1:23-cv-00135
Getty Images US Inc v. Stability Ai Inc
I. Executive Summary and Procedural Information
- Parties & Counsel:
- Plaintiff: Getty Images (US), Inc. (New York)
- Defendant: Stability AI, Ltd. (United Kingdom); Stability AI, Inc. (Delaware); and Stability AI US Services Corporation (Delaware)
- Plaintiff's Counsel: Young Conaway Stargatt & Taylor, LLP
- Case Identification: Getty Images (US), Inc. v. Stability AI, Ltd., 1:23-cv-00135, D. Del., 07/08/2024
- Venue Allegations: Venue is alleged to be proper in the District of Delaware because Defendants Stability AI, Inc. and Stability AI US Services Corporation are incorporated in Delaware, and Defendant Stability AI, Ltd. is alleged to be subject to personal jurisdiction in the District through its operation of websites accessible to users in Delaware.
- Core Dispute: Plaintiff alleges that Defendant's artificial intelligence image-generation models were trained by unlawfully copying more than 12 million of Plaintiff's copyrighted photographs and associated metadata, and that the resulting models and their output infringe Plaintiff's copyrights and trademarks.
- Technical Context: The dispute is centered on generative artificial intelligence, specifically text-to-image diffusion models, which represent a rapidly developing technology for creating novel digital visual content from user prompts.
- Key Procedural History: This filing is a Second Amended Complaint. The complaint does not reference any prior litigation between the parties, Inter Partes Review (IPR) proceedings, or licensing negotiations related to the intellectual property at issue.
Case Timeline
| Date | Event |
|---|---|
| 1948-09-24 | '652 Patent Priority Date |
| 1953-10-27 | '652 Patent Issue Date |
| 1954-02-10 | '647 Patent Priority Date |
| 1955-01-28 | '851 Patent Priority Date |
| 1956-04-12 | '208 Patent Priority Date |
| 1958-06-03 | '208 Patent Issue Date |
| 1958-07-15 | '851 Patent Issue Date |
| 1958-07-22 | '647 Patent Issue Date |
| 1970-03-31 | '335 Patent Priority Date |
| 1971-09-07 | '335 Patent Issue Date |
| 1986-12-17 | '414 Patent Priority Date |
| 1988-11-11 | '997 Patent Priority Date |
| 1988-12-01 | '996 Patent Priority Date |
| 1990-11-06 | '996 Patent Issue Date |
| 1990-11-06 | '997 Patent Issue Date |
| 1993-04-06 | '414 Patent Issue Date |
| 1995-01-01 | Getty Images founded (approx.) |
| 2019-01-01 | Stability AI, Ltd. founded (approx.) |
| 2020-01-01 | Stability AI, Inc. founded (approx.) |
| 2021-03-01 | Stability AI, Inc. completes Y Combinator incubator program (approx.) |
| 2022-08-01 | Stability AI launches Stable Diffusion and DreamStudio (approx.) |
| 2023-03-01 | Stability AI US Services Corporation incorporated (approx.) |
| 2024-07-08 | Complaint Filing Date |
II. Technology and Patent(s)-in-Suit Analysis
U.S. Patent No. 2,656,652, "APPARATUS FOR TREATMENT WITH LIQUIDS" (Issued Oct. 27, 1953)
The Invention Explained
- Problem Addressed: The patent describes the technical challenge of efficiently and thoroughly cleaning, degreasing, or otherwise treating machine parts and similar articles with liquid solvents in a continuous, automated process ʼ652 Patent, col. 1:1-12 A key difficulty is ensuring both effective agitation within the solvent and complete removal of the solvent afterward ʼ652 Patent, col. 1:40-49
- The Patented Solution: The invention proposes a conveyor system that carries articles in rotatable, reticulate holders through a series of solvent wells ʼ652 Patent, col. 1:35-49 A novel mechanism positively rotates the holders during submersion to agitate the contents for thorough treatment and again after emergence from the liquid to ensure complete drainage ʼ652 Patent, col. 4:48-64 This automated, multi-stage processing with controlled agitation and clearing steps is the core of the solution ʼ652 Patent, Fig. 1
- Technical Importance: The apparatus provided a significant improvement in the efficiency and consistency of industrial parts-cleaning operations over prior manual or static immersion methods.
