Andersen v. Stability AI

Case 3:23-cv-00201 Document 1 Filed 01/13/23 Page 1 of 46
1 Joseph R. Saveri (State Bar No. 130064)

Cadio Zirpoli (State Bar No. 179108)
2 Christopher K.L. Young (State Bar No. 318371)
Elissa A. Buchanan (State Bar No. 249996)
3 Travis Manfredi (State Bar No. 281779)
JOSEPH SAVERI LAW FIRM, LLP
4 601 California Street, Suite 1000
San Francisco, California 94108
5 Telephone: (415) 500-6800
Facsimile: (415) 395-9940
6 Email: [email protected]
[email protected]
7 [email protected]
[email protected]
8 [email protected]
9 Matthew Butterick (State Bar No. 250953)
1920 Hillhurst Avenue, #406
10 Los Angeles, CA 90027
Telephone: (323) 968-2632
11 Facsimile: (415) 395-9940
Email: [email protected]
12
Counsel for Individual and Representative
13 Plaintiffs Sarah Andersen, Kelly McKernan,
Karla Ortiz, and the Proposed Class
14
[Additional Counsel Listed on Signature Page]
15
16 UNITED STATES DISTRICT COURT

NORTHERN DISTRICT OF CALIFORNIA
17 SAN FRANCISCO DIVISION
18 SARAH ANDERSEN, an individual; Case No.
KELLY MCKERNAN, an individual;
19 KARLA ORTIZ, an individual, COMPLAINT
20 Individual and Representative Plaintiffs, CLASS ACTION
21 v.
22 STABILITY AI LTD., a UK corporation; DEMAND FOR JURY TRIAL
STABILITY AI, INC., a Delaware
23 corporation; MIDJOURNEY, INC., a
Delaware corporation; DEVIANTART, INC.,
24 a Delaware corporation,
25 Defendants.
26
27
28
Class Action Complaint

TABLE OF CONTENTS
1
2 I. AI IMAGE GENERATORS ARE 21ST-CENTURY COLLAGE TOOLS THAT

VIOLATE THE RIGHTS OF MILLIONS OF ARTISTS .................................... 1
3
II. JURISDICTION AND VENUE ........................................................................... 2
4
III. INTRADISTRICT ASSIGNMENT .................................................................... 3
5
IV. DEFINITIONS ..................................................................................................... 3
6
V. PARTIES ............................................................................................................... 6
7
VI. AGENTS AND CO-CONSPIRATORS ............................................................... 8
8
VII. CLASS ALLEGATIONS ...................................................................................... 9
9
A. Class Definitions........................................................................................ 9
10
B. Numerosity................................................................................................ 9
11
C. Typicality..................................................................................................10
12
D. Commonality & Predominance .................................................................10
13
E. Adequacy .................................................................................................. 11
14
F. Other Class Considerations ...................................................................... 12
15
VIII. FACTUAL ALLEGATIONS ............................................................................... 12
16
A. Stability AI................................................................................................ 12
17
B. Midjourney ............................................................................................... 14
18
C. DeviantArt................................................................................................ 14
19
D. How Stable Diffusion Works: A 21st-Century Collage Tool ..................... 14
20
E. The source of the Stable Diffusion training data: LAION ........................ 22
21
F. DeviantArt’s betrayal of its artist community by embracing generative AI
22 images...................................................................................................... 24
23 G. Midjourney: the 21st-century collage tool (in)famous for its artistic style 27
24 IX. CLAIMS FOR RELIEF....................................................................................... 30
25 XI. JURY TRIAL DEMANDED .............................................................................. 43
26
27
28
i Class Action Complaint

1 Plaintiffs Sarah Andersen, Kelly McKernan, and Karla Ortiz (“Plaintiffs”), on behalf of
2 themselves and all others similarly situated, bring this Class Action Complaint (the
3 “Complaint”) against Defendants Stability AI Ltd. and Stability AI, Inc. (collectively
4 “Stability”); Midjourney, Inc. (“Midjourney”); and DeviantArt, Inc. (“DeviantArt”) (all
5 collectively “Defendants”) for direct and vicarious copyright infringement under 17 U.S.C.
6 § 501; violation of the Digital Millennium Copyright Act, 17 U.S.C. §§ 1201–1205 (the
7 “DMCA”); violation of Plaintiffs’ statutory and common law rights of publicity, Cal. Civ. Code
8 section 3344; violation of Unfair Competition law, Cal. Bus. & Prof. Code §§ 17200, et seq.; and
9 declaratory relief.
10 I. AI IMAGE GENERATORS ARE 21ST-CENTURY COLLAGE TOOLS
THAT VIOLATE THE RIGHTS OF MILLIONS OF ARTISTS
11
12 1. Stable Diffusion is a software product—defined below as an AI Image Product—

13 maintained and sold by Stability.
14 2. Stability downloaded or otherwise acquired copies of billions of copyrighted
15 images without permission to create Stable Diffusion, including Plaintiffs’. These images are
16 defined below as “Training Images.”
17 3. By training Stable Diffusion on the Training Images, Stability caused those images
18 to be stored at and incorporated into Stable Diffusion as compressed copies. Stability made them
19 without the consent of the artists and without compensating any of those artists.
20 4. When used to produce images from prompts by its users, Stable Diffusion uses the
21 Training Images to produce seemingly new images through a mathematical software process.
22 These “new” images are based entirely on the Training Images and are derivative works of the
23 particular images Stable Diffusion draws from when assembling a given output. Ultimately, it is
24 merely a complex collage tool.
25 5. These resulting derived images compete in the marketplace with the original
26 images. Until now, when a purchaser seeks a new image “in the style” of a given artist, they must
27 pay to commission or license an original image from that artist. Now, those purchasers can use
28 the artist’s works contained in Stable Diffusion along with the artist’s name to generate new
1 Class Action Complaint

1 works in the artist’s style without compensating the artist at all. As used herein, the phrase “in
2 the style of,” refers to a work that others would accept as a work created by that artist whose
3 “style” was called upon, not the general category of work, such as fantasy or impressionism. Only
4 a very small number of incredibly talented artists are capable of this same feat for a single other
5 artist (i.e., reproducing art that is convincingly in that artist’s style), let alone for countless other
6 artists. AI Image Products do so with ease by violating the rights of millions of artists.
7 6. All AI Image Products operate in substantially the same way and store and
8 incorporate countless copyrighted images as Training Images.
9 7. Defendants, by and through the use of their AI Image Products, benefit
10 commercially and profit richly from the use of copyrighted images.
11 8. The harm to artists is not hypothetical—works generated by AI Image Products
12 “in the style” of a particular artist are already sold on the internet, siphoning commissions from
13 the artists themselves.
14 9. Plaintiffs and the Class seek to end this blatant and enormous infringement of their
15 rights before their professions are eliminated by a computer program powered entirely by their
16 hard work.
17 II. JURISDICTION AND VENUE
18 10. Plaintiffs bring this action on their own behalf and as representatives of a Class of
19 similarly situated individuals and entities. They seek to obtain injunctive relief and recover
20 damages as a result and consequence of Defendants’ unlawful conduct.
21 11. Jurisdiction and venue are proper in this judicial district pursuant to Defendants’
22 unauthorized use of Plaintiffs’ and the Class’s intellectual property in violation of the Copyright
23 Act, 17 U.S.C. § 501, the Digital Millennium Copyright Act, 17 U.S.C. §§ 1201–1205 (the
24 “DMCA”); Unjust Enrichment, and Unfair Competition; California’s right of publicity,
25 contract, negligence, privacy, and unfair competition statutes and case law.
26 12. A substantial part of the events giving rise to Plaintiffs’ claims occurred in this
27 District.
28

1 13. Defendant Midjourney is headquartered in San Francisco. Plaintiff Karla Ortiz

2 resides in San Francisco, California, a substantial portion of the affected interstate trade and
3 commerce was carried out in this District, and the Defendants are licensed to do business in this
4 District.
5 14. Each Defendant has transacted business, maintained substantial contacts, and/or
6 committed illegal acts that harmed Plaintiffs and the Class throughout the United States,
7 including in this District. Defendants’ conduct has had the intended and foreseeable effect of
8 causing injury to persons residing in, located in, or doing business throughout the United States,
9 including in this District.
10 III. INTRADISTRICT ASSIGNMENT
11 15. Pursuant to Civil Local Rule 3.2 (c) and (e), assignment of this case to the San
12 Francisco Division of the United States District Court for the Northern District of California is
13 proper because Plaintiff Karla Ortiz and a large portion of the Class reside in this District. In
14 addition, a substantial amount of the development of the Stable Diffusion product as well as of
15 the interstate trade and commerce involved and affected by Defendants’ conduct giving rise to
16 the claims herein occurred in this Division.
17 IV. DEFINITIONS
18 16. “AI Image Product” refers to the allegedly AI-based image generation products that
19 were created, maintained, marketed, sold, and/or distributed by Defendants, namely Stable
20 Diffusion, the Midjourney Product, DreamStudio, and DreamUp.
21 17. “Artificial Intelligence” or “AI.” AI is a software program that algorithmically
22 simulates human reasoning or inference, often using statistical or mathematical methods.
23 18. “Derivative Work” as used herein refers to the output of AI Image Products as
24 well as the AI Image Products themselves—which contain compressed copies of the copyrighted
25 works they were trained on.
26 19. “Diffusion” is a specific machine-learning application that results in a model that
27 stores a compressed copy of each item in the training dataset. A more detailed description of
28

1 diffusion appears below. Stable Diffusion is an AI software program that includes a diffusion
2 model.
3 20. “Generative AI” is a subset of machine learning where the program copies training
4 data and uses it to produce derivative works of that training data. Stable Diffusion is an example
5 of a generative AI system, because it is trained using copied images, and produces similar images.
6 Other generative AI systems exist that produce conversational text, software code, and music, in
7 each case similar to the respective training data.
8 21. “Machine Learning” or “ML” is a type of AI process in which the behavior of the
9 software program is derived from copying a corpus of material called training data. In this
10 context, the term “learning” is metaphorical. The process bears very little similarity to human
11 learning. In this context, it denotes a technique for developing a software program through
12 massive data input and statistical operations, calculations, and linear algebra, rather than line-by-
13 line coding using a programming language. Machine-learning programs can find patterns or make
14 calculations based on datasets or training data. The operator of the algorithm is sometimes called
15 a “trainer.” Machine learning will be referred to as “AI” unless further distinction is necessary.
16 Stable Diffusion, the Midjourney Product, DreamStudio, and DreamUp all include both AI
17 programs and ML programs.
18 22. “Model” denotes the software component of an AI program that is the output of a
19 machine-learning process. The model is a computer file that contains all the information that has
20 been extracted from analyzing the training data, including the rules and data structures produced
21 by the algorithm.
22 23. A “Software Library” is a self-contained software program that is designed to
23 provide certain functions or services to another software program, such as a desktop or
24 smartphone application, thereby reducing the development time. When used as a Software
25 Library, Stable Diffusion provides image-generating services to the other program. Stable
26 Diffusion has been used as a Software Library within multiple programs, including DreamStudio,
27 DreamUp, and, on information and belief, the Midjourney Product.
28

