, in the alternative, has acquired

23 secondary meaning.24 However, plaintiffs pleaded secondary meaning. See FAC ¶ 328 (“Each of

24

25 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
26 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
27 Cir.2001)).
24
28 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
25
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 26 of 33

1 the Midjourney Named Plaintiffs’ trade dress possesses secondary meaning because the trade

2 dress of their art products invoke a mental association by a substantial segment of potential

3 consumers between the trade dress and the creator of the art product.”). Proof of intentional

4 copying by Midjourney has been alleged, especially considering the allegations regarding the

5 CLIP model functioning as a trade dress database, the express use of plaintiffs’ names in the

6 Midjourney Names List to promote the product and the use of some plaintiffs’ names and

7 likenesses of their works in the Midjourney showcase, and the “mental recognition” of

8 Midjourney’s calling out of plaintiffs by name.

9 Finally, Midjourney contests the sufficiency of plaintiffs’ vicarious trade dress claim,

10 arguing that plaintiffs have failed to allege facts of “joint ownership or control” over the infringing

11 product by Midjourney and any end user. Plaintiffs allege:

12 Midjourney exercises control over the infringing images by including

Northern District of California
United States District Court

the CLIP model in its image pipeline, and by marketing artist-name

13 prompts as a key feature of its image generator via the Midjourney
Name List. Without the CLIP model, Midjourney’s users would not
14 be able to infringe on the Midjourney Named Plaintiffs’ trade-dress
rights or those of the other artists on the Midjourney Name List.
15
FAC ¶ 326. These allegations support the claim for vicarious trade dress infringement. See, e.g.,
16
Y.Y.G.M. SA v. Redbubble, Inc., No. 219CV04618RGKJPR, 2020 WL 3984528, at *9 (C.D. Cal.
17
July 10, 2020 (“To impose vicarious liability the shared control must extend in some way to the
18
infringed intellectual property itself.”).
19
Midjourney’s motion to dismiss the Lanham Act claims is DENIED.25
20

21
not limited to) direct consumer testimony; survey evidence; exclusivity, manner, and length of use
22 of a mark; amount and manner of advertising; amount of sales and number of customers;
established place in the market; and proof of intentional copying by the defendant.’ Filipino
23 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
24 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
25 omitted).” 581 F.3d at 1145.

26
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
27 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
28 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
26
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 27 of 33

C. Request for Judicial Notice

1
Midjourney requests judicial notice of: (1) messages from a Discord thread, where
2
Midjourney CEO David Holz released “our style list” and posted a link to a spreadsheet on
3
Google Docs called “Midjourney Style List” that contained the named of 4700 artists, under the
4
doctrine of incorporation by reference, Dkt. No. 170-1, Ex. 1; (2) messages in a different Discord
5
thread that plaintiffs rely on in part, where Holz discusses Midjourney’s image-prompt tool, Dkt.
6
No. 170-1, Ex. 2; (3) another Discord thread discussing the image-prompt tool; and (4) a transcript
7
from the hearing before Judge Chhabria in the Kadrey case.
8
As noted before, judicial notice of documents from the Kadrey case is not necessary, as
9
either side can point to the orders from that case to argue their persuasiveness or differences with
10
respect to the sufficiency of the allegations in this case. See supra at 15-16. Judicial notice is also
11
not appropriate for any of the three Discord threads to dispute the facts plausibly alleged. For
12
Northern District of California

example, Midjourney does not dispute that the list of artists was disclosed and promoted by its
United States District Court

13
CEO, which is the central reason for which plaintiffs rely on that Discord thread and its
14
attachment. Midjourney seeks to rely on the thread to dispute the meaning of the comments by
15
Holz and point to messages that are not relied on by plaintiffs (and whose meanings are disputed
16
by plaintiffs) to foreclose plaintiffs’ claims. That is not appropriate.
17
Midjourney’s request for judicial notice is DENIED.
18
V. DEVIANTART MOTION TO DISMISS
19
A. Copyright Claim
20
DeviantArt moves again to dismiss the copyright claims asserted against it by attempting
21
to set itself apart from the other defendants. It points out that it has not been alleged to have
22
trained any AI model, but simply to have implemented and used the AI tools provided by Stability
23
and others. It argues that holding it liable for that conduct would make the millions of third parties
24

25

26 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
27 “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,
28 combined with Midjourney’s use of plaintiffs’ names and showcase examples calling out named
plaintiffs. This argument is better determined on a full record.
27
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 28 of 33

1 who have downloaded and implemented the AI products challenged in this case – many of which

2 are open source software – liable for infringement, which is unsupportable.