- Analogy: The patented apparatus can be analogized to a modern data processing pipeline. The "trays" carrying "articles" are akin to data batches (e.g., images) being fed into a system. The "solvent wells" represent different algorithmic processing stages. The key mechanism for "rotating" the holders to agitate and drain them is analogous to the iterative transformation of data in an AI training process, such as the repeated adding and removing of "noise" in a diffusion model to teach the model how to construct an image.
Key Claims at a Glance
- The complaint does not specify which claims are asserted, but an analysis would likely begin with Independent Claim 1.
- Independent Claim 1 of the '652 Patent recites these essential elements:
- A well containing a treating liquid.
- A conveyor movable over the well.
- A carrier with a holder for articles, where the holder is rotatively supported.
- A wheel to divert the conveyor and immerse the holder in the liquid.
- A stationary track beneath the wheel.
- A roller on the holder's pivot axis.
- Control means to guide the roller onto the track.
- Means to rotate the holder as it travels along the track while submerged.
- The complaint does not preclude the assertion of dependent claims.
U.S. Patent No. 2,837,208, "SHIPPING SWING SUSPENSION FOR FRAGILE ARTICLES" (Issued June 3, 1958)
The Invention Explained
- Problem Addressed: The patent addresses the problem of shipping delicate or fragile articles, such as electronic equipment, in a way that protects them from damage caused by shock and impact during handling and transport ʼ208 Patent, col. 1:15-22
- The Patented Solution: The invention is a container system featuring an internal suspension structure. This structure supports a flexible "sling" or "hammock" that holds the fragile article ʼ208 Patent, col. 1:23-32 This design suspends the article away from the rigid outer walls of the container, effectively isolating it and allowing shocks to be absorbed by the flexible sling rather than being transmitted to the article itself ʼ208 Patent, col. 1:60-65 '208 Patent, Fig. 1
- Technical Importance: The invention provided a novel packaging method that improved the survivability of fragile goods during shipping by "floating" them within a protective outer shell.
Key Claims at a Glance
The complaint does not specify which claims are asserted, but an analysis would likely begin with Independent Claim 1.
Independent Claim 1 of the '208 Patent recites these essential elements:
- A container with vertical side walls and a bottom wall.
- A suspension device fitting within the container, with its own side wall portions that engage the container's walls.
- The suspension device is folded to create vertical corners and horizontal bends.
- These bends and folds are collapsed to form an inwardly projecting horizontal channel member on each side.
- An article-carrying means connected to ledges on opposite sides of the device.
The complaint does not preclude the assertion of dependent claims.
Multi-Patent Capsule: U.S. Patent No. 2,842,851, "ELECTROMECHANICAL MEMORY OR SYNCHRONIZING DEVICE" (Issued Jul. 15, 1958)
- Technology Synopsis: This patent describes a system for measuring a property of a material at a first station, storing that measurement value for a predetermined time delay, and then comparing it to a new measurement taken at a second station downstream in a production line ʼ851 Patent, abstract The system effectively synchronizes two measurements taken at different times and locations to isolate a change, such as the thickness of a newly applied coating. This is analogous to an AI model being trained to recognize and isolate specific features, such as a watermark, from a base image.
- Asserted Claims: The complaint does not specify, but analysis would likely begin with Independent Claims 1 and 7.
- Accused Features: The complaint alleges that Stable Diffusion can reproduce, modify, or remove watermarks Compl. ¶14 Compl. ¶84 The '851 patent's method of isolating a "coating" (watermark) from a "base material" (image) by using a delayed, synchronized measurement could be argued to be conceptually similar to the process the AI model learns.
(Additional patents were provided but not analyzed due to the two-patent limit for full analysis.)