1 24. A “Text Prompt” is a textual description that is used as an interface to certain

2 generative AI systems to produce output, including the AI Image Products. For instance, a user
3 of Stable Diffusion can input the text prompt “a dog wearing a baseball cap while eating ice
4 cream,” and Stable Diffusion will attempt to generate an image that corresponds to that text
5 prompt. Because randomness is injected into the generative process alongside the text prompt, a
6 particular text prompt will usually produce a varied set of results. In a generative AI system like
7 Stable Diffusion, a text prompt is not part of the training data. It is part of the end-user interface
8 for the tool. Thus, it is more akin to a text query passed to an internet search engine. Just as the
9 internet search engine looks up the query in its massive database of web pages to show us
10 matching results, a generative AI system uses a text prompt to generate output based on its
11 massive database of training data.
12 25. A “Training Image” is an image, or image paired with a descriptive text caption,
13 that is included among the training data for a machine-learning process. Training images are often
14 gathered through web scraping. For its training data, Stable Diffusion has taken billions of
15 Training Images scraped from public websites.
16 26. “Web Scraping” refers to the harvesting, copying, or extracting data from websites
17 by using automated tools, including bots or web crawlers. Usually, the scope and quantity of data
18 so “scraped” is massive. Web scraping can be used to harvest any kind of data available on public
19 websites, especially copyrighted data such as text, images, or software code. These collections of
20 scraped copyrighted data are used as input for other computer programs, such as search engines
21 and machine-learning processes. The training data for all AI Image Products are collected via web
22 scraping. For example, the training data for Stable Diffusion—a database of billions of captioned
23 images—was collected via web scraping.
24 27. “Work” or “Works” refers to any image that was used to train any version of
25 Stable Diffusion that was offered directly and/or incorporated into another product by one or
26 more Defendants during the Class Period.
27
28

1 V. PARTIES
2 A. Plaintiffs
3 28. Plaintiff Sarah Andersen is a resident of the State of Oregon. Ms. Andersen is a
4 full-time cartoonist and illustrator and relies on the income therefrom. Her semi-autobiographical
5 comic strip, Sarah’s Scribbles, finds the humor in living as an introvert. Her graphic
6 novel FANGS was nominated for an Eisner Award. Ms. Andersen has created and owns a
7 copyright interest in over two hundred Works included in the Training Data.1 Ms. Andersen has
8 complied with the statutory requirements for registration and has applied for and owns copyright
9 registrations for sixteen collections that include Works used as Training Images. Copies of these
10 registrations as reflected in the Copyright Office’s records are attached as Exhibits 1 through 16
11 and are valid and enforceable. Ms. Andersen was, and continues to be, injured during the Class
12 Period as a result of Defendants’ unlawful conduct alleged herein.
13 29. Plaintiff Kelly McKernan is a resident of the State of Tennessee. Mx. McKernan is
14 a full-time artist and relies on their income therefrom. Kelly creates original watercolor and acryla
15 gouache paintings for galleries, private commissions, and their online store. Mx. McKernan has
16 created and owns a copyright interest in over thirty Works used as Training Images.2 Mx.
17 McKernan was, and continues to be, injured during the Class Period as a result of Defendants’
18 unlawful conduct alleged herein.
19 30. Karla Ortiz is a resident of the State of California. Ms. Ortiz is a Puerto Rican,
20 internationally recognized, award winning full-time artist and relies on the income therefrom. Ms.
21 Ortiz is renowned for her exceptional design sense, realistic renders, and character-driven
22 narratives, and has contributed to many big-budget projects in the film, television, and video-
23 game industries. Ms. Ortiz is also a regular illustrator for major publishing and role playing game
24 companies. Lastly, Ms. Ortiz is a recognized fine artist, and her deeply personal fine art has been
25 1
Examples of Ms. Andersen’s Works included in the Training Data can be found here:
26 https://haveibeentrained.com/?search_text=sarah%20andersen.
2
27 Examples of Mx. McKernan’s Works included in the Training Data can be found here:
https://haveibeentrained.com/?search_text=kelly%20mckernan and https://laion-
28 aesthetic.datasette.io/laion-aesthetic-6pls/images?_search=kelly+mckernan&_sort=rowid.

1 showcased in notable galleries such as Spoke Art and Hashimoto Contemporary in San Francisco;
2 Nucleus Gallery, Thinkspace, and Maxwell Alexander Gallery in Los Angeles; and Galerie
3 Arludik in Paris. Ms. Ortiz has created and owns a copyright interest in at least twelve Works that
4 were used as Training Images.3 Ms. Ortiz was, and continues to be, injured during the Class
5 Period as a result of Defendants’ unlawful conduct alleged herein.
6 B. Defendants
7 31. Defendant Stability AI Ltd. is a UK corporation with its principal place of business
8 located at 88 Notting Hill Gate, London, England, W11 3HP. Stability AI Ltd. is a party to the
9 unlawful conduct alleged herein.
10 32. Defendant Stability AI, Inc. is a Delaware corporation with its principal place of
11 business located at 88 Notting Hill Gate, London, England, W11 3HP. Stability AI, Inc. is a party
12 to the unlawful conduct alleged herein. Stability AI, Inc. conducts business in this judicial district.
13 On information and belief, Defendant Stability AI, Inc. is a wholly owned subsidiary of Defendant
14 Stability AI Ltd.
15 33. Stability AI Ltd. and Stability AI, Inc. jointly created, trained, and maintain Stable
16 Diffusion, an AI Image Product. Stable Diffusion is used to derive the output images of Stability’s
17 DreamStudio product. DreamStudio is a web-based app that outputs images in response to text
18 prompts. DreamStudio requires Stable Diffusion to function; the images are produced by Stable
19 Diffusion. DreamStudio provides a user interface and access to a trained version of Stable
20 Diffusion. As noted above, Defendant Stability AI, Inc. is referred collectively with Defendant
21 Stability AI Ltd. as “Stability.”
22 34. Defendant Midjourney, Inc.4 is a Delaware corporation with its principal place of
23 business located at 333 Harrison Street, Apt. 605, San Francisco, California 94105. Midjourney
24 created, sells, markets, and distributes the Midjourney Product, which is an AI Image Product.
25 3
Examples of Ms. Ortiz’s Works included in the Training Data can be found here: https://laion-
26 aesthetic.datasette.io/laion-aesthetic-6pls/images?_search=karla+ortiz.
4
27 To avoid confusion between Midjourney’s eponymous product and the entity itself,
Midjourney, Inc. is referred to herein as “Midjourney,” and the image-generating product the
28 company offers is referred to as the “Midjourney Product.”

1 Like Stable Diffusion, the Midjourney Product is a commercial product that produces images in
2 response to text prompts. On information and belief, Stable Diffusion was used in iterations of the
3 Midjourney Product. On information and belief, the version of the Midjourney Product currently
4 available was trained on a subset of the images used to train Stable Diffusion. Midjourney is a
5 party to the unlawful conduct alleged herein.
6 35. Defendant DeviantArt, Inc. (“DeviantArt”) is a Delaware corporation with its
7 principal place of business located at 100 Gansevoort Street, New York, New York 10014.
8 DeviantArt created, sells, markets, and distributes DreamUp, which is an AI Image Product. Like
9 Stable Diffusion and the Midjourney Product, DreamUp is a commercial product that relies on
10 Stable Diffusion to produce images. DreamUp is sold directly on the internet as well as other
11 sales channels throughout the United States, including in this District. DeviantArt released
12 DreamUp on November 9, 2022. DeviantArt is a party to the unlawful conduct alleged herein.
13 VI. AGENTS AND CO-CONSPIRATORS
14 36. The unlawful acts alleged against the Defendants in this class action complaint
15 were authorized, ordered, or performed by the Defendants’ respective officers, agents,
16 employees, representatives, or shareholders while actively engaged in the management, direction,
17 or control of the Defendants’ businesses or affairs.
18 37. The Defendants’ agents operated under the explicit and apparent authority of
19 their principals.
20 38. Each Defendant, and its subsidiaries, affiliates and agents operated as a single
21 unified entity.
22 39. Various persons and/or firms not named as Defendants herein may have
23 participated as co-conspirators in the violations alleged herein and may have performed acts and
24 made statements in furtherance thereof.
25 40. Each acted as the principal, agent, or joint venture of, or for other Defendants with
26 respect to the acts, violations, and common course of conduct alleged herein.
27
28

1 VII. CLASS ALLEGATIONS

2 A. Class Definitions
3 41. Plaintiffs bring this action for damages and injunctive relief on behalf of
4 themselves and all others similarly situated as a class action pursuant to Rules 23(a), 23(b)(2), and
5 23(b)(3) of the Federal Rules of Civil Procedure, on behalf of the following Classes:
6 “Injunctive Relief Class” under Rule 23(b)(2):
7 All persons or entities nationalized and/or domiciled in the United
States that own a copyright interest in any work that was used to
8 train any version of an AI Image Product that was offered directly
and/or incorporated into another product by one or more
9 Defendants during the Class Period.
10 “Damages Class” under Rule 23(b)(3):

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

15 All members of the Damages Class who (1) maintained an account
on DeviantArt; (2) posted copyrighted work on DeviantArt; and
16 (3) had that work used to train any version of an AI Image Product.
17 These “Class Definitions” specifically exclude the following person or entities:

18 a. Any of the Defendants named herein;
19 b. Any of the Defendants’ co-conspirators;
20 c. Any of Defendants’ parent companies, subsidiaries, and affiliates;
21 d. Any of Defendants’ officers, directors, management, employees,
22 subsidiaries, affiliates, or agents;
23 e. All governmental entities; and
24 f. The judges and chambers staff in this case, as well as any members of their
25 immediate families.
26 B. Numerosity
27 42. Plaintiffs do not know the exact number of Class members, because such
28 information is in the exclusive control of Defendants. Plaintiffs are informed and believe that