3 In the October 2023 Order, I dismissed the copyright infringement claim asserted against

4 DeviantArt, explaining:

5 In addition to providing clarity regarding their definition of and theory

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

when users access DreamUp through the app or through DeviantArt’s

United States District Court

13 website.
14 October 2023 Order at 10.

15 DeviantArt argues, first, that plaintiffs’ copyright act claims are barred by the October

16 2023 Order because plaintiffs still do not allege that DeviantArt itself copied or used their works

17 to train Stable Diffusion or trained any other program. The FAC alleges only that DeviantArt

18 incorporates and relies on Stable Diffusion for its DreamUp product. FAC ¶¶ 6, 387-392, 395-397.

19 Plaintiffs have added allegations to their FAC, however, regarding how copies or protected

20 elements of their works remain, in some format, in Stable Diffusion and how those works can be

21 invoked by use of all of the Stable Diffusion versions. The actual operation of Stable Diffusion

22 1.4 and whether the amount of any plaintiff’s copyrighted works in that program suffices for

23 copyright infringement or a fair use defense concerning DeviantArt remains to be tested at

24 summary judgment.

25 Moreover, while plaintiffs admittedly did not include assertions or examples in the FAC of

26 outputs from DreamUp that appear to copy elements of their works (as they did for the other

27 defendants), their added allegations regarding the use of the LAION datasets to train Stable

28 Diffusion and how Stable Diffusion versions operates, including specific examples and academic
28
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 29 of 33

1 references to the operation of Stable Diffusion 1.4 used by DreamUp, suffice. See FAC ¶¶ 132-

2 137, 388-397. For example, in paragraph 393 plaintiffs assert:

3 On information and belief, by the end of training, Stable Diffusion 1.4

was capable of reproducing protected expression from each of the
4 LAION-5B Registered Works that was in each case substantially
similar to that registered work, because— a. In the Carlini Paper,
5 Nicholas Carlini tested Stable Diffusion 1.4 and found that it could
emit stored copies of its training images; b. The training procedure
6 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
7 H: Runway image prompts to be capable of emitting stored copies of
protected expression.
8
See also ¶ 394 (“Therefore, like Stable Diffusion 1.5, Stable Diffusion 1.4 also qualifies as an
9
infringing Statutory Copy of the LAION-5B Registered Works. Because Stable Diffusion 1.4
10
represents a transformation of the LAION-5B Registered Works into an alternative form, Stable
11
Diffusion 1.4 also qualifies as an infringing Statutory Derivative Work.”).
12
Northern District of California

DeviantArt asks me to review to the full content of one of the academic articles plaintiffs
United States District Court

13
rely on, the Carlini Study, in particular the article’s conclusion that for the 350,000 training
14
images studied, only 109 output images were “near-copies” of the training images. See DeviantArt
15
Reply at 4-5. DeviantArt argues that it is simply not plausible that LAION 1.4 can reproduce any
16
of plaintiffs’ copyrighted works given that the set of training images Carlini selected were the
17
“most-duplicated” examples, numbering in the millions of impressions in the datasets. Id.
18
Plaintiffs’ reference to Carlini, however, is only one part of the allegations that help make
19
plaintiffs’ allegations plausible regarding how these products operate and how “copies” of the
20
plaintiffs’ registered works are captured in some form in the products. As noted above, it is not
21
appropriate to rely on the defendants’ assertions regarding the method or sample size used by
22
Carlini and their conclusions from it to foreclose plaintiffs’ claims; plaintiffs aggressively dispute
23
the implications from the sample size and results of that study. Finally, while the differences
24
between Stable Diffusion 1.4 and the other Stable Diffusion versions might be legally significant
25
based on what the evidence shows after discovery, at this juncture the allegations about the
26
common training of those versions and how they all operate are sufficient to keep plaintiffs’
27