III. The Accused Instrumentality
Product Identification
The accused instrumentalities are Defendants' generative artificial intelligence models, collectively referred to as "Stable Diffusion" (including but not limited to Stable Diffusion 2.0, 3.0, SDXL 1.0, and SDXL Turbo), and the associated commercial interface "DreamStudio" Compl. ¶¶9-10
Functionality and Market Context
Stable Diffusion is a text-to-image and image-to-image AI model that generates novel, computer-synthesized images in response to user prompts Compl. ¶9 The complaint alleges the model was created by copying over 12 million photographs and associated metadata from Getty Images' websites without a license Compl. ¶1 Compl. ¶11 This data was allegedly used to "train" the model through a process involving encoding images, adding "visual noise," and then teaching the model to remove the noise by comparing its output to the original, stored images Compl. ¶¶59-60
DreamStudio is a revenue-generating web interface that allows users to access the Stable Diffusion models without needing specialized hardware or coding knowledge Compl. ¶10 The complaint alleges that outputs from these models sometimes reproduce distorted versions of Getty Images' watermarks, creating confusion and tarnishing Plaintiff's brand Compl. ¶14 Compl. ¶85 The complaint includes a screenshot of the Stable Diffusion 2.1 Demo interface to illustrate the product's function Compl. ¶74
IV. Analysis of Infringement Allegations
'652 Patent Infringement Allegations
| Claim Element (from Independent Claim 1) | Alleged Infringing Functionality | Complaint Citation | Patent Citation |
|---|---|---|---|
| a well containing treating liquid | The collection of billions of image-text training data pairs loaded into computer memory, which are subjected to the algorithmic training process. | ¶59 | col. 3:1-4 |
| a conveyor movable over the well | The automated software process that ingests and passes batches of image-text data through the various stages of the training algorithm. | ¶59 | col. 3:9-12 |
| a holder for articles to be treated rotatively supported by the arm | The data structure or memory allocation representing a single image-text pair, which is held and algorithmically transformed during the training process. | ¶59 | col. 3:33-35 |
| means whereby the holder is rotated as it continues its travel along the track while submerged in the liquid | The core training steps of adding "visual noise" to an encoded image and subsequently teaching the model to "remove the noise" by comparing the result to the original, which constitutes an iterative transformation analogous to rotation for treatment. | ¶59 | col. 4:48-52 |
'208 Patent Infringement Allegations
| Claim Element (from Independent Claim 1) | Alleged Infringing Functionality | Complaint Citation | Patent Citation |
|---|---|---|---|
| a container having vertical side walls and a bottom wall | The overall software architecture of the Stable Diffusion model, which contains and organizes the data representations used during training and inference. | ¶9 | col. 2:10-12 |
| a suspension device fitting within said container... forming an inwardly projecting horizontal channel member | The model's "latent space," an abstract, multi-dimensional space where compressed representations (embeddings) of images are organized and structured, isolating them from the raw pixel data. | ¶59 | col. 2:25-28 |
| article-carrying means connected with ledges at opposite sides of said device | A specific vector or embedding that represents a single image, which is "carried" or "suspended" within the model's latent space, isolated from the outer container (the full model architecture). An example of a copied Getty Images photo that would be so represented is shown in the complaint. Compl. ¶77 | ¶59 | col. 2:66-68 |
Identified Points of Contention
- Scope Questions ('652 Patent): A central issue may be whether the term "apparatus for treatment with liquids", which is described in a purely mechanical context, can be construed to read on a software-based AI training pipeline. The interpretation of terms like "conveyor", "holder", and "rotated" will be critical, as Plaintiff may argue they cover the algorithmic processing of digital data, while Defendant may argue they are limited to the physical movement of physical objects.
- Scope Questions ('208 Patent): Similarly, a dispute may arise over whether a "shipping swing suspension for fragile articles" can be construed to cover a software model's architecture and its use of a latent space. The key question will be whether abstract data structures like "embeddings" and "latent space" fall within the scope of claim terms like "article-carrying means" and "suspension device".
- Technical Questions: A factual question for the court may be whether the function performed by the accused AI training process is technically analogous to the function described in the patents. For instance, does the iterative denoising process in Stable Diffusion's training achieve the same functional purpose (agitation and treatment) as the physical rotation of a parts-holder claimed in the '652 patent?
V. Key Claim Terms for Construction
'652 Patent: "means whereby the holder is rotated"
- Context and Importance: This term is central to the infringement theory. To prove infringement, Plaintiff would need to argue that the algorithmic manipulation of data within the Stable Diffusion training process constitutes "rotation" as claimed. Practitioners may focus on this term because its construction will determine whether a patent for a mechanical process can extend to a purely digital, abstract process.