1 there are at least thousands of Class members geographically dispersed throughout the United
2 States such that joinder of all Class members in the prosecution of this action is impracticable.
3 C. Typicality
4 43. Plaintiffs’ claims are typical of the claims of their fellow Class members because
5 Plaintiffs’ claims arise out of the same course of conduct from which their injuries result.
6 Plaintiffs and all Class own copyrights in the Works. Plaintiffs and the Class created or owned
7 Works that were published on the internet by themselves or others. The Works were used to train
8 various AI Image Products without permission. Plaintiffs and absent Class members were
9 damaged by this and other wrongful conduct of Defendants as alleged herein. Damages and the
10 other relief sought herein are common to all members of the Class.
11 D. Commonality & Predominance
12 44. Numerous questions of law or fact common to the entire Class arise from
13 Defendants’ conduct—including, but not limited to those identified below:
14 i. Direct Copyright Infringement
15  Whether Defendants violated the copyrights of Plaintiffs and the Class
16 when they downloaded and stored copies of the Works.
17  Whether Defendants violated the copyrights of Plaintiffs and the Class
18 when they used copies of the Works to train AI Image Products.
19 ii. Vicarious Copyright Infringement
20  Whether Defendants vicariously violated the copyrights of Plaintiffs and
21 the Class when third parties used Defendants’ products to create Fakes, as
22 defined herein.
23 iii. DMCA Violations
24  Whether Defendants violated the DMCA by removing copyright
25 management information (“CMI”) from the Works and/or causing their
26 respective AI Image Products to omit CMI from their output images.
27
28

1 iv. Right of Publicity Violations

2  Whether Defendants violated Plaintiffs’ and the Class’s rights of publicity
3 when they designed their AI Image Products to respond to prompts
4 requesting output images “in the style” of specific individuals, namely
5 Plaintiffs and the Class.
6 v. Unlawful-Competition
7  Whether Defendants’ AI Image Products are being used by Defendants to
8 engage in Unfair Competition under the Lanham Act and/or California
9 law.
10 vi. Injunctive Relief
11  Whether this Court should enjoin Defendants from engaging in the
12 unlawful conduct alleged herein. And what the scope of that injunction
13 would be.
14 vii. Anticipated Defenses
15  Whether any affirmative defense excuses Defendants’ conduct, including
16 but not limited to whether some or all of Defendants’ conduct is allowed
17 under the Fair Use Doctrine.
18 45. These and other questions of law and fact are common to the Class and
19 predominate over any questions affecting the Class members individually.
20 E. Adequacy
21 46. Plaintiffs will fairly and adequately represent the interests of the Class because
22 they have experienced the same harms as the Class and have no conflicts with any other members
23 of the Class. Furthermore, Plaintiffs have retained sophisticated and competent counsel (“Class
24 Counsel”) who are experienced in prosecuting federal and state class actions throughout the
25 United States and other complex litigation and have extensive experience advising clients and
26 litigating intellectual property, competition, contract, and privacy matters.
27
28

1 F. Other Class Considerations

2 47. Defendants have acted on grounds generally applicable to the Class, thereby
3 making final injunctive relief appropriate with respect to the Class as a whole.
4 48. This class action is superior to alternatives, if any, for the fair and efficient
5 adjudication of this controversy. Prosecuting the claims pleaded herein as a class action will
6 eliminate the possibility of repetitive litigation. There will be no material difficulty in the
7 management of this action as a class action.
8 49. The prosecution of separate actions by individual Class members would create the
9 risk of inconsistent or varying adjudications, establishing incompatible standards of conduct for
10 Defendants.
11 VIII. FACTUAL ALLEGATIONS
12 50. This class action against Defendants concerns a DeviantArt software product
13 called DreamUp, a Midjourney software product, and a Stability software product called
14 DreamStudio, all of which are AI-Image Products and, upon information and belief, built on a
15 Stability Software Library called Stable Diffusion.
16 A. Stability AI
17 51. Stability was founded in London, England in 2020 by Mohammad Emad
18 Mostaque, a former hedge-fund manager. Mostaque is currently the Chief Executive Officer of
19 Stability.
20 52. Stability released Stable Diffusion in August 2022. Stable Diffusion is an AI Image
21 Product that produces images in response to Text Prompts. Stable Diffusion is being updated
22 rapidly, and has had several major releases: version 1.1, version 1.2, version 1.3, version 1.4, and
23 the current version is 2.1. Stability is developing an updated version 3.0.
24 53. Stable Diffusion is software released under a permissive open-source license.5
25 Under this open-source license, programmers and users can download for free the software and
26
27
28 5
See https://github.com/Stability-AI/stablediffusion/blob/main/LICENSE.

1 its associated machine-learning models derived from the Training Images and then use the
2 software according to the terms of the open-source license.
3 54. Stability’s choice to release Stable Diffusion under an open-source license—rather
4 than under a traditional paid license—has led to rapid adoption of Stable Diffusion, with many
5 programmers devising and releasing their own software based on Stable Diffusion.
6 55. In August 2022, the same month that Stable Diffusion was released, Stability
7 released DreamStudio (https://dreamstudio.ai). DreamStudio is a web-server-based AI Image
8 Product through which users can generate images with Text Prompts. DreamStudio relies on
9 Stable Diffusion as its underlying Software Library, meaning that DreamStudio relies on Stable
10 Diffusion to generate images from Text Prompts.
11 56. DreamStudio is billed in packages of “credits,” priced at $1 for 100 credits, with a
12 minimum purchase of 1000 credits for $10. New DreamStudio users receive a certain number of
13 credits for free, after which they must buy more. The credits generally represent computer-
14 processing resources on Stability’s cloud-based servers. For each image generated with
15 DreamStudio, a certain number of credits are redeemed. DreamStudio represents that the
16 number of credits consumed per image depends on user-controlled settings related to the quality
17 and size of the requested image, and the computing resources used. With 1000 credits, Stability
18 estimates that a user can make “approximately 5000 images with default settings.”
19 57. Stability scraped, and thereby copied over five billion images from websites as the
20 Training Images used as training data for Stable Diffusion. Stability did not seek consent from
21 either the creators of the Training Images or the websites that hosted them from which they were
22 scraped.
23 58. Stability did not attempt to negotiate licenses for any of the Training Images.
24 Stability simply took them. Stability has embedded and stored compressed copies of the Training
25 Images within Stable Diffusion.
26 59. Stable Diffusion uses the compressed copies in generating its output in response to
27 Text Prompts. Since launching its DreamStudio app or Stable Diffusion, Stability has not
28

1 attempted to negotiate any licenses for any of the Training Images and is not sharing any of the
2 revenue with the artists who created the Training Images nor any other owners of the Works.
3 60. DreamStudio has been lucrative for Stability. In October 2022, Stability
4 announced it had raised $100 million, led by Coatue and Lightspeed Venture Partners. At the
5 time, Stability was valued at approximately $1 billion.
6 B. Midjourney
7 61. Midjourney was incorporated on September 16, 2020 by David Holz. Midjourney
8 launched an open beta of the first public version of its Midjourney Product on July 12, 2022.
9 Since then, it has released Versions 2 and 3, and, most recently, an alpha iteration of Version 4.
10 C. DeviantArt
11 62. DeviantArt was founded in 2000, DeviantArt has been primarily known as an
12 online community (https://deviantart.com) where digital artists post and share their work,
13 primarily in the form of digital images. Today DeviantArt bills itself as “the world’s largest art
14 community.” DeviantArt hosts millions of such images.
15 63. Thousands—and possibly millions—of the Training Images for Stability’s Stable
16 Diffusion product were scraped and copied from DeviantArt.
17 64. In November 2022, DeviantArt released DreamUp
18 (https://deviantart.com/dreamup). Like Stability’s DreamStudio, DreamUp is a web-based app
19 that generates images in response to Text Prompts. Like DreamStudio, DreamUp relies on
20 Stability’s Stable Diffusion software as its underlying software engine.
21 D. How Stable Diffusion Works: A 21st-Century Collage Tool
22 65. As mentioned above, Stable Diffusion is an AI Image Product released by Stability.
23 It has been incorporated as an image-generating engine into many other software programs,
24 including DreamStudio (by Stability), the Midjourney Product, and DreamUp (released by
25 DeviantArt). Thus, the description that follows of how Stable Diffusion works also describes the
26 operation of DreamStudio, the Midjourney Product, and DreamUp, because they rely on Stable
27 Diffusion as an embedded image-generating engine.
28

1 66. The word “diffusion” in its name refers to the technique the software uses to
2 generate output images that are similar to those found in its training data.
3 67. The diffusion technique was invented in 2015 by a team of researchers led by
4 Jascha Sohl-Dickstein at Stanford University and introduced in their paper “Deep Unsupervised
5 Learning Using Nonequilibrium Thermodynamics” (2015).6 The technique can be applied to any
6 kind of data, but the paper focuses on its application to digital images.
7 68. Diffusion operates in two phases. The first phase of diffusion is to take an image
8 and progressively add more noise to it in a series of steps. In this case, “noise” refers to
9 something seen rather than heard, but the connotation is the same: random fluctuations that we
10 perceive as chaotic and unstructured. At each step, the program records how the addition of noise
11 changes the image. By the last step, the image has been “diffused” into essentially random noise.
12 69. The second phase is like the first but reversed. Having recorded the process of
13 turning a certain image into noise over many steps, the program can then run the sequence
14 backwards. Starting with some random noise, the program applies the steps in reverse order. As it
15 progressively removes noise (or “denoises”) the data, the program is eventually able to
16 reconstruct the original image.
17 70. The program relies on complicated mathematics, linear algebra, and a series of
18 algorithms and requires powerful computers and computer processing to recognize underlying
19 relationships in the data.
20
21
22
23
24
25
26
27
28 6
Available at https://arxiv.org/abs/1503.03585