28
29
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 30 of 33

1 copyright infringement allegations against DeviantArt alive.26

2 Finally, DeviantArt contends that any use of plaintiffs’ works in Stable Diffusion 1.4

3 should be considered fair use as a matter of law, given the huge size of the training datasets and

4 plaintiffs apparent inability to use Stable Diffusion 1.4 to reproduce any works that look similar to

5 their copyrighted works. Whether DreamUp operates in a way that could draw upon or otherwise

6 reproduce plaintiffs’ works to an extent that violates the Copyright Act and whether a fair use

7 defense applies are issues that must be tested on an evidentiary basis. See, e.g., Authors Guild v.

8 Google, Inc., 804 F.3d 202, 226 (2d Cir. 2015) (applying fair use defense upon summary judgment

9 record).

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

11 B. Breach of Contract
12 DeviantArt moves again to dismiss the breach of contract claim that plaintiffs assert based
Northern District of California
United States District Court

13 on the same provision of DeviantArt’s Terms of Service (“ToS”) that I considered in the October

14 2023 Order. In that Order I explained:

15 In opposition, plaintiffs focus on § 16 of the TOS, a provision not

identified or quoted in their Complaint:
16
16. Copyright in Your Content
17 DeviantArt does not claim ownership rights in Your Content.
For the sole purpose of enabling us to make your Content
18 available through the Service, you grant to DeviantArt a non-
exclusive, royalty-free license to reproduce, distribute, re-
19 format, store, prepare derivative works based on, and publicly
display and perform Your Content. Please note that when you
20 upload Content, third parties will be able to copy, distribute
and display your Content using readily available tools on their
21 computers for this purpose although other than by linking to
your Content on DeviantArt any use by a third party of your
22 Content could violate paragraph 4 of these Terms and
Conditions [preserving copyright rights in the original owner
23

24 26
Given this conclusion, I need not separately reach the question of whether plaintiffs can assert a
25 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
26 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,
27 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).
28 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.
30
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 31 of 33

of the copyright and disclaiming any ownership interest of

1 DeviantArt in the posted work] unless the third party receives
permission from you by license.
2
Plaintiffs argue that DeviantArt breached this provision when “it
3 incorporated Stable Diffusion into its own AI Image product knowing
that Stability had scraped DeviantArt’s artists’ work.” Oppo to
4 DeviantArt MTD [Dkt. No. 65] at 22. However, section 16 provides
a limited license to DeviantArt and warns that third parties may be
5 able to copy and violate content-owners’ rights. It does not clearly
cover the conduct that plaintiffs accuse DeviantArt of in this suit;
6 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
7 alleged supporting an allegation that DeviantArt itself exceeded the
scope of the limited license.
8
DeviantArt also challenges the ability of plaintiff McKernan and the
9 unspecified “others” to sue DeviantArt for breach claims based on
contractual provisions prohibiting other users (presumably here,
10 Stability) from using DeviantArt content for commercial uses. The
Complaint does not allege and is devoid of facts supporting the
11 inference that Stability is bound by the TOS or that plaintiff
McKernan or others are third party beneficiaries of specific
12 provisions in the TOS who may sue to enforce terms of agreements
Northern District of California
United States District Court

entered between DeviantArt and Stability.

13
The breach of contract claim is DISMISSED with leave to amend. If
14 plaintiffs attempt to amend this claim, they must identify the exact
provisions in the TOS they contend DeviantArt breached and facts in
15 support of breach of each identified provision. To the extent plaintiffs
rely on provisions that appear to protect or benefit DeviantArt but not
16 the users, or contracts.
17 In the FAC, plaintiffs reallege a violation of Section 16, based on the same theory as

18 above, and DeviantArt argues it fails for the same reason. See FAC ¶¶ 420-422. I agree. For the

19 reasons discussed in the October 2023 Order, plaintiffs cannot state a breach of contract claim as a

20 matter of law based on allegations that DeviantArt knew Stable Diffusion was trained on LAION

21 datasets that had been scraped in part from DeviantArt’s website. As before, there are no

22 allegations that DeviantArt did anything to permit the scraping of images from its site. Plaintiffs

23 admit that DeviantArt played no role in the scraping or training. Nothing in the TOS precludes

24 DeviantArt from using Stable Diffusion, even if it had knowledge that some of the images used in

25 its training were scraped from its own site and its members’ works.