- Intrinsic Evidence for Interpretation:
- Evidence for a Broader Interpretation: The patent states the purpose of rotation is to "agitate the work" ʼ652 Patent, col. 1:11 Plaintiff may argue that any process that "agitates" or transforms the subject matter for treatment, such as the denoising algorithm described in the complaint Compl. ¶¶59c-d, meets the functional goal of the claim, irrespective of the physical mechanism.
- Evidence for a Narrower Interpretation: The specification consistently describes a physical mechanism for rotation, including a "toothed pinion 85" on the holder and a "segmental track... formed as a rack with a series of pins 103" ʼ652 Patent, col. 3:27-28 ʼ652 Patent, col. 5:1-12 The figures, such as Figures 6-9, exclusively depict this mechanical gear-and-rack system, suggesting the term is limited to physical rotation.
'208 Patent: "article-carrying means"
- Context and Importance: This term's definition is critical for determining if a patent for a physical packaging system can read on abstract data representations. Plaintiff's infringement theory would likely depend on this term being construed to cover an image's embedding in a latent space.
- Intrinsic Evidence for Interpretation:
- Evidence for a Broader Interpretation: The patent describes this element as a "flexible hammock or sling-like carrying device" ʼ208 Patent, col. 1:26-27 intended to suspend an article in a "floating condition" ʼ208 Patent, col. 1:62 Plaintiff may argue that a latent space embedding is a "flexible" digital representation that "carries" the image's core information and "floats" it in an abstract space, functionally isolating it from raw data noise, which is analogous to the patent's goal of isolating an object from physical shocks.
- Evidence for a Narrower Interpretation: The specification describes the "article-carrying means" (sling 13) as being made of physical materials like "plastic, or cloth, or combinations" thereof for carrying "fragile articles" such as "electronic equipment" ʼ208 Patent, col. 2:66-73 All disclosed embodiments are physical, which may support a narrower construction limited to tangible objects and carriers. The complaint's visual evidence of a specific photo of soccer players could be presented by Plaintiff as the "fragile article" that is processed and held by the "article-carrying means" of the AI system Compl. ¶77
VI. Other Allegations
Indirect Infringement
The complaint alleges that Stability AI provides its Stable Diffusion models to the public, including through open-source releases and the DreamStudio interface, which "permits third party developers to access, use, and further develop the model" and "enables users to obtain images" Compl. ¶10 Compl. ¶78 These allegations could form the basis of a claim for induced infringement, suggesting that Defendants provide the tool and encourage or instruct others to perform infringing acts.
Willful Infringement
The complaint alleges that "Stability AI was well aware that the content it was scraping without permission from Getty Images' websites was protected by U.S. copyright law" Compl. ¶13 and knew that its acts "were in violation of the terms of use of Getty Images' websites" Compl. ¶94 It further alleges that Getty Images licenses its content for AI training, a commercial option that Defendants allegedly bypassed, suggesting knowledge of the value of the intellectual property being used Compl. ¶6 Compl. ¶82 These allegations of pre-suit knowledge and "callous disregard" for Plaintiff's rights could be used to support a claim for willful infringement.
VII. Analyst's Conclusion: Key Questions for the Case
The resolution of this dispute, as framed by an analysis of the provided patents against the complaint's allegations, may turn on several key questions for the court:
Definitional Scope and Analogy: A central issue will be one of claim construction and whether claim terms rooted in mid-20th century mechanical engineering can be interpreted to cover modern, abstract software processes. Can a patent for a physical "apparatus for treatment with liquids" ('652 patent) be construed to read on an algorithmic training pipeline for a generative AI model, and can a "shipping swing suspension" ('208 patent) read on the use of a latent space to manage digital data representations?
Doctrine of Equivalents: If literal infringement is not found, a key evidentiary question will be one of functional equivalence. Does the accused AI training process-which uses algorithmic "noise" addition and removal to teach a model Compl. ¶59-perform substantially the same function (iterative treatment), in substantially the same way (algorithmic transformation), to achieve substantially the same result (a processed output) as the claimed physical rotation of a parts holder in a solvent bath ('652 patent)?
The Nature of a "Copy": While primarily a copyright concept, the nature of the "copies" made during the AI training process will be a foundational technical question. The complaint alleges that multiple, non-transitory copies of images are made, encoded, and stored during training Compl. ¶59b Compl. ¶66 The characterization of these intermediate data structures will be critical in determining whether they constitute an infringing "use" of any patented methods for processing or holding "articles."