1 71. The diagram below, taken from the Sohl-Dickstein paper, illustrates the two
2 phases of the diffusion process using a spiral image as the example training data.
3 72.
4
10
11
12 73. The first row of the diagram (with the blue spiral) reads left to right. It depicts the
13 first phase of diffusion, with noise being progressively added to the spiral image (not every step is
14 shown). The middle image shows the spiral halfway through the diffusion process. The rightmost
15 image shows the end of the diffusion process—the spiral has become a field of random noise.
16 74. The second row of the diagram (with the red spiral) reads right to left. It shows
17 the reverse process: a patch of random noise (second row, rightmost image) is progressively un-
18 diffused, or “denoised” by reversing the sequence of steps learned in the first phase. The middle
19 image in the second row shows the denoising process at the halfway point. The leftmost image in
20 the second row shows the end result of the denoising process: the spiral has reappeared.
21 75. Three facts about the diffusion technique are apparent from this diagram.
22 a. Diffusion is a way for a machine-learning model to calculate how to reconstruct a
23 copy of its Training Images. For each Training Image, a diffusion model finds the
24 sequence of denoising steps to reconstruct that specific image. Then it stores this
25 sequence of steps. The diagram above shows a spiral as an example. In practice,
26 this training would be repeated for many images—likely millions or billions. A
27 diffusion model is then able to reconstruct copies of each Training Image.
28 Furthermore, being able to reconstruct copies of the Training Images is not an
1 incidental side effect. The primary goal of a diffusion model is to reconstruct

2 copies of the training data with maximum accuracy and fidelity to the Training
3 Image. It is meant to be a duplicate.
4 b. These reconstructed copies do not perfectly match the originals. For instance, in
5 the diagram, the reconstructed spiral (in red) has some fuzzy parts in the lower
6 half that the original spiral (in blue) does not. Though the red spiral is plainly a
7 copy of the blue spiral, in computer terms it is known as a lossy copy, meaning
8 small, unimportant, or insignificant details are lost as the data is compressed into a
9 smaller size. This is true of many digital data formats, including MP3, AAC, and
10 JPEG, that also make highly compressed copies of digital data by omitting small,
11 unimportant, or insignificant details. This technique is called lossy compression. A
12 diffusion model is a form of lossy compression applied to the Training Images.
13 c. Because a trained diffusion model can produce a copy of any of its Training
14 Images—which could number in the billions—the diffusion model can be
15 considered an alternative way of storing a copy of those images. In essence, it’s
16 similar to having a directory on your computer of billions of JPEG image files. But
17 the diffusion model uses statistical and mathematical methods to store these
18 images in an even more efficient and compressed manner.
19 76. In December 2020, the diffusion technique was improved by a team of researchers
20 at UC Berkeley led by Jonathan Ho. These ideas were introduced in their paper “Denoising
21 Diffusion Probabilistic Models”.7
22 77. Ho’s paper described two improvements to the diffusion technique.
23 78. First, Ho introduced what he called “progressive lossy compression”, a way for a
24 diffusion model to store its training data more efficiently without impacting its ability to
25 reconstruct high-quality copies of the training data. These compressed versions of Training
26
27
28 7
Available from https://arxiv.org/abs/2006.11239

1 Images have come to be known as latent image representations (or just latent images). Ultimately, a
2 latent image is just another copy of an image from the training dataset.
3 79. Second, Ho showed how a latent image could be interpolated—meaning, blended
4 mathematically—to produce new derivative images. Rather than combine two images pixel by
5 pixel—which gives unappealing results—Ho showed how Training Images can be stored in the
6 diffusion model as latent images and then interpolated as a new latent image. This interpolated
7 latent image can then be converted back into a standard pixel-based image.
8 80. The diagram below, taken from Ho’s paper, shows how this process works, and
9 demonstrates the difference in results between interpolating pixels and interpolating latent
10 images.
81.
11
12
13
14
15
16
17
18
19
20
21
22
23 82. In the diagram, two photos are being blended: the photo on the left labeled
24 “Source x0,” and the photo on the right labeled “Source x'0.”
25 83. The image in the red frame has been interpolated pixel by pixel, and is thus labeled
26 “pixel-space interpolation.” This pixel-space interpolation simply looks like two translucent face
27 images stacked on top of each other, not a single convincing face.
28

1 84. The image in the green frame, labeled “denoised interpolation”, has been
2 generated differently. In that case, the two source images have been converted into latent images
3 (illustrated by the crooked black arrows pointing upward toward the label “Diffused source”).
4 Once these latent images have been interpolated (represented by the green dotted line), the newly
5 interpolated latent image (represented by the smaller green dot) has been reconstructed into
6 pixels (a process represented by the crooked green arrow pointing downward to a larger green
7 dot). This process yields the image in the green frame. Compared to the pixel-space interpolation,
8 the difference is apparent: the denoised blended interpolation looks like a single convincing
9 human face, not an overlay or combination of images of two faces.
10 85. A enlarged detail of the two interpolated images is shown below.
11
12
13
14
15
16
17
18
19 86. Despite the difference in results, these two modes of interpolation are equivalent:
20 they both generate derivative works from the source images. In the pixel-space interpolation (the
21 red-framed image), the source images themselves are being directly interpolated to make a
22 derivative image. In the denoised interpolation (the green-framed image), (1) the source images
23 are being converted to latent images, which are lossy-compressed copies; (2) those latent images
24 are being interpolated to make a derivative latent image; and then (3) this derivative latent image
25 is decompressed back into a pixel-based image.
26 87. In April 2022, the diffusion technique was further improved by a team of
27 researchers led by Robin Rombach at Ludwig Maximilian University of Munich. These ideas
28 were introduced in his paper “High-Resolution Image Synthesis with Latent Diffusion Models.”

1 88. Rombach is also employed by Stability as one of the primary developers of Stable
2 Diffusion, which is a software implementation of the ideas in his paper.
3 89. Rombach’s diffusion technique offered one key improvement over previous efforts.
4 Rombach devised a way to supplement the denoising process by using extra information, so that
5 latent images could be interpolated in more complex ways. This process is called conditioning. The
6 most common tool for conditioning is short text descriptions, previously introduced as Text
7 Prompts, that might describe elements of the image, e.g.—“a dog wearing a baseball cap while
8 eating ice cream”. This metric uses Text Prompts as conditioning data to select latent images that
9 are already associated with text captions indicating they contain “dog,” “baseball cap,” and “ice
10 cream.” The text captions are part of the Training Images, and were scraped from the websites
11 where the images themselves were found.
12 90. The resulting image is necessarily a derivative work, because it is generated
13 exclusively from a combination of the conditioning data and the latent images, all of which are
14 copies of copyrighted images. It is, in short, a 21st-century collage tool.
15
16
17
18
19
20
21
22
23
24
25
26
27
28

1 91. The result of this conditioning process may or may not be a satisfying or accurate
2 depiction of the Text Prompt. Below is an example of output images from Stable Diffusion (via
3 the DreamStudio app) using this Text Prompt—“a dog wearing a baseball cap while eating ice
4 cream”. All these dogs in the resulting images seem to be wearing baseball hats. Only the one in
5 the lower left seems to be eating ice cream. The two on the right seem to be eating meat, not ice
6 cream.
7 92.
8
10
11
12
13
14
15
16
17
18
19
20
21 93. In general, none of the Stable Diffusion output images provided in response to a
22 particular Text Prompt is likely to be a close match for any specific image in the training data.
23 This stands to reason: the use of conditioning data to interpolate multiple latent images means
24 that the resulting hybrid image will not look exactly like any of the Training Images that have been
25 copied into those latent images.
26 94. But it is also true that the only thing a latent-diffusion system can do is interpolate
27 latent images into hybrid images. There is no other source of visual information entering the
28 system.

1 95. Every output image from the system is derived exclusively from the latent images,
2 which are copies of copyrighted images. For these reasons, every hybrid image is necessarily a
3 derivative work.
4 96. A latent-diffusion system can never achieve a broader human-like understanding of
5 terms like “dog,” “baseball hat,” or “ice cream.” Hence, the use of the term “artificial
6 intelligence” in this context is inaccurate.
7 97. A latent-diffusion system can only copy from latent images that are tagged with
8 those terms. The system struggles with a Text Prompt like “a dog wearing a baseball cap while
9 eating ice cream” because, though there are many photos of dogs, baseball caps, and ice cream
10 among the Training Images (and the latent images derived from them) there are unlikely to be any
11 Training Images that combine all three.
12 98. A human artist could illustrate this combination of items with ease. But a latent-
13 diffusion system cannot because it can never exceed the limitations of its Training Images.
14 99. In practice, the quality of the latent-diffusion images depends entirely on the
15 breadth and quality of the Training Images used to generate the latent images. If that weren’t
16 true, then it wouldn’t matter where Stable Diffusion (or any other AI-Image Product) got its
17 Training Images.
18 100. In actuality, the provenance of an AI-Image-Product’s Training Images matters
19 very much. According to Emad Mostaque, CEO of Stability, Stable Diffusion has “compress[ed]
20 the knowledge of over 100 terabytes of images.”8 Though the rapid success of Stable Diffusion
21 has been partly reliant on a great leap forward in computer science, it has been even more reliant
22 on a great leap forward in appropriating copyrighted images.
23 E. The source of the Stable Diffusion training data: LAION
24 101. LAION (acronym for “Large-Scale Artificial Intelligence Open Network”) is a
25 nonprofit organization based in Hamburg, Germany. LAION is led by Christoph Schuhmann.
26
8
27 See Kyle Wiggers, This Startup is Setting a DALL-E 2-Like AI Free, Consequences Be Damned,
TechCrunch (Aug. 12, 2022, 3:55 PM), https://techcrunch.com/2022/08/12/a-startup-wants-to-
28 democratize-the-tech-behind-dall-e-2-consequences-be-damned/

1 LAION’s stated goal is “to make large-scale machine learning models, datasets and related code
2 available to the general public.” All of LAION’s projects are made available for free.
3 102. One of LAION’s most well-known projects is the image datasets it used train AI
4 systems.
5 103. In August 2021, LAION released LAION-400M, a dataset of 400 million Training
6 Images that included text captions. The Training Images in LAION-400M and their text captions
7 were copied or scraped from web pages or other sources without the consent of the image owners
8 or website operators. At the time, LAION-400M was the largest freely available dataset of its
9 kind.
10 104. Stability paid LAION to create LAION-5B, a new dataset of 5.85 billion Training
11 Images—more than 14 times bigger than LAION-400M. The only reason LAION-5B exists is
12 because Stability paid for it, so that Stability could have sufficient Training Images for Stable
13 Diffusion.
14 105. At the time, Stability was in the process of developing Stable Diffusion. As
15 admitted by Mostaque, Stability needed a bigger set of training images for training Stable
16 Diffusion.
17 106. Mostaque has publicly acknowledged the importance of using licensed training
18 images, saying that future versions of Stable Diffusion would be based on “fully licensed” training
19 images.9 But for the current version, he took no steps to obtain or negotiate suitable licenses.
20 107. Stability also paid LAION to create LAION-Aesthetics, a subset of LAION-5B
21 containing the images rated most highly for beauty and visual appeal by testers of Stable
22 Diffusion. To improve the quality of the output images, Stable Diffusion received more
23 concentrated training on version 2.5 of the LAION-Aesthetics dataset,10 which contains 600
24 million Training Images.
25 9
See @EMostaque, Twitter (Dec. 15, 2022, 8:03 AM),
26 https://twitter.com/EMostaque/status/1603390169192833027.
10
27 See Stable Diffusion v1 Model Card, GitHub, https://github.com/CompVis/stable-
diffusion/blob/main/Stable_Diffusion_v1_Model_Card.md#training (last visited Dec. 21,
28 2022).