26 Plaintiffs argue that DeviantArt breached this provision because it is now using – through

27 its use of Stable Diffusion – its members’ works for purposes beyond “the sole purpose of

28 enabling us to make your Content available through the Service.” But nothing in Section 16 limits
31
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 32 of 33

1 DeviantArt’s ability to use plaintiffs’ works that are available from another source, i.e., the

2 LAION datasets as incorporated into Stable Diffusion. The breach claim fails, again, as a matter

3 of law.

4 Plaintiffs added a breach of the implied covenant claim in the FAC, premised also on a

5 breach of Section 16 of the ToS. See ¶ 422b; Dkt. No. 177 at 16.27 “[T]o state a claim for breach

6 of the implied covenant of good faith and fair dealing, a plaintiff must identify the specific

7 contractual provision that was frustrated.” Rockridge Tr. v. Wells Fargo, N.A., 985 F. Supp. 2d

8 1110, 1156 (N.D. Cal. 2013). DeviantArt argues that the implied covenant claim fails because

9 plaintiffs do not identify any contractual provision that obligated DeviantArt to protect its users

10 from competition from DeviantArt or anyone else, or to offer protection from data scraping, or

11 that would otherwise preclude DeviantArt from using a tool developed by third parties even if that

12 tool was based in part on works scraped from its site without its involvement. I agree. There is no
Northern District of California
United States District Court

13 basis to find any provision of the ToS was frustrated by DeviantArt’s alleged conduct.

14 The breach of contract claim is DISMISSED. As plaintiffs did not contest this claim at oral

15 argument, despite my tentative ruling order identifying my intent to dismiss this claim (Dkt. No.

16 193), and did not suggest any facts they could allege to salvage their breach claim in an amended

17 complaint, this claim is DISMISSED with prejudice.

18 C. Unjust Enrichment
19 Consistent with the analysis above, the newly added unjust enrichment claim against

20 DeviantArt based on use of plaintiffs’ “works to develop and promote DreamUp and the

21 DreamUp–CompVis Model” that deprived plaintiffs “the benefit of the value of their works” FAC

22 ¶ 433, is dismissed as preempted under the Copyright Act. In their opposition to DeviantArt’s

23 motion, plaintiffs imply that the unjust enrichment claim against DeviantArt could be based upon

24

25
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
26 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
27 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
28 Scraping & Machine Learning Activities’ provision to its Terms of Service after DeviantArt’s
members complained about the unfairness of DreamUp.”
32
 Case 3:23-cv-00201-WHO Document 223 Filed 08/12/24 Page 33 of 33

1 a different ground; misrepresentations made by DeviantArt to plaintiffs. Plaintiffs’ Oppo. to

2 DeviantArt at 19. Plaintiffs are given one last attempt to amend their unjust enrichment claims

3 against each defendant. If the theory underlying plaintiffs’ unjust enrichment claim against

4 DeviantArt rests on different facts and theories from the unjust enrichment claim asserted against

5 the other defendants, plaintiffs should make that clear.

6 CONCLUSION

7 Defendants’ motions to dismiss the DMCA claims are GRANTED and the DMCA claims

8 are DISMISSED WITH PREJUDICE. Defendants’ motions to dismiss the unjust enrichment

9 claims are GRANTED and those claims are DISMISSED with leave to amend. Defendants’

10 motions to dismiss the Copyright Act claims are DENIED. Midjourney’s motion to dismiss the

11 Lanham Act claims is DENIED. DeviantArt’s motion to dismiss the breach of contract and breach

12 of the implied covenant of good faith and fair dealing claims is GRANTED and those claims are
Northern District of California
United States District Court

13 DISMISSED WITH PREJUDICE.

14 Dated: August 12, 2024

15

16
William H. Orrick
17 United States District Judge
18
19

20

21

22

23

24

25

26
27

28
33