1 108. Because LAION releases its datasets to the public, it is possible to study the
2 sources of the data, including the websites that the Training Images were scraped or copied from.
3 109. The LAION-Aesthetics dataset is heavily reliant on scraping and copying images
4 from commercial image-hosting services: according to one study, 47% of the images in the dataset
5 were scraped from only 100 web domains.11 The sources of some of the copies and scrapes are
6 stock-image sites, including Getty Images, Shutterstock, and Adobe Stock, as well as shopping
7 sites (like Shopify, Pinterest, Wix, and Squarespace). Significantly, websites featuring user-
8 generated content were a huge source of images, including sites like Smugmug, Flickr,
9 Wikimedia, Tumblr, and DeviantArt.
10 110. DeviantArt is the source of a significant portion of the LAION-Aesthetic dataset
11 amounting to an estimate of one out of every fifty images. Reasonable estimates show that there
12 are likely 3.3 million images from DeviantArt in the LAION-Aesthetics dataset.
13 F. DeviantArt’s betrayal of its artist community by embracing generative AI images
14 111. In 2000, Angelo Sotira, Scott Jarkoff, and Matthew Stephens founded DeviantArt.
15 112. In 2017, Wix acquired DeviantArt. Wix acquired all of DeviantArt’s corporate
16 stock for $36 million.
17 113. Shortly afterward, in March 2017, Moti Levy became Chief Operating Officer
18 (COO) of DeviantArt. In April 2022, Levy was promoted to Chief Executive Officer (CEO) of
19 DeviantArt.
20 114. Since its founding in 2000, DeviantArt has held itself out as an online community
21 friendly to artists, colloquially known on the site as “deviants.” A primary activity of artists on
22 DeviantArt is sharing digital images of their artwork, colloquially called “deviations.” Today,
23 DeviantArt bills itself as “the world’s largest art community,” hosting millions of such images.
24 115. DreamUp is a commercial product for DeviantArt and available only to customers
25 who pay DeviantArt. DeviantArt offers paid subscriptions to its artist members called “Core
26
11
27 Andy Baio, Exploring 12 Million of the 2.3 Billion Images Used to Train Stable Diffusion’s Image
Generator, Waxy (Aug. 30, 2022), https://waxy.org/2022/08/exploring-12-million-of-the-images-
28 used-to-train-stable-diffusions-image-generator/.

1 Plans.” Custom Core Plans typically range in price from $3.95 to $14.95 per month. To use
2 DreamUp, a member must first subscribe to a Core Plan. A Core Plan subscriber is allowed to use
3 DreamUp for a certain number of Text Prompts per month. For instance, the $9.95 “Pro” level
4 permits 200 DreamUp Text Prompts per month. Core Plan members can purchase additional
5 Text Prompts by purchasing packages of “points.” DeviantArt charges $1 for 80 points, with a
6 minimum purchase of 400 points for $5.
7 116. Stability copied thousands—and possibly millions—of the Training Images from
8 DeviantArt created by artists and other DeviantArt subscribers.
9 117. DeviantArt claims that DreamUp “lets you create AI art knowing that creators
10 and their work are treated fairly.”
11 118. This statement is false and misleading.
12 119. Like Stability, DeviantArt has not attempted to negotiate licenses for any of the
13 Training Images. Like Stability, DeviantArt is not sharing any of the revenue from the DreamUp
14 app with the artists or other owners of the Training Images.
15 120. DeviantArt has betrayed its artist community. Rather than standing up for the
16 rights of its members by rejecting Stable Diffusion and other sources of AI-generated art,
17 DeviantArt has gone the opposite direction: it has built an app called DreamUp that is based on
18 Stable Diffusion.
19 121. By releasing DreamUp as a paid product, DeviantArt is a co-conspirator in the
20 illegal use of copyrighted works.
21 122. In addition, by offering for sale AI-generated work based on Stable Diffusion,
22 DeviantArt is in fact competing with and displacing the work of the artists and other subscribers
23 to DeviantArt.
24 123. DeviantArt’s choice to embrace Stable Diffusion by incorporating it into their
25 website via the DreamUp app violates their own terms of service and privacy policy and
26 represents unfair competition against their artist customers.
27 124. DeviantArt’s terms of service have long contained a strict prohibition against
28 using content on the site “for any commercial purpose,” and also says no “commercial activities

1 are permitted on or through the Service without DeviantArt’s written approval.” The terms of
2 service elsewhere state that “You may not reproduce, distribute, publicly display or perform, or
3 prepare derivative works based on any of the [DeviantArt-hosted artworks] without the express,
4 written consent of DeviantArt or the appropriate owner of copyright in such works.”
5 125. DeviantArt was aware or reasonably should have been aware that Stability was
6 acting in violation of those terms. Thus, having been put on notice that DeviantArt images had
7 been used for a commercial purpose—namely, training Stable Diffusion—DeviantArt could have
8 taken legal action against Stable Diffusion for violating those terms.
9 126. DeviantArt did not do so.
10 127. There is no evidence that DeviantArt, despite its professed dedication to its terms
11 of service, has ever challenged Stability’s violation of the terms of service by training Stable
12 Diffusion on the work of DeviantArt members.
13 128. The scope of DeviantArt’s betrayal of its artist community by embracing Stable
14 Diffusion was evident in a group audio session held by DeviantArt management on November 11,
15 2022, from approximately 1:00–2:30 pm Pacific Time. DeviantArt scheduled the discussion
16 specifically to allay the well-founded concerns of DeviantArt members that DeviantArt’s embrace
17 of AI art was a complete repudiation of its longstanding community principles, as well as
18 economically and legally unfair.
19 129. At one point in the audio session, CEO Moti Levy explicitly took ownership of the
20 decision to bring Stable Diffusion onto DeviantArt via the DreamUp app: “The reason why we're
21 using Stable Diffusion because it's the only option for us to take an open source [software engine]
22 and modify it . . . . The other platforms or the other companies do not allow it. . . . [A]nd by the
23 way, that was my decision. That’s our decision by me as the CEO. That's my decision to take
24 Stable Diffusion.” (Emphasis added.)
25 130. Shortly after the end of this audio session, DeviantArt updated its terms of
26 service. DeviantArt added a new paragraph about “Data Scraping & Machine Learning
27 Activities” that explicitly permits this kind of usage under certain circumstances, so that Stable
28 Diffusion and future generative AI services can continue to scrape DeviantArt for images. In so

1 doing, DeviantArt has reneged on its promises. It plainly switched its loyalties from its artist
2 members to the AI companies, like Stability, infringing Plaintiffs’ and the Class’s intellectual
3 property rights in the work of those members. According to the Internet Archive, this new data-
4 scraping provision was added to the DeviantArt terms of service on November 11, 2022,
5 sometime between 1:41pm and 4:22pm Pacific Time.
6 131. Furthermore, although the new “Data Scraping” provision acknowledges that
7 certain kinds of data scraping will continue to be an “unauthorized use” of the DeviantArt
8 website, that “owners of the works are responsible for policing their own works”. In other words,
9 despite its professed interest in using its terms of service to protect artists, DeviantArt is washing
10 its hands of the matter. Instead of standing up for artists and using its resources to combat illegal
11 AI data scraping, it is forcing artists to take matters into their own hands.
12 G. Midjourney: the 21st-century collage tool (in)famous for its artistic style
13 132. Midjourney is a generative AI company based in San Francisco, California.
14 133. Midjourney was founded in August 2021 by David Holz, who also serves as CEO.
15 According to its website, Midjourney is “self-funded” and employs “11 full-time staff”.
16 134. Midjourney’s main product is an online AI-based image generator offered under
17 the name “Midjourney.” Like DreamUp and DreamStudio, the Midjourney image generator uses
18 Text Prompts as input and produces digital images as output. Just like DreamUp and
19 DreamStudio, Midjourney relies on Stable Diffusion as its underlying software engine for
20 generating images.
21 135. Midjourney released the first version of its service in March 2022, and has
22 continued to update the product continually since. Version 2 was launched in April 2022. Version
23 3 was launched in July 2022. On November 10, 2022, the alpha iteration of Version 4 was
24 launched.
25 136. Midjourney has deployed its service through an internet-chat system called
26 Discord. Users can visit certain Discord servers where Midjourney is enabled and use the
27 command “\imagine . . .” in a group chat room to introduce a Text Prompt. Midjourney will take
28

1 this Text Prompt and return an image within the chat-room window. These chat rooms are shared
2 by other users, so everyone can see each others’ Text Prompts, and the images that result.
3 137. Midjourney allows anyone to sample its service for free by providing a small
4 number of image outputs in response to Text Prompts. Midjourney offers a number of paid
5 subscription plans. For instance, its “Standard” plan costs $30 per month and allows unlimited
6 Text Prompts and digital image outputs. For an additional $20 per month, a customer can get
7 “Private Visibility”, allowing users to keep private Text Prompts, images, and digital output.
8 Midjourney also has a “Corporate” plan for $600 per year per person that is “required for
9 employees” of companies with “over $1Million/year [sic] in gross revenue.”
10 138. Though Holz has described Midjourney as a “diverse research lab” that is “not
11 really financially motivated,” Holz intends for Midjourney to make money by enrolling corporate
12 and other professional customers to generate images. Holz has said that “Millions are using
13 [Midjourney] … maybe 30%-50% of our users right now are professionals.”
14 139. Midjourney subscribers also receive access to the Midjourney web app, similar to
15 DreamStudio or DreamUp, which lets users access the Midjourney service through a web
16 interface.
17 140. In its terms of service, Midjourney calls the images generated by its service
18 “Assets.” The terms of service require that users grant Midjourney an “irrevocable copyright
19 license . . . [in the] Assets produced by the service at your direction.” Midjourney is therefore
20 necessarily asserting that the images generated by its system are copyrightable, and that the
21 copyright inheres in the subscriber who makes the image.
22 141. According to Holz, Midjourney distinguishes itself from competitors like
23 DreamUp and DreamStudio by being “focused toward making everything beautiful and artistic
24 looking.”
25 142. In September 2022, New York resident Kris Kashtanova sought and received U.S.
26 copyright registration for a comic book titled Zarya of the Dawn, featuring images generated by
27 Midjourney. In December 2022, the U.S. Copyright Office revoked this registration, deeming the
28 work ineligible for registration because it was generated by AI.

1 143. Also in September 2022, Colorado resident Jason Allen used Midjourney to
2 generate an image that he submitted to an art competition at the Colorado State Fair, which later
3 won. In response to artists who felt he had cheated, Allen later told the New York Times, “Art is
4 dead, dude. It’s over. A.I. won. Humans lost.” He also said that while using Midjourney, “I felt
5 like … some otherworldly force was involved.”
6 144. But the secret to Midjourney isn’t some “otherworldly force”. Just like DreamUp
7 and DreamStudio, Midjourney relies on appropriating millions of copyrighted images created by
8 artists and using these images as Training Images.
9 145. Thus, just like DreamUp and DreamStudio, Midjourney is a collage tool, only
10 capable of producing images that are remixed and reassembled from the copyrighted work of
11 others.
12 146. Holz has been sanguine about the copyright-infringement aspect of Midjourney,
13 saying that “To my knowledge, every single large AI model is basically trained on stuff that’s on
14 the internet. And that’s okay, right now. There are no laws specifically about that.” (Emphasis
15 added.)
16 147. That statement is false. There are a number of laws that protect and preserve the
17 rights and interests with respect to their art.
18 148. Holz has been cagey when asked direct questions about the source of
19 Midjourney’s Training Images. When asked how the dataset of Training Images was built, he said
20 “It’s just a big scrape of the internet. We use the open data sets that are published and train
21 across those. And I’d say that’s something that 100% of people do. We weren’t picky.”
22 149. On information and belief, because the LAION image datasets are the only large
23 “open data sets that are published,” Holz’s comment implies that Midjourney has used the
24 LAION image datasets for training. In August 2022, Midjourney released a beta version that used
25 SD.
26 150. When asked whether he sought consent from the creators of the Training Images,
27 Holz said “No. There isn’t really a way to get a hundred million images and know where
28 they’re coming from. . . . There’s no way to find a picture on the internet, and then

1 automatically trace it to an owner and then have any way of doing anything to authenticate it.”
2 (Emphasis added.)
3 151. Holz’s statement is false. LAION and other open datasets are simply lists of URLs
4 on the public web. Many of those URLs are derived from a small handful of websites that
5 maintain records of image ownership. Thus, many images could be traced to their owner. Holz
6 and LAION possess information sufficient to perform such tracing.
7 152. But Holz is correct that the project of licensing artworks ethically and complying
8 with copyright is not automatic—on the contrary, it is difficult and expensive. This is why Holz
9 was able to say in August 2022, one year after Midjourney’s founding: “To be honest, we're
10 already profitable, and we’re fine.” This stands to reason: Midjourney skipped the expensive part
11 of complying with copyright and compensating artists, instead helping themselves to millions of
12 copyrighted works for free.
13 IX. CLAIMS FOR RELIEF
14 COUNT I
15 DIRECT COPYRIGHT INFRINGEMENT
17 U.S.C. §§ 106, et seq.
16 (All Defendants)
17 153. Plaintiffs and the Class hereby repeat and incorporate by reference each preceding
18 and succeeding paragraph as though fully set forth herein.
19 154. As the owners of the copyright rights associated with the Works and/or Training
20 Images, Plaintiffs and the Class hold the exclusive rights under 17 U.S.C. § 106.
21 155. Defendants had access to but were not licensed by Plaintiffs or the Class to train
22 any machine learning, AI, or other computer program, algorithm, or other functional prediction
23 engine using the Works.
24 156. Defendants had access to but were not licensed by Plaintiffs nor the Class to
25 incorporate the Works into the products offered by Stability, DeviantArt, Midjourney, or related
26 software applications.
27
28

1 157. Defendants had access to but were not licensed by Plaintiffs or the Class to
2 download, store, or distribute copies of the Works for use in training or otherwise creating AI
3 Image Products.
4 158. Defendants had access to but were not licensed by Plaintiffs nor the Class to create
5 Derivative Works based upon the Works.
7 distribute the Works.
8 160. Defendants directly infringed Plaintiffs’ and the Class’s rights because they have:
9 a. reproduced one or more of the Works in violation of 17 U.S.C. § 106(1);
10 b. prepared Derivative Works based upon one or more of the Works in violation of 17
11 U.S.C. § 106(2);
12 c. distributed copies of one or more of the Works to the public in violation of 17
13 U.S.C. § 106(3);
14 d. performed one or more of the Works publicly in violation of 17 U.S.C. § 106(4);
15 and/or
16 e. displayed one or more of the Works publicly in violation of 17 U.S.C. § 106(5).
17 161. Plaintiffs and the Class have been damaged by Defendants’ actions.
18 162. Defendants have directly and indirectly profited from their acts of infringement.
19 163. Defendants have infringed the Training Images for commercial purposes.
20 164. Defendants are using copies of the Training Images interconnected with their AI
21 Image Products to generate digital images and other output that are derived exclusively from the
22 Training Images, and that add nothing new.
23 165. Defendants’ AI Image Products produce digital images and other output that act
24 as market substitutes for the underlying Training Images, thereby competing with Plaintiffs and
25 members of the Class.
26 166. Defendants’ AI Image Products contain copies of every image in the set of
27 Training Images and are capable at any time of producing as output a copy of any of the Training
28 Images.

1 167. Defendants’ AI Image Products, because they generate images derived from the
2 Training Images, will substantially negatively impact the market for the work of Plaintiffs and the
3 Class.
4 168. The conduct of Defendants is causing and, unless enjoined and restrained Court,
5 will continue to cause Plaintiffs and the Class great and irreparable injury that cannot fully be
6 compensated or measured in money and have no adequate remedy at law.
7 COUNT II
8 VICARIOUS COPYRIGHT INFRINGEMENT
17 U.S.C. §§ 106, et seq.
9 (All Defendants)
12 170. As the owners of the copyright rights, Plaintiffs and the Class hold the exclusive
13 rights under 17 U.S.C. § 106.
14 171. Individuals have used AI Image Products to create works using the names of
15 Plaintiffs and the Class in prompts and passed those works off as original works by the artist
16 whose name was used in the prompt. Such individuals are referred to herein as “Imposters” By
17 using a particular artist’s name, Imposters can cause the AI Image Product to rely more heavily
18 on that artist’s prior works to create images that can pass as original works by that artist. These
19 output images are referred to herein as “Fakes.”
20 172. The ability of AI Image Products to respond to prompts containing specific artists’
21 names was designed by Defendants with either the knowledge of or reckless disregard for the fact
22 that this functionality could easily be used to create Fakes.
23 173. Imposters have sold Fakes on ArtStation, Kickstarter, the Unreal Engine
24 Marketplace, and elsewhere.
25 174. Plaintiffs and the Class have been damaged by Imposters’ actions.
26 175. The Defendant-owner of the AI Image Product used to create each Fake is
27 vicariously liable for any infringements committed by Imposters.
28

1 176. Defendants have directly and indirectly profited from acts of infringement by
2 Imposters.
3 177. The conduct of Defendants and Imposters is causing and, unless and until
4 enjoined and restrained by this Court, will continue to cause Plaintiffs and the Class great and
5 irreparable injury that cannot fully be compensated or measured in money and have no adequate
6 remedy at law.
7 COUNT III
8 VIOLATION of the DIGITAL MILLENNIUM COPYRIGHT ACT
17 U.S.C. §§ 1201–1205
9 (All Defendants)
12 179. Plaintiffs and members of the Class own the copyrights to Works used to train AI
13 Image Products. The AI Image Products were trained on billions of images found on the internet.
14 180. Plaintiffs and members of the Class included the following Copyright
15 Management Information (as defined in Section 1202(c) of the DMCA) (“CMI”) in the Works:
16 a. copyright notices;
17 b. the title and other information identifying the Works;
18 c. the name of, and other identifying information about, the creators of the Works;
19 and
20 d. the name of, and other identifying information about, the copyright owners of the
21 Works.
22 181. Defendants did not contact Plaintiffs nor the Class to obtain authority to remove
23 or alter CMI from the Works within the meaning of the DMCA.
24 182. Defendants knew that they did not contact Plaintiffs nor the Class to obtain
25 authority to remove or alter CMI from the Works within the meaning of the DMCA.
26 183. As part of the scheme, Defendants did not attempt to contact Plaintiffs to obtain
27 authority to remove or alter CMI from the Works within the meaning of the DMCA. In fact,
28 Defendants’ removal of CMI made it difficult or impossible to contact Plaintiffs and the Class to

1 obtain authority to remove or alter CMI from the Works within the meaning of the DMCA.
2 Rather, Defendants removed or altered CMI from images that are owned by Plaintiffs and the
3 Class by training Stable Diffusion on those images and designing it to omit any CMI as part of the
4 output.
5 184. Without the authority of Plaintiffs and the Class, Defendants intentionally
6 removed or altered CMI from the Works after they were posted on DeviantArt or other websites.
7 185. Defendants had access to but were not licensed by Plaintiffs nor the Class to train
8 any machine learning, AI, or other pseudo-intelligent computer program, algorithm, or other
9 functional prediction engine using the Works.
11 incorporate the Works into their AI Image Products.
12 187. Defendants had access to but were not licensed by Plaintiffs nor the Class to create
13 Derivative Works based upon the Works.
15 distribute the Works as they do through Defendants’ AI Image Products.
16 189. Without the authority of Plaintiffs and the Class, Defendants distributed CMI
17 knowing that the CMI had been removed or altered without authority of the copyright owner or
18 the law with respect to the Works.
19 190. Defendants distributed copies of the Works knowing and intending that CMI had
20 been removed or altered without authority of the copyright owner or the law, with respect to the
21 Works.
22 191. Defendants removed or altered CMI from the Works knowing and intending that it
23 would induce, enable, facilitate, or conceal infringement of copyright. Both in the dataset used to
24 source the Training Images as well as the places on the Internet where the Training Images were
25 found set forth CMI, such as the creator’s name. CMI is also incorporated into the Works in the
26 form of artist’s signatures. When any of Defendants’ AI Image Products output an image, the
27 CMI that was previously included with the Works the image is based upon is removed.
28

1 192. Without the CMI associated with the Works, users of AI Image Products are
2 induced or enabled to copy the Works and/or make Derivative Works based on them. Without the
3 CMI, copyright infringement is facilitated or concealed, because Plaintiffs and the Class are
4 prevented from knowing or learning that the Output is based upon one or more of the Works.
5 193. Defendants removed or altered CMI from Works owned by Plaintiffs and the Class
6 while possessing reasonable grounds to know that it would induce, enable, facilitate, and/or
7 conceal infringement of copyright in violation of the DMCA. By omitting and concealing CMI
8 from Output, Defendants have reasonable grounds to know that innocent infringers are induced
9 or enabled to copy the Works, because CMI has been removed. Without the CMI, Defendants
10 have reasonable grounds to know copyright infringement is facilitated or concealed, because
11 Plaintiffs and the Class have the difficult or impossible task of proving the Works belong to them.
12 194. The profits attributable to Defendants’ violation of the DMCA include the
13 revenue from: any AI Image Products they offer that incorporate Stable Diffusion and other AI
14 Image Products trained on copyrighted images owned by Plaintiffs and/or the Class. The Works
15 add nearly all value to these products because the purpose of them is to provide images and the
16 source of those images are the Works. Without the Works, these products would not be
17 functional.
18 195. On information and belief, Defendants could have trained their AI Image Products
19 to include any CMI associated with the Works relied on to create a given Output when providing
20 that Output.
21 196. Defendants did not request or obtain permission from Plaintiffs and the Class to
22 use the Works to train their AI Image Products.
23 197. Unless Defendants are enjoined from violating the DMCA, Plaintiffs and the Class
24 will suffer great and irreparable harm by depriving them of the right to identify and control the
25 reproduction and/or distribution of their copyrighted works and to pursue copyright-
26 infringement remedies. Defendants will not be damaged if they are required to comply with the
27 DMCA. Plaintiffs and the Class members are therefore entitled to an injunction barring
28 Defendants from violating the DMCA and impounding any device or product that is in the

1 custody or control of Defendants and that the court has reasonable cause to believe was involved
2 in a violation of the DMCA.
3 198. Plaintiffs and the Class are further entitled to recover from Defendants the actual
4 or statutory damages Plaintiffs and the Class sustained pursuant to 17 U.S.C. § 1203(c) and for
5 Plaintiffs’ and the Class’s costs and attorneys’ fees in enforcing the Licenses. Plaintiffs and the
6 Class are also entitled to recover as restitution from Defendants for any unjust enrichment,
7 including gains, profits, and advantages that Defendants have obtained as a result of their breach
8 of the Licenses.
9 199. Defendants conspired together and acted jointly and in concert pursuant to their
10 scheme to commit the acts that violated the DMCA alleged herein.
11 200. Defendants induced their customers to unknowingly violate the DMCA by
12 withholding attribution and other information as described herein.
13 COUNT IV
14 VIOLATION of the STATUTORY RIGHT of PUBLICITY
Cal. Civ. Code § 3344
15 (All Defendants)
18 202. Defendants knowingly used Plaintiffs’ names in Defendants’ AI Image Products.
19 At no time did Plaintiffs consent to Defendants’ use of their names in this capacity.
20 203. Defendants appropriated Plaintiffs’ names to Defendants’ advantage, including for
21 the purposes of advertising, selling, and soliciting purchases through Defendants’ AI Image
22 Products. Defendants’ AI Image Products can be directed to prioritize inclusion of specific
23 artists’ Works by invoking the name of the artist or artists. This was a function designed and
24 promoted by Defendants as a product feature.
25 204. Plaintiffs have invested considerable energy, effort, ingenuity, and creativity into
26 the development of their distinct artistic identities and have successfully built careers as artists.
27 Plaintiffs’ names are uniquely associated with their art and artistic styles and are recognizable to
28

1 the public. Plaintiffs have derived value from their names, identities, and distinctive artistic
2 styles.
3 205. There is a direct connection between Defendants’ misappropriation of Plaintiffs’
4 names and Defendants’ commercial purposes, because Defendants used Plaintiffs’ names to
5 advertise art “in the style” of Plaintiffs’ work. Defendants used Plaintiffs’ names and advertised
6 their AI’s ability to copy or generate work in the artistic style that Plaintiffs popularized in order
7 to sell Defendants’ products and services. Defendants’ ability to market art similar to and
8 associated with Plaintiffs’ names also enabled Defendants to establish an advantage over actual
9 and prospective competitors.
10 206. Defendants’ use of Plaintiffs’ names was not incidental. Rather, Defendants
11 specifically and knowingly used Plaintiffs’ names because these names were uniquely related to
12 specific artistic styles, and Defendants generated valuable business from their ability to sell
13 artworks “in the style” that Plaintiffs popularized. Thus, the use of Plaintiffs’ names contributed
14 value to Defendants’ platform and services.
15 207. Defendants used Plaintiffs’ names to link and associate the art generated by its AI
16 with Plaintiffs’ specific styles and artistic accomplishments. This link uniquely enhanced the
17 marketability of Defendants’ AI art-generating services to consumers and the public.
18 208. Defendants’ emphasized the ability of AI Image Products to create images based
19 on “in the style” prompts that included specific Class members’ names. This functionality was
20 prominent and used throughout Defendants’ apps, website, and social media posts.
21 209. Thus, Defendants’ misappropriation of Plaintiffs’ names is directly connected
22 with Defendants’ advertising and sale of their products and services.
23 210. Because Defendants advertise the ability of their systems to generate artwork “in
24 the style” of Plaintiffs’ work—and explicitly used Plaintiffs’ work to train their AI algorithms—
25 the art generated by Defendants’ AI products is not transformative. Defendants’
26 misappropriation merely capitalizes on Defendants’ theft of Plaintiffs’ artistic work and the
27 associated value of Plaintiffs’ names.
28

1 211. Defendants appropriated Plaintiffs’ names exclusively for commercial purposes.

2 Defendants’ appropriation was not done in connection with any news, public affairs, sports
3 broadcast or account, or political campaign.
4 212. Because of Defendants’ unlawful appropriation of Plaintiffs’ names, Plaintiffs have
5 suffered injury. Plaintiffs have a right to protect the goodwill that is associated with their names,
6 and that goodwill is compromised by a proliferation of AI-generated art created without Plaintiffs’
7 consent but associated with Plaintiffs’ names. Further, the value of Plaintiffs’ name recognition—
8 and thus the value of their art itself—is diluted in a market flooded with AI-generated copies
9 associated with Plaintiffs’ names and artistic styles. Rather than generating revenue by licensing
10 their own images, Plaintiffs also suffer injury through having to compete with knock-off images
11 generated from Plaintiffs’ work and associated with Plaintiffs’ names.
12 COUNT V
13 VIOLATION of the COMMON LAW RIGHT of PUBLICITY
Common Law
14 (All Defendants)
17 214. Plaintiffs’ identities are distinctly tied to their work as artists and their specific
18 artistic styles. Plaintiffs have invested considerable energy, effort, ingenuity, and creativity into
19 the development of their distinct artistic identities and name-recognition. Plaintiffs have also built
20 careers as artists and have derived value from their names, identities, and distinctive artistic
21 styles.
22 215. Plaintiffs’ names and artistic identities are not limited to a specific copyrighted
23 image or work developed by Plaintiffs. Rather, they extend to Plaintiffs’ entire corpus of work and
24 allow consumers and the public to identify work “in the style of” Plaintiffs. Thus, Defendants did
25 not only misappropriate work fixed in a tangible medium of expression, but also misappropriated
26 Plaintiffs’ names and artistic identities.
27 216. Plaintiffs did not consent to Defendants’ use of their names or identities.
28

1 217. Defendants knowingly used Plaintiffs’ names and identities to further Defendants’
2 commercial advantage, including for the purposes of advertising, selling, and soliciting purchases
3 through Defendants’ AI art-generating system.
4 218. Defendants used Plaintiffs’ names and distinct artistic identities to link and
5 associate the art generated by its AI with Plaintiffs’ specific styles and artistic accomplishments.
6 This link uniquely enhanced the marketability of Defendants’ AI art-generating services to
7 consumers and the public.
8 219. Defendants’ use of Plaintiffs’ names and identities was prominent and used
9 throughout Defendants’ apps, website, and social media posts.
10 220. Thus, Defendants’ misappropriation of Plaintiffs’ names and identities is directly
11 connected with Defendants’ advertising and sale of their products and services.
12 221. Because Defendants advertise the ability of their systems to generate artwork “in
13 the style” of Plaintiffs’ work—and explicitly used Plaintiffs’ work to train the algorithms—the art
14 generated by Defendants’ AI products is not transformative. Defendants’ misappropriation
15 merely capitalizes on Defendants’ theft of Plaintiffs’ artistic work and the associated value of
16 Plaintiffs’ names and identities.
17 222. Because of Defendants’ unlawful appropriation of Plaintiffs’ names and identities,
18 Plaintiffs have suffered injury. The goodwill associated with Plaintiffs’ names and distinct
19 identities is compromised by a proliferation of AI-generated art associated with Plaintiffs’ names
20 and identities, but created without Plaintiffs’ consent. The value of Plaintiffs’ name recognition
21 and Plaintiffs’ distinct artistic styles—and thus the value of their art itself—is diluted in a market
22 flooded with AI-generated copies built on Plaintiffs’ unique artistic identities. Plaintiffs also suffer
23 injury through having to compete with knock-off images generated from Plaintiffs’ work and
24 associated with Plaintiffs’ names.
25
26
27
28

1 COUNT VI
2 UNFAIR COMPETITION
15 U.S.C. § 1125; Cal. Bus. & Prof. Code §§ 17200, et seq.; and Common Law
3 (All Defendants)
6 224. Defendants have engaged in unlawful business practices, including:
7 a. Infringement of Plaintiffs’ and the Class’s copyrights; and
8 b. Violations of Plaintiffs’ and the Class’s rights under the DMCA.
9 The details of the unlawful business practices are set forth herein.
10 225. Plaintiffs and the Class have suffered monetary damages as a result of Defendants’
11 conduct.
12 226. The conduct of Defendants is causing and, unless enjoined and restrained by this
13 Court, will continue to cause Plaintiffs and the Class great and irreparable injury that cannot fully
14 be compensated or measured in money.
15 COUNT IX
16 BREACH of CONTRACT
VIOLATION of DEVIANTART POLICIES
17 Cal. Bus. & Prof. Code § 22575–22579; Cal. Civ. Code § 1798.150; and Common Law
(Against Defendant DeviantArt Only)
18
227. Plaintiffs and the Class hereby repeat and incorporate by reference each preceding
19
and succeeding paragraph as though fully set forth herein.
20
228. DeviantArt Plaintiffs and the DeviantArt Class are DeviantArt users who have
21
accepted DeviantArt’s Terms of Service. As a result, Plaintiffs and the Class have formed a
22
contract, the terms of which are set forth in DeviantArt’s Terms of Service.
23
229. Plaintiffs and the Class are DeviantArt users who have accepted DeviantArt’s
24
Privacy Statement. As a result, Plaintiffs and the Class have formed a contract.
25
230. DeviantArt’s Privacy Statement and Terms of Service share definitions and refer
26
to each other. As such, they are collectively referred to herein as “DeviantArt’s Policies” unless a
27
distinction is necessary. Copies of the November 11, 2022 version of the DeviantArt Terms of
28

1 Service and copies of the current DeviantArt Terms of Service and Privacy Policy are attached as
2 Exhibit 17.
3 231. Plaintiffs and the Class have performed each of the conditions, covenants, and
4 obligations imposed on them by the terms of DeviantArt’s Policies.
5 232. DeviantArt has substantially and materially breached DeviantArt’s Policies in the
6 following ways:
7 a. Sharing Plaintiffs’ and the Class’s personal data with unauthorized third parties in
8 violation of the DeviantArt Privacy Statement;
9 b. Selling and distributing Plaintiffs’ and the Class’s personal data in contravention
10 of the DeviantArt’s Policies;
11 c. Use of Plaintiffs’ and the Class’s personal data after the DeviantArt Privacy
12 Statement explicitly claims it will be deleted;
13 d. Use and distribution of Plaintiffs’ and the Class’s personal data outside the
14 limitations set forth in the DeviantArt Privacy Statement.
15 233. Plaintiffs and the Class have suffered monetary damages as a result of
16 DeviantArt’s conduct.
17 234. DeviantArt’s conduct is causing and, unless enjoined and restrained by this Court,
18 will continue to cause Plaintiffs and the Class great and irreparable injury that cannot fully be
19 compensated or measured in money.
20 235. As a direct and proximate result of these material breaches by DeviantArt,
21 Plaintiffs and the Class are entitled to an injunction requiring DeviantArt to comply with all the
22 terms of the DeviantArt Policies.
23 236. Plaintiffs and the Class are further entitled to recover from DeviantArt the
24 damages Plaintiffs and the Class sustained—including consequential damages—for Plaintiffs’ and
25 the Class’s costs in enforcing DeviantArt’s Policies. Plaintiffs and the Class are also entitled to
26 recover as restitution from DeviantArt for any unjust enrichment, including gains, profits, and
27 advantages that it has obtained as a result of its breaches of the DeviantArt Policies.
28

1 COUNT VII
2 DECLARATORY RELIEF
28 U.S.C. § 2201(a) and Cal. Code Civ. Proc. § 1060
3
(All Defendants)
4
7 238. An actual controversy exists between the Class (including Plaintiffs) and
8 Defendants due to Defendants’ operation of AI Image Products, which violate Plaintiffs’ and the
9 Class’s rights, including but not limited to their rights under the Copyright Act, DMCA Section
10 1202, Cal. Civ. Code § 3344, and Cal. Civ. Code § 17200 as alleged herein.
11 239. Plaintiffs are entitled to a judgment declaring that Defendants’ actions are
12 unlawful and, specifically, that Defendants violated the Copyright Act, DMCA Section 1202, Cal.
13 Civ. Code § 3344, and Cal. Civ. Code § 17200.
14 X. DEMAND FOR JUDGMENT
15 WHEREFORE, Plaintiffs requests that the Court enter judgment on their behalf and on
16 behalf of the Class defined herein, by adjudging and decreeing that:
17 240. This action may proceed as a class action, with Plaintiffs serving as Class
18 Representatives, and with Plaintiffs’ counsel as Class Counsel;
19 a. Judgment in favor of Plaintiffs and the Class and against Defendants;
20 b. An award of statutory and other damages pursuant to 17 U.S.C. § 504 for
21 violations of Plaintiffs’ and the Class’s copyright interests by Defendants, both
22 directly and vicariously through others;
23 c. Permanent injunctive relief, including but not limited to making changes to its
24 Defendants’ AI Image Products to ensure that all applicable information set forth
25 in 17 U.S.C. § 1203(b)(1) is included with any Output incorporating a Work that
26 had such information associated with it where Defendants found it on the internet;
27 d. An order of costs and allowable attorney’s fees pursuant to 17 U.S.C.
28 § 1203(b)(4)–(5);

1 e. An award of statutory damages pursuant to 17 U.S.C. § 1203(b)(3) and 17 U.S.C.

2 § 1203(c)(3), or, in the alternative, an award of actual damages and any additional
3 profits pursuant to 17 U.S.C. § 1203(c)(2) (including tripling damages pursuant to
4 17 U.S.C. § 1203(c)(4) if applicable);
5 f. An award of damages, including punitive damages, for harms resulting from
6 Defendants acts of unfair competition;
7 g. An award of damages sufficient to compensate Plaintiffs and the Class for harms
8 resulting from Defendants unjust enrichment; and
9 h. An award of damages, including punitive damages, for harms resulting from
10 Defendants violations of Plaintiffs’ and the Class’s rights of publicity.
11 241. Injunctive relief sufficient to alleviate and stop Defendants’ unlawful conduct
12 alleged herein.
13 242. Plaintiffs and the Class are entitled to prejudgment and post-judgment interest on
14 the damages awarded them, and that such interest be awarded at the highest legal rate from and
15 after the date this class action complaint is first served on Defendants;
16 243. Defendants are to be jointly and severally responsible financially for the costs and
17 expenses of a Court approved notice program through post and media designed to give immediate
18 notification to the Class.
19 244. Plaintiffs and the Class receive such other or further relief as may be just and
20 proper.
21 XI. JURY TRIAL DEMANDED
22 Pursuant to Federal Rule of Civil Procedure 38(b), Plaintiffs demand a trial by jury of all
23 the claims asserted in this Complaint so triable.
24
25
26
27
28

1 Dated: January 13, 2023 By: /s/ Joseph R. Saveri

Joseph R. Saveri
2
4 Joseph R. Saveri (State Bar No. 130064)

Cadio Zirpoli (State Bar No. 179108)
5 Christopher K.L. Young (State Bar No. 318371)
Elissa A. Buchanan (State Bar No. 249996)
6 Travis Manfredi (State Bar No. 281779)
7 JOSEPH SAVERI LAW FIRM, LLP
601 California Street, Suite 1000
8 San Francisco, California 94108
Telephone: (415) 500-6800
9 Facsimile: (415) 395-9940
[email protected]
11 [email protected]
[email protected]
12 [email protected]
13
Matthew Butterick (State Bar No. 250953)
14 1920 Hillhurst Avenue, #406
Los Angeles, CA 90027
15 Telephone: (323) 968-2632
16 Facsimile: (415) 395-9940
Email: [email protected]
17
18 Brian D. Clark (pro hac vice pending)

Laura M. Matson (pro hac vice pending)
19 LOCKRIDGE GRINDAL NAUEN P.L.L.P.
100 Washington Avenue South, Suite 2200
20
Minneapolis, MN 55401
21 Telephone: (612)339-6900
Facsimile: (612)339-0981
[email protected]
23
24 Counsel for Individual and Representative

Plaintiffs and the Proposed Class
25
26
27
28

Andersen v. Stability AI

Uploaded by

Copyright:

Available Formats

Andersen v. Stability AI

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Andersen v. Stability AI

Uploaded by

Copyright:

Available Formats

What is the purpose of this class action lawsuit?

What is the purpose of this class action lawsuit?

Who are the defendants named in the lawsuit?

Who are the defendants named in the lawsuit?

Case 3:23-cv-00201 Document 1 Filed 01/13/23 Page 1 of 46

1 Joseph R. Saveri (State Bar No. 130064)

16 UNITED STATES DISTRICT COURT

Class Action Complaint

2 I. AI IMAGE GENERATORS ARE 21ST-CENTURY COLLAGE TOOLS THAT

i Class Action Complaint

12 1. Stable Diffusion is a software product—defined below as an AI Image Product—

1 Class Action Complaint

2 Class Action Complaint

1 13. Defendant Midjourney is headquartered in San Francisco. Plaintiff Karla Ortiz

3 Class Action Complaint

4 Class Action Complaint

1 24. A “Text Prompt” is a textual description that is used as an interface to certain

5 Class Action Complaint

6 Class Action Complaint

7 Class Action Complaint

8 Class Action Complaint

1 VII. CLASS ALLEGATIONS

10 “Damages Class” under Rule 23(b)(3):

14 “DeviantArt Damages Subclass” under Rule 23(b)(3):

17 These “Class Definitions” specifically exclude the following person or entities:

9 Class Action Complaint

10 Class Action Complaint

1 iv. Right of Publicity Violations

11 Class Action Complaint

1 F. Other Class Considerations

12 Class Action Complaint

13 Class Action Complaint

14 Class Action Complaint

15 Class Action Complaint

1 incidental side effect. The primary goal of a diffusion model is to reconstruct

17 Class Action Complaint

18 Class Action Complaint

19 Class Action Complaint

20 Class Action Complaint

25 copied into those latent images.

21 Class Action Complaint

22 Class Action Complaint

23 Class Action Complaint

24 Class Action Complaint

25 Class Action Complaint

26 Class Action Complaint

27 Class Action Complaint

28 Class Action Complaint

29 Class Action Complaint

30 Class Action Complaint

31 Class Action Complaint

32 Class Action Complaint

33 Class Action Complaint

34 Class Action Complaint

35 Class Action Complaint

36 Class Action Complaint

37 Class Action Complaint

1 211. Defendants appropriated Plaintiffs’ names exclusively for commercial purposes.

38 Class Action Complaint